Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Automated contracts and the lawyers who don't review them: adoption and use of machine learning technology
(USC Thesis Other)
Automated contracts and the lawyers who don't review them: adoption and use of machine learning technology
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
AUTOMATED CONTRACTS AND THE LAWYERS WHO DON’T REVIEW THEM:
ADOPTION AND USE OF MACHINE LEARNING TECHNOLOGY
By
Beverly Rich
A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOHY
BUSINESS ADMINISTRATION
December 2021
Copyright 2021 Beverly Rich
ii
Acknowledgements
Thank you to the esteemed members of the USC Marshall Department of Management and
Organization for their support, dedication, and thoughtful feedback over the years. In particular, I
want to thank Kyle Mayer, Florenta Teodoridis, Nan Jia, Milan Miric, Roshni Raveendhran, and
Scott Wiltermuth for modeling how to be a both a productive researcher and a good colleague. I
owe a massive debt of gratitude to my family for their support and encouragement in this
endeavor. Thank you to my sons, Simon and Bernard, for their cuddles and giggles, and to my
husband, George, for being patient, for being funny, for always listening, and for reading paper
drafts (or at least the abstracts).
iii
TABLE OF CONTENTS
Acknowledgements……………………………………………………………………………….ii
List of Tables……………………………………………………………………………………..iv
List of Figures…………………………………………………………………………………….v
Abstract…………………………………………………………………………………………..vi
Introduction………………………………………………………………………………………..1
Chapter 1: Reconciling Technology Commercialization Literature with Artificial Intelligence…5
Subchapter: Connecting Chapters One and Two………………………...………26
Chapter 2: Technology Adoption and Commercialization Strategies in a Machine Learning
World: Examining the Use of Automated Contracts…………………………………………….28
Subchapter: Connecting Chapters Two and Three………………………………67
Chapter 3: Adoption or Acquisition? Examples from the Legal Technology Industry………….69
Subchapter: Conclusion………………………………………………………...101
References………………………………………………………………………………………103
iv
List of Tables
Table 1a. Descriptive Statistics of Adopters and Non-Adopters………………………………...60
Table 1b. Comparing Adopters and Non-Adopters……………………………………………...61
Table 1c. Adopters and Non-Adopters by Industry……………………………………………...61
Table 2a. Topics from onboarding calls: Customer responses…………………………………..63
Table 2b. Topics from onboarding calls: LawGeex responses…………………………………..64
Table 3. Customer calls chart…………………………………………………………………….65
Table 4a. Descriptives……………………………………………………………………………66
Table 4b. Descriptives…………………………………………………………………………...66
Table 5. Customer Firms………………………………………………………………………..100
v
List of Figures
Figure 1. Adopters and Non-Adopters by Geography of Headquarters…………………………62
Figure 2. Conceptual Model……………………………………………………………………..99
Figure 3. Firm Categories………………………………………………………………………..99
vi
Abstract
In this dissertation, I examine firms’ technology commercialization strategies and firms’
technology adoption strategies in light of use and proliferation of AI-enabled technologies. In
chapter one, I discuss the existing literature on technology commercialization and technology
adoption, how AI is distinct from existing technologies, and why the growth of AI necessitates
firms adapting their technology strategies to fully maximize benefits from this technology. In
chapter two, I examine technology adoption of AI within firms. Specifically, I examine how
manager and employee perceptions of the value of AI differ in firms that have adopted AI and
how this is reflected in actual use of the AI-enabled technology at the firm. I find that manager
and employee alignment about the value of the AI dictates actual usage of the product and, in
turn, renewal of the product, suggesting long-term integration into the firm’s processes. In terms
of firm strategy, these findings suggest that firms should devote significant resources to ensuring
alignment between managers and employees when deploying a new AI-enabled technology, in
order to ensure that the firm captures value from the technology. It also suggests that firm
learning may be the key mechanism by which this alignment is realized. The third chapter
discusses AI commercialization and adoption from the perspective of the innovating firm. I
explore how start-ups with novel technology pursue different strategies for commercialization. In
this third chapter, I develop a framework for understanding variation in successful AI product
commercialization using novel data related to AI-enabled software. As firms begin to use AI in
ways that extend beyond merely deploying a product in the organization, it is important to
understand how the unique characteristics of AI may alter how firms strategize around
technology commercialization and adoption.
1
Introduction
Firms’ choices about technology commercialization reflect firm strategy and dictate how
firms pursue innovation and growth. Firms’ choices about when and how to adopt a technology,
such as through licensing the technology, also dictate how firms innovate and grow. Upon
successful commercialization or adoption, new technologies can provide firms with unexpected
complementarities that, in turn, may result in the firms shifting or altering their technology
strategy. While much work has been done on the topic of technology commercialization and
technology adoption (e.g., Gans and Stern 2003, Pisano and Teece 2007, Teece 1986), I propose
that the increasing proliferation and growth of artificial intelligence (AI)-enabled technologies
1
that are commercialized and adopted by firms dictates a reexamination of the more traditional
notions of technology commercialization and adoption. This reexamination is warranted because
AI-enabled technologies are both product and process technologies, and thus they are distinct in
their ability to learn and improve with use. When a product has an AI capability, such as
machine learning (ML), it can both adapt to and to dictate firm technology strategy, both for the
innovating, inventor firms, and for the firms that license the new technology.
In this dissertation, I examine firms’ technology commercialization strategies and firms’
technology adoption strategies in light of use and proliferation of AI-enabled technologies. In
chapter one, I discuss the existing literature on technology commercialization and technology
adoption, how AI is distinct from existing technologies, and why the growth of AI necessitates
firms adapting their technology strategies to fully maximize benefits from this technology.
1
Artificial intelligence (AI) is an umbrella term for a host of technologies including machine learning (ML), natural
language processing (NLP), and deep learning. In this dissertation, the term AI-enabled technology refers to
technology that uses ML, unless otherwise noted, because ML is currently the frontier technology within the
umbrella of AI. Within this introduction, AI is used to connote broader applicability of the ideas (e.g., understanding
technology adoption in this context likely applies to all AI, not just ML), but within each chapter, ML is used in
order to be more specific.
2
In chapter two, I examine technology adoption of AI within firms. Specifically, I examine
how manager and employee perceptions of the value of AI differ in firms that have adopted AI
and how this is reflected in actual use of the AI-enabled technology at the firm. I find that
manager and employee alignment about the value of the AI dictates actual usage of the product
and, in turn, renewal of the product, suggesting long-term integration into the firm’s processes.
In terms of firm strategy, these findings suggest that firms should devote significant resources to
ensuring alignment between managers and employees when deploying a new AI-enabled
technology, in order to ensure that the firm captures value from the technology. It also suggests
that firm learning may be the key mechanism by which this alignment is realized. These findings
contribute to the existing literature in technology commercialization. The current literature on
technology commercialization discusses how successful technology commercialization depends
on how the consumers value the technology (Lo et al., 2012) and stresses that the market
commercialization of the newly developed technology often contributes to the new technology’s
success or failure in achieving market success and generating profits (Gans and Stern, 2003;
Markham and Lee, 2013). In particular, there are a number of antecedents and factors that
contribute to successful commercialization or to improving commercial outcome of a new
technology. This chapter discusses how micro-level, or individual level factors, can contribute to
a firm’s ability to capture value from a new technology, and thus, the technology’s ultimate
successful commercialization.
The third chapter discusses AI commercialization and adoption from the perspective of
the innovating firm. I explore how start-ups with novel technology pursue different strategies for
commercialization. While the technology commercialization literature typically discusses how
innovating firms can pursue a licensing strategy or a strategy in which they compete in the
3
market (e.g., Hsu 2006), the commercialization of AI-enabled technologies involves slightly
different value considerations given the unique nature of the technology. Specifically, AI is
novel and it can be a GPT, or the input to many other technologies (Cockburn et al. 2018,
Goldfarb et al. 2019), and AI enabled technologies increase in value, to the licensing firm and to
the innovating firm, through use via commercialization. Unlike other most technologies, where
the technological innovation stands on its own, AI is best leveraged as an input to other
technologies. Thus, firms that develop AI enabled technologies need to consider different
strategies for commercialization and licensing if they wish to be successful. And, because it is
critically important that they successfully commercialize and license the products, so that the AI
gets more data and improves its quality, strategies for commercialization are even more salient.
In this third chapter, I develop a framework for understanding variation in successful AI
product commercialization using novel data related to AI-enabled software consisting of
onboarding calls between customer firms and the product developer firm and data on customers’
usage and renewals. I describe four distinct categories of firms that differ according to a
customer firm’s readiness to adopt the product and the specificity of the firm’s knowledge about
the product, technology, and capabilities. I examine the degree to which these variables impact
renewals, product customization, and use, finding that high levels of specificity matters more to
renewals and continued use, whereas a high level of readiness to adopt is more likely to impact
product customization. I contribute to literature on AI commercialization and its impact on firm
strategy.
My hope with this dissertation is to expand the conversation on technology
commercialization and technology adoption in light of AI and associated technologies. As firms
begin to use AI in ways that extend beyond merely deploying a product in the organization, it is
4
important to understand how the unique characteristics of AI may alter how firms strategize
around technology commercialization and adoption. Specifically, because AI can be both a
product and process technology, depending on usage, and can simultaneously be both, it is unlike
other technologies. Additionally, because AI improves over time and with usage, its ability to
add value to a firm (both in terms of commercialization and adoption) can potentially increase at
a rate much faster than that of other technologies. Therefore, expanding our understanding of
technology commercialization and technology adoption in light of AI is important.
5
CHAPTER 1:
Reconciling Technology Commercialization Literature with Artificial Intelligence
INTRODUCTION
Technological innovations are generally considered to increase efficiencies in firms
through a variety of mechanisms, depending on the type of innovation (Chesborough & Teece,
1996). While it has also been recognized that technological innovations may, in certain contexts,
have downsides (e.g., Weigelt & Sarkar 2012), the strategic management literature on innovation
generally views use of a new technology as positive in terms of creating value for firms and
markets (Ahuja, Lampert & Tandon, 2008). As such, firms frequently engage in a process of
trying new technologies. Some firms develop and commercialize technologies for profit and
some firms pay to license these new technologies, incorporating them into the firm to capture
and create value.
Discussion of how a new technology impacts firms and markets rests on the assumption
that both the firms that develop the technology and the firms that use the technology have the
capabilities to implement and utilize the technology to the firm’s advantage. Capabilities refer
to a firm’s ability to achieve new forms of competitive advantage by appropriately adapting,
integrating, and reconfiguring skills and resources, both external and internal, to match the
changing environment (Teece et al., 1997). Firms typically invent and commercialize new
technologies to capture value, and likewise, other firms license and adopt new technologies to
capture additional value from the market (Ahuja et al., 2008).
Isolating the benefits and drawbacks of using a technological innovation and what firms
specifically benefit from a technological innovation can be difficult (Ahuja and Katila, 2001).
Firms that successfully use new technologies may be those firms that already have capabilities to
6
adopt that specific technology or firms that have the tendency – and capability -- to
innovate (Teece et al., 1997). Further, the development and use of new technology
affects multiple aspects of firm operations, making it difficult to isolate the specific benefits that
commercialization and/or adoption of the new technology confers on a firm.
However, understanding the process of how firms adopt new technologies can inform
technology commercialization strategies. Likewise, it is important to understand how firms
commercialize new technologies, in terms of the strategies firms use to encourage use and
integration of the new technology within licensee firms. Further, while a large body of research
has discussed technology commercialization and technology adoption strategies, there remain
open questions about whether existing theory on these topics applies to the development,
commercialization, and use of new technologies such as artificial intelligence (AI).
The technology commercialization literature assumes that firms engaged in technology
commercialization are doing so with the goal of capturing value from the technology in some
way. The value capture assumption applies to both firms producing and developing the
technology, as well as to firms that are engaged in using the commercialized technology. The
ability to capture value from the technology can be via a number of different mechanisms. For
example, the most typical route to value capture from technology commercialization is via
licensing and use. However, value capture can also come from reputational benefits or from
preventing a competitor from licensing the technology (e.g., Somaya, Kim & Vonortas 2011).
Research on innovation and technology adoption in the strategic management
literature tends to focus on firm-specific or market-specific characteristics that influence
innovation and adoption (see e.g., Schumpeter, 1942; Tushman & Anderson, 1986; Cohen &
Levinthal, 1990; Adner & Kapoor, 2016). The individual level of analysis is largely overlooked,
7
with the exception of work on human capital, such as star scientists (e.g., Singh & Fleming,
2010; Agrawal et al., 2014; Grigoriou & Rothaermel, 2014) and knowledge creation
(e.g., Agrawal et al., 2013). On the other hand, there is a substantial body of research in the
organizational behavior and information sciences literature that discusses innovation and
technology adoption at the individual level (see e.g., Waytz et al., 2014; Davis, 1989; Venkatesh
& Davis, 2000). However, work that recognizes the importance of how individual employees’
use of the technology affects firm level adoption, adaptation, and innovation is scarce (Ruttan,
1959). Existing research on technology adoption focuses either at the firm level or at the
individual level; work that accounts for firm structure, hierarchy, and power dynamics, and how
these factors may promote or inhibit technology development and adoption, is also scarce. I
propose that understanding how individual adoption impacts firm level commercialization and,
in particular, adoption, is key to understanding when innovation and technology
adoption is likely to be successful at the firm level, because the factors that drive individual
commercialization and adoption will eventually affect firm adoption.
The disconnect between work on firm and industry level technology commercialization
and adoption and work on individual level technology adoption is especially salient in the
context of AI enabled technologies, which are becoming increasingly widespread in business
(Council of Economic Advisers, 2016). Unlike other technologies, AI is unique in its ability to
improve itself (Agrawal et al., 2017), in its applications (Agrawal et al., 2017), and in its
purpose within a firm (Furman & Seamans, 2019). AI generates more uncertainty than other
technologies, and it does so as a technology that can be commercialized and implemented as a
product or process – or both. With many technologies, there is often uncertainty as to whether
the technology will be successful in the market or whether it will perform as expected. AI is
8
unique in that even if it performs as expected, the implications of this performance for
organizations and individual workers is still uncertain, given the technology’s potential to learn
and improve from more usage. In improving, the technology becomes more accurate and
expands the range of tasks or problems it can perform or assist with. Through this process, which
relies on usage so that the algorithm can learn from more data, the technology’s capabilities
increase. Unpacking the uncertainty in AI adoption at the individual level is critical in order to
ensure that firms are aware of challenges that may arise when commercializing and adopting AI.
In this chapter, I discuss existing literature in strategic management on technology
adoption and commercialization strategies at the firm level and at the individual level. I attempt
to reconcile these literatures with the emerging literature on AI-enabled technologies and their
function, use, and predicted impact. I hope that in doing so this will set the stage for future
scholarly work that develops theories about the nuances of ML technology adoption, particularly
in terms of other industries and its impact on firm and industry structure.
I first review literature on technology commercialization and firm capabilities, focusing
on how individual level mechanisms play a role in shaping firm capabilities for successful
technology commercialization and successful integration of new technology. I also discuss the
literatures on product and process technologies in terms of innovation, and the emerging strategic
management literature on AI and related technologies. Finally, I attempt to reconcile these
literatures and discuss potential contributions to the literatures on technology commercialization
and adoption in light of new technologies such as AI.
Technology Commercialization Strategies
Technology commercialization is “the process of acquiring ideas, augmenting them with
complementary knowledge, developing and manufacturing saleable goods, and selling the goods
9
in a market” (Mitchell and Singh, 1996: 170). It is a process that includes product conception,
definition, design, prototyping and testing, and manufacturing and marketing. Successful
technology commercialization requires firms to have a variety of capabilities from internal and
external sources (e.g., Teece, Pisano and Shuen, 1997). Technology commercialization, when
successful, refers to a firm’s ability to develop and market a large number of product and process
technologies (Zahra and Covin, 1993), including creating new products, expediting these
products to the market, and creating new knowledge (Leonard-Barton, 1995). A firm engages in
technology commercialization with the goal of developing a competitive advantage, though the
sources of this competitive advantage may be diverse. For example, the innovativeness of the
products commercialized, prior commercialization experience, or the speed of commercialization
(e.g., Eisenhardt, 1990) may contribute to competitive advantage in different ways.
Firms face two primary choices in commercializing a technology. Firms can
commercialize via contract, meaning that they license the technology or sell the patent to the
technology. Under this cooperative licensing model, it is assumed that the strength of the
appropriability regime under which the transaction is operating is high, and that there is
relatively low uncertainty about the value of the technology (Arora, Fosfuri, and Gambardella
2004). The other primary commercialization choice available to firms is to commercialize the
technology into products (Conti, Gambardella & Novelli 2019) and compete in the market.
These two strategies are not mutually exclusive. Firms can employ a dynamic
commercialization strategy in which they utilize one method of commercialization at the outset
and then pivot to another (Gans and Hsu 2013). Firms can also pursue a dynamic intermediate
strategy of joint commercialization, in which the innovator contracts with an incumbent firm to
commercialize, but remains involved in the process so as to acquire specialized complementary
10
assets (Hsu & Wakeman, 2013). The conditions of the appropriability regime in which the
commercialization is taking place can affect a firm’s commercialization strategy, as can the
availability of complementary assets (e.g., Teece, 1986; Gans, Hsu, Stern 2002). For example,
commercialization via contracting involves engaging in negotiation, and negotiation necessarily
involves disclosing aspects of the technology. In an appropriability regime where such
disclosures may expose the entrepreneur to an expropriation threat, the entrepreneur may be
more likely to follow a market based competitive strategy. However, where the appropriability
regime would support patent protection for the invention, for example, at a relatively low cost,
the entrepreneur will be more likely to engage in licensing the innovation.
The degree of uncertainty about the technology can affect the firm’s commercialization
strategy (Marx & Hsu 2015). The value of the technology being commercialized must be clearly
signaled and demonstrated in order for the market to function efficiently. However, when a
technology is new, its value may be less certain. Assessing the commercial value of a technology
varies depending on the nature of the technology and its capabilities to be measured or metrics
associated with the technology. However, parties may disagree as to the metrics, and thus the
underlying value, or, given the nature of the innovation it may be less clear how to establish the
innovation’s value. In these situations, where it is difficult to establish the value of the
technology, pursing a tech commercialization strategy that relies on cooperative negotiation with
a contracting partner to license or sell the patent to a technology may not be feasible. In these
scenarios, it may be too costly to negotiate to determine a mutually agreed upon value for the
technology, both in terms of the time spent negotiating and the potential for giving away
valuable information about the technology during licensing negotiations. Even despite potential
remedies to the contracting problem presented in a scenario with asymmetric information about
11
the technology, there are still many impracticalities that exist to support a firm pursuing this
technology commercialization strategy in a scenario with uncertainty.
Thus the protections afforded against expropriation by the appropriability regime and
how easily the technology’s value can be quantified or measured determine which technology
commercialization strategy firms will use at the outset.
Technology Commercialization and Firm Capabilities
Technology commercialization success depends on the firm’s capabilities and on external
sources (Teece, 1986). Knowledge that the firm acquires from internal and external sources
shapes the firm’s capabilities, and, in turn the firm’s strategic planning, choices, and processes
(Grant, 1998). The firm’s strategic decision making, planning, and processes also affect how
likely it is that the firm can successfully integrate internal and external sources of capabilities,
which dictates technology commercialization success.
A firm’s capabilities for technology commercialization evolve over time based on
experience with commercialization. A firm can develop these capabilities from engaging in joint
commercialization with an experienced incumbent, for example, or based on experience in prior
commercialization (Hsu & Wakeman, 2013). Internal and external capability sources within a
firm, both within the commercializing firm and the adopting firm, contribute to successful
technology commercialization (e.g., Eisenhardt and Tabrizi, 1995). However, there is little
empirical evidence to support this point. In general, the literature does not fully address the
sources of capabilities for firms’ successful technology commercialization or how internal and
external capabilities sources integrate for successful technology commercialization (Zahra &
Nielsen, 2002). An exception is the stream of research that focuses on how capabilities are
integrated within internal units within the firm (Song et al., 1997; Barney, 1999). As firm
capabilities improve and as firms’ capabilities for integrating external and internal capabilities
12
improve, firms are increasingly able to influence technological trajectories. Implicit in this notion
is that firms want to improve a given technology’s performance, meaning that firms want to
develop capabilities to facilitate the technology’s improved performance such that it can allow
them to capture or create more value. In order to do this, firms must be able to first assess the
likely trajectory of the new technology and then leverage existing capabilities and develop new
capabilities to be able to shape the trajectory and capture value from the technology. If managers,
who are the decision makers at the firm, do not understand how employees are using the
technology, it will be difficult to accurately integrate knowledge about the technology and thus
assess the technology trajectory.
A firm’s capabilities can evolve based on changes in technological trajectories. A
technology’s trajectory is the path it follows over time, depending on existing scientific
knowledge about the technology and improvements to the technology’s performance (Dosi,
1982, 1984; Winter, 1984). As the technology evolves, firms may find that their existing
capabilities and the capabilities required for successful technology commercialization do not
match. While a strong manufacturing capability, according to the research based view, can
enable firms to adapt to evolving technology and maintain successful technology
commercialization, this finding does not hold in all scenarios and technologies. For example,
external market factors can affect the availability of external resources, facilitating or inhibiting a
firm’s ability to leverage and adapt its existing capabilities (Dierickx and Cool, 1989).
I propose that the nature of the technology itself may promote or inhibit firms’ adaptation
to the technology and thus the firms’ ability to successfully engage in technology
commercialization. Specifically, if a technology, such as AI, can be a product, process, or both,
then firms may not be able to anticipate or predict the technology’s trajectory. In turn, because
13
the firms are unable to anticipate the technology’s trajectory, firms will not be able to adapt,
including develop new capabilities, in order to be able to capture value from the technology.
Unpacking this further, the uncertainty associated the nature of a technology (i.e.,
whether it is a product or a process) and thus with its anticipated trajectory can be problematic
for firms in other ways. For example, firms may adapt to the anticipated technology trajectory
but may fail to predict the full potential of the technology. For example, in terms of AI, firms
may adapt to natural language processing as the frontier of the technology’s trajectory, and
commercialize products using this technology, only to have this technology supplanted by
further advances in machine learning (ML) that may require different resources and capabilities.
Relatedly, firms may make specific investments in the technology, such as data resources, that
constrain future technology development and affect commercialization success. For example, if a
firm exhausts considerable resources in building a data gathering capability, only to have the
technology evolve such that the amount of data gathered is no longer useful to the technology
development, this may have a negative effect on commercialization. However, in taking these
actions to commercialize the technology, based on the anticipated trajectory and the firms’
capabilities, firms that commercialize the technology may have a role in shaping the technology
trajectory; that is to say that the process is endogenous.
Integration of knowledge via firms’ capabilities involves the macro and microlevels.
However, while there is a substantial, and growing, body of research that discusses firm
capabilities at the micro level (e.g., Felin, Foss & Ployhart, 2015; Helfat & Peteraf, 2015) and
much research that discusses firm capabilities at the firm level, the integration of these micro and
macro level factors is often overlooked. Additionally, the interplay of micro and macro level
14
factors within the firm and in terms of integrating outside firms within the firm (especially in
terms of tech commercialization) has not fully been explored in the literature.
A firm’s knowledge creation capability enhances value creation. Firms are heterogeneous
in terms of their willingness to formalize knowledge integration activities. The process of
formalizing knowledge integration has been examined from a variety of different theoretical
perspectives. The knowledge based view of the firm is based integrating micro-level employee
and manager learnings to help drive strategy at the firm level. Similarly, literatures on
organization learning and knowledge management discuss the integration of micro and macro-
level perspectives on knowledge within the boundaries of the firm to help create value. The
discussion of organization boundaries and technology commercialization relies on the focal
firm’s capabilities, which often involve individual level capabilities and how those translate to
firm level capabilities. The mechanism by which this translation from micro to macro levels
happens is via knowledge integration. Management, which helps facilitate knowledge
integration, has been considered to be a technology that firms can invest in and develop (Bloom,
Sadun & Van Reenen, 2016). Some firms then, are better at management than others.
Unsurprisingly, there is heterogeneity in the commercialization processes, because some firms
are better at management than others, and some firms are better at knowledge integration than
others. Sources of variation amongst firms’ abilities to successfully commercialize implies that
formal processes by which knowledge is integrated within the firm could help firms consistently
commercialize different technologies.
New knowledge must be processed in order to be useful to the firm in creating value.
Firms’ abilities to process and integrate the knowledge depends on the volume and complexity of
the knowledge. Additionally, the ability to process and integrate knowledge relies critically on
15
the relationships within the firm and how employees interact with each other and with managers.
Knowledge transfer within the firm helps facilitate new knowledge processing and deployment,
which in turn can facilitate successful technology commercialization and adoption.
Individual factors
Within the firm that is adopting the commercialized technology, individual level factors,
such as the status of the employees who are actually using the technology, can affect the
technology commercialization success. From the agency theory perspective, there are a number
of reasons that these micro-level dynamics exist within the firm. Most firms are by nature
hierarchical, and thus principal agent incentives problems exist within firms (Eisenhardt 1989).
Managers often make the decision to adopt a new technology (e.g., Tucker 2008). However,
employees are often the ones using this technology. While managers may understand the value
that a new technology is poised to add, the managers may not be able to successfully translate or
integrate this information to their employees. Knowledge integration across hierarchical levels
(e.g., manager to employee) within the firm is especially difficult. Trust (Hart & Saunders 1997),
capabilities (Grant 1996), and specializations or expertise (Parmigiani & Mitchell 2009) have all
been shown to augment the process of knowledge integration across levels within the firm (e.g.,
Coff & Kryscynski, 2011; Felin, Foss & Ployhart 2015). Ideally, the firm is able to successfully
integrate the new technology into its processes. Successful knowledge integration allows the firm
to capture and create value from the new technology. Knowledge integration involves the firm
leveraging its capabilities to ensure that it is able to capture value from new technologies (Zahra,
Neubaum & Hayton 2020). Firms then leverage their capabilities to integrate new technologies
into firm processes.
16
In the firm commercializing the technology, individual level perspectives on knowledge
integration involve cognitions, skills, experience, and aspirations that can affect how knowledge
is processed within the firm. Similar to within firms using the technology, multiple levels of the
organization in the commercializing firm must align their capabilities and processes in order to
successfully commercialize the technology. First, organizations must ensure that different levels
are aligned as to the goal of technology commercialization. Then, organizations must facilitate
processes for the different hierarchical divisions to interact.
Firm Level Factors
While individuals carry out important roles in knowledge integration, especially in terms
of their individual cognitive biases, knowledge, and capabilities, the role of group-oriented
knowledge integration also is a critical factor in ensuring successful technology
commercialization. Managers in groups can set the stage for how a technology
commercialization is initiated within the firm. In turn, the group attitudes can influence
individual employees’ perceptions and attitudes towards the technology (Qian, Agarwal, &
Hoetker, 2012).
Business models that facilitate tech commercialization emphasize the adaptability of the
technology. For example, some firms may develop general purpose technologies (GPTs) in order
to license these GPTs to other downstream firms with specializations in particular aspects of the
commercialization value chain (Gambardella and McGahan, 2010). The adaptability of the GPTs
to different settings and uses can accommodate the differences in strategies that downstream
firms may employ. The commercializing firm is able to capture more value from the technology
because it can be used by a larger group of downstream firms.
17
The speed and quality of a firm’s knowledge integration of a new technology, whether it
is a technology that the firm is engaged in commercializing or whether the firm is benefiting
from this commercialization, determines the success of the technology commercialization.
Micro-level interactions within the firms influences this process, as individual cognitions, skills
and experiences can significantly influence knowledge integration. However, bridging the gap
between how these micro level processes and factors that facilitate knowledge integration affect
value creation is largely left unanswered by the existing literature.
Technology Adoption
There are a number of different frameworks and models that discuss technology adoption
at the individual and firm levels and from a variety of different theoretical perspectives, drawing
from, for example, the information science (IS) literature and the psychology literature. In the
strategic management literature, technology adoption is treated as a dichotomous variable. That
is, either the firm or individual adopts (here, uses) the technology, or not (e.g., Hannan &
McDowell, 1984). In many empirical contexts, technology adoption is a proxy for use, meaning
that when a firm has licensed the technology it is considered to be using the technology (Priem,
Li & Carr, 2011). Technology adoption has been examined from macro and micro economic
perspectives (e.g., Parente & Prescott 1994) and at firm and individual levels, however there is
little work that reconciles firm and individual level technology adoption.
Technology adoption is examined extensively in strategic management, though often
through related topics that address the drivers of technology adoption (e.g., strategic alliances;
licensing; competition) and the impact of adoption (e.g., technology commercialization;
capabilities). Research has examined the role of contractual completeness in technology
adoption, finding that greater incompleteness results in the adoption of less advanced
18
technologies (Acemoglu, Andtras & Helpman 2007). Additionally, a number of papers address
the barriers and enablers to technology adoption, especially in the manufacturing technology
sector (Stornelli, Ozcan & Simms, 2021). There are few studies that directly estimate the returns
to profits arising from increased use or adoption of new technologies. The direct returns to profit
from technology adoption are difficult to measure because a large amount of data is needed and
profit, though conceptually straightforward, can be difficult to measure. Additionally, it is
difficult to isolate the direct impact of the technology on profits without other firm variables
confounding the analysis.
Summary: Technology Commercialization and Adoption
Firm level factors, such as capabilities and knowledge integration determine the success
or failure of a technology commercialization strategy for a firm that is the inventor of a
technology and for firms that hope to profit off the technology. Research in strategic
management thus far has focused on the technology commercialization strategies of firms that
are designing and inventing the technology. However, firms that intend to license and use the
new technologies do so with a specific strategy in mind. While the technology commercialization
literature focuses on firms inventing the technologies, there is a body of literature in other
disciplines such as organization theory, psychology, and information science that also discusses
the use and adoption of new technologies by firms, and the strategies that these firms use to
facilitate successful integration, use, and value capture from new technologies..
Firms that invent a technology and then seek to capture value from its commercial
applications have a choice in strategies for how to pursue commercialization. Firms can pursue
commercialization of their technology via contracting, whether via licensing or patenting.
Alternatively, firms can choose to compete in the market. These technology commercialization
19
strategies do not have to be exclusive; firms in some environments, given threats from
expropriation, for example, may benefit from first competing in the market and then moving to a
cooperative technology commercialization strategy as the value of the technology and/or the
external environment (e.g., expropriation concerns) become more certain (Marx & Hsu 2015).
Firms that want to adopt and use a new technology, yet do not create or invent that
technology, may do so with the goal of commercialization. While these firms do not capture
value from direct commercialization, meaning from invention to market or licensing the
technology, they nonetheless attempt to capture value from the technology. The primary strategy
for value capture amongst firms that license or use (adopt) the new technology is via usage..
Firms may also license a technology or purchase a patent to prevent a competitor from doing so.
Further, firms may also license a technology for reputational benefits, such as signaling that the
firm is technologically advanced.
For both firms that invent and firms that license new technologies, the technology
commercialization strategy used determines how likely it is that the technology will actually be
used by downstream firms. The inventor firms have an interest in promoting downstream use and
adoption of the technology, whether through licensing or through competing in the market and
having the technology gain dominance. Depending on the nature of the technology, the firms that
use the technology may benefit from a closer or more distant relationship with the inventor of the
technology. For example, many technology products licensed to firms that involve large amounts
of the licensor’s data, such as Salesforce, require significant and continued oversight and
assistance from the licensing firm. Extending the Salesforce example, in many instances
Salesforce teams will embed within the licensor firm to help with successful use and adoption of
the technology, raising questions about where the boundaries of the firm lie for some
20
technologies and their use. While some technological inventions can be easily commercialized in
a product, others require more complementary inputs and infrastructure in order to capture value,
and thus the commercialization strategy is slightly different.
Product and Process Innovation
The literature on technology management and the technology life cycle describes the
dynamics of product and process innovation (Utterback and Abernathy 1975; Clark 1985;
Klepper 1986). The underlying premise is that initially, product design of new technology is
fluid and varied. At this stage, product innovation is the dominant form of innovation and this
product innovation attempts to improve performance. As different product innovations occur, a
dominant design eventually emerges and the “optimal” product configuration is reached. Process
innovation, on the other hand, is initially minor; early production processes occur at a small
scale, are flexible, and occur in highly skilled labor sectors. As the product stabilizes and a
dominant design emerges, automated production is increasingly used and process innovation
becomes dominant as firms seek to lower costs. Both product and process innovation ultimately
slows. The dominant design in product and process innovation focuses on the reduction of
uncertainty that allows firms to make technology specific investments.
As the product life cycle evolves, the technological innovations change from product to
process developments, because innovating firms are less able to appropriate returns from their
investments in innovation (Klepper 1996). As an industry matures and the innovating firms get
bigger, firms then have increasing incentives to pursue process innovations (Henderson 1995).
Similarly, because product innovations are related to acquiring new customers, rather than the
magnitude of the existing customer base, the return to product innovation decreases over time
(Christiansen 1992). The rise in returns to process innovations and the decreasing returns to
21
product innovations over time are largely exogenous, especially in contexts where product
innovation occurs before process innovations. Product innovations are largely determined by the
innovating firm’s expected profit from the new product and the firm adopting the product’s
expected value capture from deploying the product (Boone 2000). Process innovation, on the
other hand, is aimed at reducing the adopting firm and the innovating firm’s costs. For the
adopting firm, this is a function of the firm’s profit with respect to its own costs; for the
innovating firm, this may vary depending on the technology commercialization strategy.
Product and Process Technologies and Technology Commercialization
An industry’s evolution from product to process innovation for a given technology
reflects the strategic decisions of firms in that industry. Technology commercialization decisions
are made at the firm level based on the firm’s expected value capture from a given strategy. As
discussed above, the firm weighs which commercialization strategy to pursue based on a number
of factors, including the external environment (Adner and Kapoor, 2010). However, the
technology’s projected trajectory also makes a difference. As more is known about the
technology, this will affect the degree of uncertainty associated with one commercialization
strategy or another. Firms can adopt a dynamic commercialization strategy that allows them to
switch from a market-based competitive strategy to one where they license the technology, for
example. However this switching fails to fully account for firms’ understanding and projecting
the technology trajectory of the innovation. Entrepreneurs’ understanding of the technology S-
curve is endogenous to the strategies they employ for commercializing and developing their
products (Gans, Kearney, Scott & Stern 2021).
ARTIFICIAL INTELLIGENCE
22
The innovation literature discusses product and process innovations as related to
technologies (products) and inputting and manipulating these technologies in existing systemic
frameworks (Damanpour & Gopalakrishnan 2001). Typically, an innovation is either a product
or a process technology, but not both. Innovations can help spur the development of a product
technology, and in turn, at a later stage of the technology S-curve, can help improve processes.
However, these innovations with dual purposes are rare.
In the case of artificial intelligence (AI), the technology’s unique attributes may provide
more information as to how to predict its trajectory and optimal strategies for firms’ technology
commercialization efforts involving AI products. AI is an umbrella term that encompasses a
large number of technologies including big data, natural language processing, machine learning,
and neural networks, for example. Use of AI throughout this work includes all the associated
technologies that fall under the AI umbrella.
There is much recent research about AI and its impact. AI is a general purpose
technology (GPT) (Cockburn et al. 2018, Goldfarb et al. 2019) meaning that it is used as the
input for a large number of technologies, and that co-invention and coordination costs are
associated with its use. AI is unique in its ability to learn and improve with more usage. More
usage of AI allows it to collect more data, and by processing this data, the algorithm improves its
predictions. Thus, with enough data covering a broad enough range, AI predictions can (in its
current iteration) approximate the predictions that humans make and, in many cases, can improve
on these predictions because of the amount and quality of the data used. Unlike other
technologies, AI’s ability to learn from more use means that the value that AI adds to a firm
critically relies on usage. For example, in many legal technology applications of AI (specifically
machine learning) for contracts, the data that the algorithm trains on is limited by the availability
23
of contracts to use as training data. Much of the training data for these machine learning (ML)
tools comes from publicly available sources. However, this introduces bias in the algorithm’s
results and limits its ability to learn about different clauses and scenarios that may arise in a
contracting exchange. The reliance on a large set of data varies depending on the required
robustness of the output (predictions), however in most AI applications, a large set of data is
required. Thus for firms commercializing AI technology, it is beneficial not just to sell the
technology but to also have a mechanism by which the users’ data can improve the technology
for the users who have already licensed the tech and for future potential users.
However, the technology trajectory of AI remains somewhat unclear. Despite AI having
been around since the 1950s, recent advances in data processing and storage capabilities, as well
as access to additional data, has spurred AI growth. Technological trajectories are structured yet
evolutionary processes by which emerging technologies change and improve over time (Dosi
1982), with implications for the distribution of value from innovation (Teece 1986), the
coevolution of industries and competition (Utterback 1994), and the choice of commercialization
strategy (Teece 1986, Gans and Stern 2003). The technological trajectory framework, and related
frameworks (e.g., the concept of the technology S-curve (Foster 1986)), propose that even in the
face of uncertainty, the evaluation of technology follows a structured process. The impact of a
technological trajectory depends on more than the technology’s performance (Adner and Kapoor
2016), however the ability of firms, whether start-ups or incumbents, to commercialize the
technology also critically affects the technology’s trajectory.
A firm’s strategic choice in commercialization then may not simply have implications for
its ability to capture value from the technology, but also for how the technology itself develops
and whether it is impactful. As research on the technology S-curve discusses, managers can
24
harness the evolution of technology for competitive advantage, provided they can accurately
gauge where on the S-curve the technology is at that time. By making investments into the
technology, especially the investments required to facilitate the utility of a GPT such as AI, firms
may influence the technology trajectory by prioritizing certain aspects of the technology over
others. By investing in the technology either as an inventor or as a licensee, firms contribute to
what aspects of the technology get prioritized by firms that develop the technology.
TECHNOLOGY COMMERCIALIZATION STRATEGIES FOR AI
Successful innovation requires investments in exploration and exploitation (March 1991;
Gans et al. 2021). Firms’ successful technology commercialization strategies also consider
exploration and exploitation, but in terms of the technology’s S-curve. Firms that have invented a
technology and that seek to license it or to compete in the market pursue one strategy or the other
based, in part, on where they judge the technology to lie on its S-curve. Similarly, firms choose
to license a technology based on where they anticipate it to be on the technology S-curve and
how that relates to their intended strategy for value capture from that technology.
For inventor firms that have developed an AI technology, licensing may be preferable to
competing in the market. Because AI improves with use, licensing the technology allows for the
inventor firms to contract with licensees. Contracts that govern these transactions can articulate
how control rights over the data used in the technology will be apportioned, so that both parties
can benefit from the AI learning from new data. Additionally, licensing allows both firms to
potentially impact the technology trajectory as they use it.
Competing in the market, on the other hand, may yield profit in the short term but may
ultimately constrain the growth of the technology. If the firm competes in the market with its
technology then it may adapt the technology to market demand, which may serve to flatten the
25
growth as competitive forces act upon the inventor firms, ultimately resulting in less investment
in the technology’s development.
For firms that want to benefit from using AI, understanding how the micro-level, or the
employees at the adopting firms, interact with the new technology is key. Employees are actually
using the technology, and thus directly contributing to the degree to which the firm captures
value from the technology. However because managers hold the decision rights for licensing and
renewal, the interactions between employees and managers and how information and knowledge
flow within the firm can help predict how successful firms will be at adopting and using the new
technology, particularly AI.
Because AI’s value is tied to use, and use of new technology occurs at the employee
level, it is critical to understand the factors that promote and inhibit employees’ use of AI. While
the technology adoption literature in strategic management focuses on the macro level and the
micro level separately, the proliferation of AI technologies necessitates an expansion of this
literature.
26
CONNECTING CHAPTERS ONE AND TWO
The literature on technology commercialization focuses on the perspective of the firm
that is inventing and bringing the innovation to market. Though technology commercialization
involves the firms that will be purchasing or using the product, the technology commercialization
literature tends to be relatively agnostic as to the actions and characteristics of firms that will be
using and adopting the technology once commercialized. On the other hand, the literature on
technology adoption focuses on the firms that will be using and integrating the new technology,
while remaining relatively agnostic on how commercialization strategies may affect the adoption
process.
However, it is likely that the processes of technology commercialization and technology
adoption are interdependent. A firm’s commercialization efforts for a new technology, for
example, may be impacted by whether or not the firms likely to use the new technology actually
do so. Similarly, firms’ demand for technology may shape what technologies are successfully
commercialized.
In the first chapter, I explore the literatures on technology commercialization and
adoption and attempt to reconcile and discuss them both. I pay particular attention to how these
two literatures may address the trajectories of new technologies, such as AI, that function as both
a product and a process. Because AI functions as both a product and a process, the strategy for
commercializing AI relies on usage. Thus, understanding the interaction between
commercialization and usage of the technology is critical to its continued development.
Thus, in the second chapter, I attempt to begin to explore the interdependence between
firms focused on developing and commercializing new technology and firms that are using
(adopting) this technology. To do so, I examine the process of technology use and adoption
within a firm that has licensed AI-enabled technology from a commercializing firm. I explore
what factors promote and inhibit sustained use of the new technology by employees such that the
managers elect to renew the technology license. In doing so, I discuss the dynamic between
27
managers and employees, and how manager and employee alignment between with regard to the
value of the technology can result in sustained use of the technology over time. By exploring the
individual level factors that affect technology adoption, I hope to add to the conversation on
technology commercialization and technology adoption strategies given technologies where the
value that the technology adds increases with its sustained usage. In these situations, such as with
AI enabled technologies in particular, the interaction between technology commercialization and
adoption, and, more specifically between firms developing and commercializing technology and
firms licensing the technology, can yield insights about how the technology is likely to develop
and how firms are likely to use it to capture and create value.
28
CHAPTER TWO: Technology Adoption and Commercialization Strategies in a Machine
Learning World: Examining the Use of Automated Contracts
ABSTRACT
Firms’ transitions from existing technologies to new technologies are often challenging and
incremental. Machine learning (ML) is distinct from many technologies in that it is a general
purpose technology (GPT) and thus co-invention and coordination costs are associated with its
use. ML also reduces the cost of prediction. However, the growth and diffusion of ML can be
inhibited by individual level factors such as user uncertainty and lack of incentives to use the
technology. To explore the endogenous process of ML adoption within a firm and how to
motivate employees’ ML use, I use a unique data set from a company with ML contract
automation technology that includes eighty onboarding calls from thirty customer firms,
employee usage data, customer firm renewal data, and interviews. Based on topic modeling
analysis of the call transcripts, I develop propositions that discuss how focusing on the long-term
value added to the employees’ jobs and to the firm, instead of on the short term efficiency gains
to a specific task, can motivate employees to use ML. To ensure that firms use and capture value
from new technologies, I propose that managers should focus on the long-term value added to
the firm and to the employees’ jobs when certain tasks are automated, rather than focusing on
short-term efficiency gains from automating one task. I also theorize about how employees’
orientations towards ML can change by learning more about the technology and when
employees’ and managers’ orientations towards ML are aligned. I conduct preliminary tests of
the propositions and find support.
“Our problem is that once we close a sale and go through the whole onboarding process, the
customer – their employees – fail to actually use the product.”
– former Chief Marketing Officer, LawGeex
Introduction
In an attempt to maintain a competitive advantage, firms continuously evaluate new
technologies and business models. The decision of whether to license or to develop a new
technology in-house (Kapoor and Adner 2012, Williamson 1975), the decision of how to allocate
ownership rights to innovation (Lerner and Merges 1998, Magelssen 2020), and the role of firm
resources in achieving competitive advantage (Barney 1991) are, separately and jointly, the
subjects of substantial bodies of research that help explain the pursuit of competitive advantage.
Yet once a firm licenses a new technology or adopts a new business model, whether or not that
decision creates value depends on the firm’s organizational structure and routines (Argyres and
29
Silverman 2004, Eisenhardt and Martin 2000). A firm’s organizational structure and routines
depend largely on its resources and capabilities to motivate structural change (Acs and Audretsch
1987, Galbraith 1956, Schumpeter 1942), including the firm’s ability to adapt (Hill and
Rothaermel 2003, Teece et al. 1997) and to learn (Mayer and Argyres 2004, Milgrom and
Roberts 1992).
The decision to license a new technology typically occurs at the managerial level rather
than at the level of the individual (employee) user (Anderson and Joglekar 2005, Chao and
Kavadias 2008, Loch and Tapper 2002). However, individual employees may have unique
motivations for using or not using the technology once it has been licensed (Hall and Khan
2002). Thus, the manager’s ability to adequately motivate the employee will influence the degree
to which the technology will or will not be used. Here, I refer to new technology as any
commercialized technological innovation that allows the focal firm (licensee) to capture value
that accrues from use of the technology.
In the case of technologies that incorporate machine learning (ML), the manager’s ability
to incentivize employees to use the ML technology is critical to determining whether the firm
will be able to capture as much value as possible from the technology. This is because ML,
unlike many other technologies, improves its capabilities with more data (more use). To put it
differently, ML is a general purpose technology (GPT) and thus there are coordination costs
involved in its use (Cockburn et al. 2018). Additionally, ML, like most technologies, reduces the
cost of something – here, prediction. In reducing the cost of prediction, ML can be a substitute
for prediction tasks; ML can also be a complement to other high-skill tasks involving judgment,
for example (Agrawal et al. 2018). The more the employees use the ML, the more value the firm
will capture from the ML, relative to use.
30
Therefore, the question becomes how managers should structure incentive schemes to
motivate employees to use ML technology. There is an extensive literature on how to motivate
employees to innovate (Aghion et al. 2013, Manso 2011) across a number of different firm and
industry contexts (e.g., Terwiesch and Xu 2008; Galasso and Simcoe 2011). This literature
distinguishes between pay-for-performance schemes used to induce effort and incentives
schemes that account for how innovation can be nonlinear. In addition, the agency theory
literature more broadly also discusses how to optimally incentivize agents (Eisenhardt 1989,
Gibbons and Murphy 1992), including how to do so in environments with varying degrees of
information (Weigelt and Sarkar 2009), and across a number of tasks of varying difficulty
(Kaarboe and Olsen 2006, Siemsen 2008). There is also an extensive literature on how
capabilities shape a firm’s ability to capture value from a new technology (Chao et al. 2009,
Rothaermel and Hess 2007) and how entrepreneurs can facilitate commercialization and
adoption of their technological innovations (Karp 2020, Kirtley and O’Mahony 2020).
In this paper, I explore how managers can shape incentive schemes for using ML, how
employees learn about the ML technology, and how this impacts whether the firm is successful
in adopting and integrating the ML technology into the firm. Innovative technologies that help a
firm maintain a competitive advantage may initially appear, to the firm’s employees, to be
licensed at the employees’ expense (Acemoglu and Autor 2010, Acemoglu and Restrepo 2019).
Often, new automation technologies, such as ML, have the potential to replace some tasks,
rendering the employees redundant (Acemoglu and Restrepo 2018a, 2019). Anticipating this
task-specific substitution effect upon use of the new technology, employees may not be
motivated to learn about the new technology or to use it.
31
I find that when employees focus on long-term, rather than short-term, benefits of the
new technology use, and when there is alignment between employees and managers about the
benefits of the technology, it is more likely that the firm will use the ML and more likely that the
firm will renew its license for the ML. To do this, I use proprietary data from LawGeex, a
company with a machine learning (ML) contract automation tool, which includes eighty
onboarding calls for thirty customer firms, customer (individual user) usage data, customer firm
renewal data, customer firms’ user data, and interviews with LawGeex executives. Given the
robust nature of this data, I am able to explore the mechanisms that explain why some firms
successfully capture value from ML (i.e., use the ML more and then renew the product license)
while others are unable to do so. I conduct a topic modeling analysis on the onboarding call
transcripts. Based on the results of the topic modeling analysis, I develop propositions to explain
how focus on long-term, rather than short-term, value added for the employees can motivate
employees to use ML, based on employees’ preconceived notions of ML and their jobs, and how
employees’ motivations to use ML can change over time. Then, I conduct a preliminary test of
the propositions using employee use and firm renewal data.
I show, first, that fear of job loss from automation contributes employees failing to use
the ML; second, that a way to overcome this barrier to use is to emphasize the long-term, rather
than short-term, benefits to using ML; and third, that while learning about the more technical
aspects of ML can increase long-term ML use, learning about the commercialized product itself
does not increase long-term use. Additionally, heterogeneities in terms of managers’ and
employees’ initial motivations for using the technology may provide insights into firms’
likelihood of long-term use and renewal.
32
This paper has implications for how firms can address incentive issues that arise in
motivating employees to use a new technological innovation, in particular machine learning. The
literature on incentive mechanisms discusses the interplay of implicit and explicit incentives in
resource allocation decisions (Gibbons and Murphy 1992, Grossman and Hart 1983, Murphy
1992), multi-task settings (Kaarboe and Olsen 2006), and multiple innovation initiatives (Chao et
al. 2009). There is also an extensive literature on the nature of formal versus real authority within
the firm (Aghion and Tirole 1997, Baker etl. al. 1999). However, there is little work that directly
connects how non-pecuniary employee incentives affect the value that a firm can extract from a
technological innovation. These incentive issues are particularly salient for ML technologies
given the nature of the ML. If firms do not capture value from ML, they will not continue to use
it, and the underlying ML algorithm will fail to capture the data needed to improve. This paper
also explores issues related to nuances of firms’ technology adoption and value capture from
ML, and how non-pecuniary employee incentives and firm capabilities can affect this process.
Licensing a technology, at least from an empirical perspective, is often considered to be
sufficient to satisfy the condition of adopting the technology (Gans and Stern 2003). This paper
presents a more detailed view into how merely licensing ML may not be sufficient to be ensure
that a firm will capture value from the ML technology, especially if the firm largely fails to use
it. In some contexts, this firm would be considered to have adopted the technology, yet a
dichotomous adoption measure does not fully capture whether and how the technology is likely
to affect firm strategy. I propose that, for ML in particular, measuring adoption and diffusion
should look beyond mere acquisition and should measure how the technology impacts, or is
likely to impact, firm performance and strategy, based on whether and to what degree it is used
33
within the firm. A more dynamic measure of technology use allows for better predictions about
how the technology will affect firm strategy and value creation.
Machine Learning Reduces the Cost of Prediction
Machine learning (ML) uses large amounts of data to predict outcomes. Typically, ML
predicts more accurately than humans and with much fewer costs associated with making the
prediction. Therefore, it can be said that ML reduces the cost of prediction. Applications of ML
can be framed as prediction problems, that is, ML can solve problems or queries in which one
attempts to predict an outcome. Thinking about ML as a prediction technology helps better
anticipate its potential impact (Agrawal et al. 2018, Agrawal et al. 2019).
As with most technologies that reduce the cost of something, ML reduces the cost of
prediction. In turn, the value of substitutes to prediction decreases while the value of
complements to prediction increases (Acemoglu and Restrepo 2018a). For example, in the legal
context (the subject of this paper), an attorney’s job is largely to make predictions about, for
example, the outcome of a case or the implications from drafting a contract a certain way. Nearly
every aspect of legal practice, whether it is the work of a judge, an attorney, or a legal support
professional, relies on prediction at its core. Thus, ML is poised to have a large impact on the
legal industry.
In this paper, I examine in particular a legal technology that uses ML to review contracts.
Reviewing contracts is a time-intensive, detail oriented task for attorneys and legal professionals
that, despite the time and attention required, is often very routine. Contracts must be reviewed to
ensure (relative) completeness, clarity, and that the client’s interests are met. The attorney or
legal professional reviewing the contract must predict how to create the best contract in order to
best serve the client’s interests. Using ML to review contracts will reduce the value of substitutes
34
to ML contract review automation (i.e., lawyers and legal professionals) and increase the value
of complements to ML contract review automation (i.e., lawyers and legal professionals).
Therefore use of ML contract review technology is likely to change, rather than substitute, the
tasks that lawyers and legal professionals perform during the course of their work. In fact, most
professions that involve high-skilled labor are likely to be similarly impacted in the scope of
their jobs by increasing use of ML that performs prediction tasks (Acemoglu and Restrepo
2018a).
Legal Tech and LawGeex: Context, Company, and Data
Because law relies heavily on prediction, use of ML in law is growing rapidly. Indeed,
investments in the “legal tech” industry have grown over 700% from 2017 to 2019 (Pivovarov
2019) and legal tech start-ups continue to receive venture funding at relatively high rates
(McEvoy 2019). Legal tech start-ups traditionally tend to focus on a few different areas: legal
research and analytics, practice management, and contracting. Within the contracting area of
legal tech, start-ups typically have products that perform contract management, review, or
drafting functions (Rich 2018).
2
Contracts govern transactions and are therefore central to firm
operations and strategy. Additionally, because many contracts are relatively consistent and
routine, contracts present a prime opportunity to use ML to predict what clauses should and
should not be in the contracts.
Indeed, there are many start-ups that use ML for contract review. One preeminent ML
contract review automation start-up is LawGeex. LawGeex was founded in 2015 and uses a
proprietary ML algorithm to review contracts such as non-disclosure agreements (NDAs), sales
agreements, and service agreements. Their customers are primarily organizations, not law firms.
2
A few new start-ups (e.g., Pactum) are focusing on negotiations as well.
35
LawGeex provided access to much of its internal customer data. The data used for this
paper includes the number, type, and date of contract uploads (usage), customer firm renewal
rates, and eighty onboarding calls with customers. Additionally, I conducted two LawGeex
customer surveys and multiple interviews with LawGeex managers. Further, LawGeex also
provided data on its customer firms as well as on firms that were considered “closed-lost” firms,
or those firms that came close to licensing LawGeex’s product but ultimately chose not to
(reasons were specified).
Empirical work on ML in high-skilled industries such as law is still relatively scant as
access to data is difficult and the technology is still relatively new. Additionally, there are a lack
of empirical studies on the legal technology industry for similar reasons. However, the data from
LawGeex on customers and firms that were almost customers allows for a limited, descriptive
picture of the emerging legal technology industry. The analysis herein in this first part of the
paper, while limited in scope to one ML contract automation firm, provides additional context
about characteristics of firms that license legal ML as compared to those firms that come close to
licensing but ultimately do not.
3
LawGeex Customers versus Almost-Customers
To examine the differences between LawGeex customers (adopters) and the almost-
customer firms (non-adopters), I compared the groups across variables related to firm size, such
as number of employees, EBITDA, and total revenue, because early adopters typically tend to be
smaller firms than later adopters (Christensen 1992, Foster 1986). Early adopters also tend to be
firms that will realize a more immediate productivity gain from the new technology due to firm
3
I use proprietary data from the Association of Corporate Counsel’s Chief Legal Officer survey to confirm that
firms that LawGeex considered “non-adopters” did not adopt a similar ML contract review product in the year
following their designation as a “non-adopter”.
36
capabilities and resources (Adner 2002, Henderson and Clark 1990). Therefore, I also analyzed
the number of lawyers at each firm and the number of lawsuits in each firm.
The firms designated as “adopters” were those firms that licensed LawGeex’s product for
at least a year. The firms designated as “non-adopters” were those firms that engaged in the
LawGeex sales cycle but ultimately decided not to pay to license LawGeex’s product. LawGeex
had well over a thousand of these “closed-lost” firms and had indicated reasons for each firm’s
failure to adopt. The analysis is limited to the non-adopter firms that did not license LawGeex
due to having “no clear timeline” for adoption. Other reasons that non-adopter firms gave were,
for example: insufficient resources; no clear need for LawGeex’s product (ML automated
contract review); and chose a competitor instead. The samples of non-adopter firms and adopter
firms are both sets of firms that showed interest in LawGeex’s product, as opposed to a random
sample of firms that did not show interest in an ML automation contracting product. However,
comparing the non-adopters and adopters allows for making some initial observations about the
firms that use ML automation contracting products, as compared to those that come close to
adopting but ultimately do not.
There are 91 adopter firms in the sample and 415 non-adopter firms. Data was collected
from LawGeex (specifically, from their sales team) as well as from CompuStat and
BloombergLaw. Given the sample size constraints, logits were used to compare the populations
with the dependent variable of adoption (or not).
There are no differences between adopters and non-adopters (before adoption) along the
dimensions of firm size, lawyers, and lawsuits. See Tables 1a and 1b. Further, in terms of
industry comparisons, there are roughly three times more adopters in financial investment
industries compared to non-adopters. Otherwise, adopter and non-adopter firms tend to be
37
distributed evenly across other industries. See Table 1c. Firms with higher risk exposure, those
firms that engage in more transactions and that have more contracts, are likely to be firms that
would see an immediate and long-term benefit from adoption of ML contract automation, so it is
not surprising that financial investment firms would be adopters. And finally, adopters are more
concentrated in United States (US) cities (Los Angeles, San Francisco, New York, and some
cities in Texas) as compared to non-adopters; non-adopters tend to be spread throughout the
country. See Figure 1. ML is a GPT, and GPTs need complementary assets for diffusion
(Cockburn et al. 2018, Goldfarb et al. 2019), so it follows that firms that adopt ML automation
tend to locate where they have or have access to complementary resources such as human capital
and other technology firms to be able to realize the benefits from ML.
This analysis, though limited by sampling issues related to size and endogeneity, helps to
provide background and to frame the issues explored in this paper related to ML adoption. This
analysis emphasizes the similarities between the adopter and the non-adopter firms, at least along
the variables of interest. This suggests that firms that initially license LawGeex but then do not
continue to renew may not be significantly different along these variables from firms that do
renew or that never adopt LawGeex. While there are surely heterogeneities within the set of
adopter firms along variables not examined herein, there may also be relevant similarities that
support conclusions drawn from the subsequent analyses regarding incentives for ML use.
Methodology
I use LawGeex data to examine the main research question of how to incentivize
employees to use ML, and how employees’ learning about the technology promotes use.
Specifically, I use eighty onboarding call transcripts between LawGeex and thirty customer
firms.
38
Data: Customer Firms and Onboarding Calls
The customer firms represented in the data I use were randomly selected. LawGeex is a
start-up, founded in 2015, and thus did not have internal processes for data organization. In
conversations with managers at LawGeex, they discussed how they shared with me the audio
from as many onboarding calls as they had, and that they could not locate the majority of the
onboarding call recordings due to a lack of internal organization. Additionally, the managers
discussed how the available call recordings were random, meaning that there was no reason why
they could find the recordings from some customers but not others. LawGeex only had call audio
available for multiple onboarding calls from thirty customers. They had one onboarding call
available for eight other firms. These eight firms were not used in the analyses.
The number of calls per firm varies, as does the length of time between the calls. The
maximum number of calls per firm was five and the minimum number of calls per firm was two.
The average amount of time between calls was sixteen days. In each onboarding call, there is at
least one LawGeex manager and at least one employee of the customer firm. For each of the
thirty firms, I have a “kickoff” call, which is the first call between the customer firm and
LawGeex after the sales cycle is complete. I also have at least one “playbook validation” call
between each customer firm and LawGeex. In playbook validation calls, the customer firm
employees on the call discuss with LawGeex concerns, issues, or questions they have related to
use of the product. The ‘playbook’ refers to the key that a customer firm creates to help train the
ML for a specific contract type. For example, if customer firm Alpha wants to use LawGeex for
its sales agreements, Alpha’s employees create a playbook that includes all the clauses specific
to Alpha’s specifications for a sales agreement contract. The ML uses the playbook to learn what
clauses to flag as potentially problematic – and what not to flag – for a given contract type.
39
However, the ML needs feedback from employees (users) in order to perform optimally. During
the playbook validation calls, LawGeex and users discuss how the ML can be adjusted to best
serve the users’ needs. The calls last, on average, seventy-two minutes. The purpose of the
onboarding calls is to introduce employees (users) to the technology and to assist them in
learning to use the technology.
I used topic modeling to analyze the onboarding call transcripts. I also read through all
the transcripts multiple times. Topic modeling is a well-developed ML technique to summarize
and organize large collections of text. Specifically, I used LDA (Latent Dirichlet Allocation)
topic modeling (see e.g., Blei et al. 2003). LDA is a generative probabilistic model of a
document corpus containing words (phrases). The model output consists of two tables: the first
shows the probability of selecting a particular word in the corpus when sampling from a
particular topic and the second table is the probability of selecting a particular topic when
sampling from a particular document. I used Mallet (Machine Learning for Language Toolkit), a
Java-based package from UMass Amherst (McCallum 2002). Mallet is an LDA implementation
that uses Gibbs sampling and is more precise than the standard Gensim which uses a Variational
Bayes Sampling. The output, however, is essentially the same.
First, I separated the calls by when they took place. Customer firms’ first calls (kickoff
calls) with LawGeex were one corpus; customer firms’ second calls (onboarding calls) were a
second corpus; customer firms’ third calls were a third corpus; and customer firms’ fourth calls
were a fourth corpus. Unfortunately, beyond four calls the data became too sparse to perform a
robust topic modeling analysis. Then I separated out each call by speaker and grouped responses
based on what LawGeex said and what the customer firm employees said. Therefore I ended up
with eight distinct corpuses: periods one through four, with data for LawGeex and customers in
40
each of the four periods. I observed a list of keywords that the model output for each latent topic
in each text corpus. I ran the topic modeling specifying twenty latent topics given the data. I also
ran Mallet specifying 10, 15, and 25 topics on each corpus in order to confirm that results were
generally the same. Then, I observed the term-frequency matrix for the documents for the twenty
topics in each corpus. Based on the keywords in each corpus I determined the topic labels. I
observed how these topics changed – and remained the same -- from call period one to call
period four, both for the employees and for LawGeex. See Tables 2a and 2b.
I discuss below how the results of this topic modeling and changes to the topics and
categories over time help answer how to best incentivize employees to use ML after the firm has
licensed this technology. Based on insights from agency theory and entrepreneurship theory (i.e.,
how to motivate employee innovation), I develop the propositions below which I support using
evidence from the topic modeling analysis and, where relevant, direct quotes from the
onboarding calls themselves.
Employee Payoffs for Using ML
The value of ML is not immediately apparent. ML improves its capabilities relative to the
amount it is used. At the outset, ML may have a limited scope of capabilities. For example, when
LawGeex first launches at a customer firm, the ML is limited in its ability to review contracts
based on the playbook and based on previous contracts it has reviewed. As the users at the firm
review more contracts using LawGeex, the ML improves.
ML, Tasks, and Jobs
ML improves with use, but it remains complementary to other tasks even as its
capabilities improve. The process of scaling ML use at a firm can be considered a two-step
process. In stage one, the ML is limited in its capabilities and requires direct human input in
41
order to hone and develop the algorithm. In stage one, the more the ML is used, the more data it
has from which to learn and improve. The ML is complementary to human-performed, behavior
based tasks. At this stage, while the ML can still offer efficiency gains in a given task, such as
contract review, the heavy lifting of checking the ML’s output and providing feedback to the ML
falls to the human. Within the specific task, such as contract review of a non-disclosure
agreement, the ML is strictly complementary to the human’s work. The ML does not offer
complementarities beyond the specific task for which it is specified at this stage.
In stage two, the ML technology is complementary to multiple tasks that the user
performs during the course of her job. The complementary aspects of ML are not limited to one
task. Instead, the ML for the specific task, such as contract review, is capable of doing much of
the contract review based on all the data it used. Because the ML can effectively review
contracts, the human user can perform other tasks that are part of her job. In the case of an
attorney, the other tasks that an attorney can focus on may include tasks that are more aligned
with case strategy or firm strategy, that is, tasks that are of higher value to the firm and to the
individual attorney in terms of career development.
In many sectors with low-skilled work, it is the case that automation has replaced many
of the low-skill tasks (Acemoglu and Restrepo 2018b). Tasks are comprised of two dimensions:
engineering complexity and training requirements (Feng and Graetz 2018). These dimensions are
based in the notion that complexity differs from an engineering point of view to a human activity
and behavior point of view (Moravec 1988). Based on this distinction, routine tasks may be more
easily automated because they are repetitive and thus can be engineered, while nonroutine tasks
with high engineering complexity and either low or high training requirements will be more
difficult to automate. Low-skill tasks are at the low end of the complexity distribution; high-skill
42
tasks are at the high end (Acemoglu and Autor 2010, Acemoglu and Restrepo 2018b). Tasks can
be performed by low-skill labor, high skill-labor, or capital (Acemoglu and Restrepo 2018b).
However, employees who perform high-skill tasks are less likely to be supplanted by ML
automation. Because ML reduces the cost of prediction and thus increases the value of
complements to prediction, high-skilled workers may change the tasks that they perform, but
ultimately are less likely to be made redundant than low-skilled workers who are more limited in
their ability to perform complementary tasks (and complementary tasks may also be automated)
(Acemoglu and Restrepo 2018a).
Task Efficiency Orientation
Initially, the value of ML technology may appear to the firm to be marginal. Similarly, a
firm’s innovation initiatives may not immediately capture value for the firm (Anderson and
Tushman 1990, Teece 1986, Abernathy & Utterback 1978). However, the value of ML
technology comes from long-term use of the technology.
Yet, while managers at firms that license ML technology may recognize that ML
technology captures more value for the firm in the long-term, the employees may not recognize
this trade-off. Employees may expect that ML technology will immediately make a difference in
their work and productivity. Employees’ expectations of immediate productivity gains may be
based on the popular press hype around ML and the principals’ excitement over the technology’s
capabilities (Acemoglu and Restrepo 2018b).
For example, in the first round of LawGeex calls (kickoff calls), one topic from the
employee response had the following words: busy; full; ready; improve; impact; effective;
survival; language; monthly; heavy-lifting. These words suggest that the employees expect that
the ML technology will lead to an immediate benefit in their current job tasks. Indeed, in
43
confirming this with the transcripts, an employee (attorney) at a customer firm in the financial
services industry (SIC 6219) states in the firm’s “kickoff” call with LawGeex: “We’re trying to
streamline and make our process more efficient for the review of particular NDAs, pretty
immediately. And we really see this as an opportunity to dip our toe in the water of what AI
might be able to do within our business and for our baseline efficiency, and take this as an
opportunity for other things that might arise in the future”. An employee (attorney) at another
customer firm, in the telecommunications industry (SIC 4841) states during a kickoff call: “Our
volume, in terms of NDAs, is . . . overwhelming. We need this [technology] to help so we can
get to other work.”
Then, when the ML tech fails to offer immediate benefits in the short-term, and thus fails
to live up to the employees’ expectations short-term benefits, they are not motivated to continue
to use the technology. For example, in even the second round of onboarding, one topic from the
customer responses includes the words: question; work; review; miss; time; benefit; trust; use;
train; busy. It seems that in even as little as two weeks beyond the initial optimism of the kickoff
call, some employees have already begun to lose interest because the technology has failed to
yield short term benefits. Specifically, the technology has failed to improve the employees’
efficiency in the task of contract review. For example, an employee (attorney) at a consulting
firm (SIC 7379) states during the firm’s second onboarding call: “We haven’t had a ton of time
to use the playbook, to test the playbook out. It’s been quite busy here, and we’ve had quite a
volume of contracts for review. But because we didn’t feel that [LawGeex] was quite ready to go
yet, even after our calls, we just ended up, we just ended up doing it ourselves. It took a lot of
time and we were disappointed that the product didn’t seem to be saving us much time.” During
44
this call another attorney at the same firm stated: “I thought that there would be a much more
immediate reduction in my work load that, as of now, I am just not seeing.”
In another example, the employee listed above from the telecommunications firm stated
during the playbook validation call a few weeks later: “We were excited to use AI, but to be
honest it’s been a bit more of a pain that we thought. We don’t have the time to constantly check
with the playbook, to check things over, and, to be honest, we did a little comparison and it is
taking us much longer to review a contract with LawGeex than without, so we mostly stopped
using it.” Another employee (support staff) at a biotech firm (SIC 8731) states during a (second)
playbook validation call: “The set up and, just, seeing the benefits, is taking a lot longer than we
thought. I don’t know if it’s the AI or us, or what, but this.. I didn’t think I would still be
reviewing contracts. But I am. Definitely.” Even as few as two weeks after the kickoff call
during which enthusiasm was expressed for the ML technology, many employees seem to have
changed course. Topic modeling and the qualitative examples suggest that the benefits from ML
are not immediate and the tasks that the ML is helping to automate may not be immediately
automated.
Many benefits to ML use accrue if the focus is on identifying complementarities, rather
than on pursuing automation (i.e., focusing on substitution), with the caveat that each
technology, including ML, substitutes for something, and, in doing so, complements something
else. ML can be developed in a way that emphasizes the complementarities, or in a way that
pursues substitution and lets complementarities naturally occur. This argument for ML
development is similar to the literature on product versus process technologies, in that when a
firm that adopts new technology focuses on the complementarities as the outcome of interest that
produces value, this necessitates revisiting the production process. On the other hand, if a firm
45
adopts a new technology and uses a substitution approach, meaning it tries to produce the same
result in the same way but with automation of current tasks facilitated by the technology, it is
more likely that revisiting the products used will be necessary.
In some firms, it is clear from the topic modeling results and review of the transcripts that
employees were focused on the ML’s ability to perform the task of contract review more
efficiently, rather than contemplating the complementarities, which may have been less
immediate. Firms in this orientation emphasized how the ML would help with the specific task
of contract review and that help was needed. Employees view ML as a substitute and lacked
understanding of ML as a process technology that could ultimately help their work. As seen by
the usage rates and renewal rates, these firms lacked the necessary capabilities to implement
more complex process (relative to product) re-design.
Proposition 1: When employees view ML as a product technology that is a substitute to their
jobs, the employees will be less likely to use the ML and the firm will be less likely to renew its
license for ML.
Job Efficiency Orientation
While some employees may focus on the short-term, task efficiency gains in their job
tasks, others focus on the long-term gains from using a new ML product, particularly in terms of
how use of the ML product will change the employees’ job role. For example, in the context of
ML contract review automation, employees can focus on how use of this technology can
facilitate legal professionals spending a larger percentage of their time negotiating over contract
terms, working with clients, or drafting memos, which all may be tasks that are valued more
highly by the firm than contract review. This ultimately may enhance the employee’s skill set
and job security at the firm. For example, in one third round playbook validation call with a firm
in the finance industry (SIC 6211), the employee, an attorney, states: “Overall it’s going well so
46
far; I’ve had the opportunity to devote more of my time to tasks that, frankly are higher value for
our clients. Not having to review forty or fifty contracts in detail is so valuable.” In the first
round of employee calls (kickoff calls), one topic includes the words: “efficient; value; firm;
client; time; long-term; immediate; future; counsel; management”. This seems to support the idea
that employees notice and distinguish between short-and long-term benefits from ML use. To
support this point, an attorney at a firm in the manufacturing industry (SIC 3593) states during a
kickoff call: “My boss, the GC, really wants us to get this right. She thinks that AI is the future
and that we can take as much time as we need to learn this, and use it, because it’s important”.
Later, during a third round onboarding call, this employee reiterates: “We wish this was going
faster, but, I mean basically we’re in it for the long haul, we want to get this to work.” In another
second round onboarding call, an attorney at a law firm (SIC 8111) stated: “Once I finally had
the time to build up the playbook, to review a bunch of NDAs with LawGeex, it helped a lot
because then I could see how the .. how LawGeex was able to do better at reviewing. And I
would check it over and find fewer things that I had to change manually. So yes, I do see this
eventually really helping out and time savings. But I believe it’s going to take a bit of time”.
In firms with a job efficiency orientation, employees focus on ML’s ability to add
efficiencies to their work. For example: “I am interested in this tool and standardizing the claims
because sometimes we do have disputes, and dealing with corporate housekeeping, and this will
just make my job a lot easier”; “All I have to do is just drop [the contract] into LawGeex and
then I’m able to do other stuff. So that’s definitely a positive”; “Volume, in terms of NDAs, is
overwhelming. We need this[technology] to help so we can get to other work”; “I do see this
eventually really helping out and [with]time savings across the board”; “We actually have a
47
senior lawyer assisting with third-party reviews. So the main driver to implement LawGeex is so
that paralegals, or contract managers, can do the work. And she can do other work.”
Given the nature of ML, benefits of adoption and use of the technology accrue when
identifying and capitalizing on complementarities. In this orientation, employees have identified
that continued use of ML will promote complementarities. While they still do not fully grasp the
long-term benefits that ML will confer to the firm, they do observe that using ML can help them
perform their jobs. Additionally, these employees will be more open to learning and re-designing
processes in order to benefit their job efficiency. In turn, the employees firm will benefit from
the employees’ use of ML as a process technology. Thus,
Proposition 2: When employees view ML as a process technology that promotes
complementarities in their jobs, the employees will be more likely to use the ML and the firms
will be more likely to renew their license for the ML.
Strategic Value Orientation
In firms with a strategic value orientation, employees see that the value of ML to the firm
will accrue over the long-term. Instead of focusing on the benefits that ML use would confer to
the employees’ job, they understand that ML will help the firm capture more value in the long-
term, and that this will have benefits for their job and tasks as well. For example: “[LawGeex] is
just great from a consistency point of view, from the whole-firm perspective”; “The goal with
this whole process is to get the NDAs to a point where we use the data for other processes and
deals, not just in legal” and “Long-term a benefit would be reduced use of outside counsel; we’re
invested in the larger business case.”
Employees at these firms understand that ML is a process technology and thus, because
ML automates judgment processes, it is harder to define specifically where and how to best use it
48
within a firm. Employees see value in developing processes that complement ML and that, in
turn, will benefit the firm.
Proposition 3: When employees view ML as a process (vs product) technology that benefits the
firm in the long-term, the employees will be more likely to use the ML and the firms will be more
likely to renew their license for ML.
Manager and Employee Alignment Regarding Use of ML
Use of ML in firms will result in changes to the tasks that high-skilled workers perform
(Acemoglu and Restrepo 2019). Yet this transition has been slower than anticipated. There are a
number of potential mechanisms to explain why use of ML within firms has not accelerated at a
fast pace given advances in the technology and the increase in relatively easily deployable
products that use ML. This paper focuses on one of these potential mechanisms: misaligned
incentives for ML use between managers and employees. Misaligned incentives have been
shown to stifle innovation (Aghion and Tirole 1997, Leiponen 2008) and slow organizational
change (Makri et al. 2006).
In conversations with managers and employees at firms that license ML products, the
issue of misaligned incentives within the firm causing slow or no adoption arises frequently.
Additionally, employees and founders of start-ups with ML products also flag misaligned
incentives internal to their customer firms to be a major issue inhibiting use of the ML product.
For example, the former Chief Marketing Officer of LawGeex, discusses in an interview how the
mismatch between management and employee expectations for the product contributes to
incentives issues: “What we find, frequently, with customers and prospective customers, is that
management is excited to launch the product but they either don’t clear it with their employees or
they don’t tell their employees what it's going to do. So then we have to spend a lot of time
telling the employees that this is not going to take their job, and how it will actually make their
49
job better.” Employees will lack motivation to use a new technology if their expectations about
the technology are that it will replace them. After licensing a new technology, managers may be
keen for their employees to use it. However, they must take the correct approach to structuring
non-pecuniary incentive schemes for their employees given the specific capabilities of ML.
Managers evaluate and license new technologies to gain a competitive advantage.
Employees may or may not have the same goals as managers for use of the new technologies. In
order to ensure that the manager is able to achieve her goal of increasing use of the new product
within the firm to gain competitive advantage, the manager must also incentivize her employees
to use the product. Employee incentives are often based on job concerns (Gibbons and Murphy
1992) and compensation (Grossman and Hart 1983). On the other hand, managers, while still
concerned with compensation and performance, tend to also care more about the firm as a whole
(Aghion et al. 2013). Therefore, incentives between managers and employees can be misaligned
when it comes to technological innovation.
The topic modeling analysis and the data on usage and renewal showed that some firms
switched orientation from short-to long-term during the course of onboarding. In these firms, the
call transcripts showed that, during the course of onboarding, the employees ‘bought in’ to
managers’ vision about the value of ML as a process (versus product) technology. That is, the
employees became aligned with the managers’ perspective on ML.
Individual employees’ learning can influence the firm’s absorptive capacity (e.g., Yao &
Chang, 2017). Absorptive capacity is the “ability of a firm to identify, assimilate, and exploit
knowledge from the environment” (Cohen & Levinthal, 1989, p. 589). Employees’ learning
about the technology and using the technology facilitates understanding of the unique benefits it
confers in reforming processes within the firm.
50
Proposition 4a: Employees’ learning more about ML may result in employee and manager
alignment about the value of ML as a process technology.
Proposition 4b: Manager and employee alignment will result in higher usage and renewal rates
of the ML compared to firms that lack alignment.
Employees’ Short-term and Long-term Incentives
Incentives motivate behavior through expected payoffs. Expected payoffs only serve as
an incentive if employees have a degree of certainty in the payoff’s expected value (Grossman
and Hart 1983). In terms of innovation within a firm, incentive schemes are most effective for
motivating agents to be more innovative when the schemes tolerate early failure and reward
long-term success (Manso 2011). Rewarding long-term success means focusing on long-term
compensation, job security, and timely feedback on performance rather than on bonuses
(Gibbons and Murphy 1992, Manso 2011). Effectively, because the innovation process is long
and the expected payoffs are not immediate, focusing on the long-term value associated with
innovation helps motivate employees to engage in innovation.
Managers want to incentivize employees to perform tasks in support of the firm’s goals
(Grossman and Hart 1983). To do this, managers provide incentive schemes to employees in the
form of compensation, benefits, etc. (Gibbons and Murphy 1992). Managers also provide,
ideally, clear parameters about the expectations and tasks that merit the benefits and
compensation. Under many conditions, the outcome of employees performing the specified tasks
is clear. However, tasks performed under some conditions have uncertain outcomes, whether in
terms of performance or employment impact. When the outcome of performing these tasks is
uncertain, it can be more difficult to incentivize employees to act. Incentivizing employee
innovation should focus on long-term rather than short-term success (Manso 2011). That is, the
incentives should be framed in such a way to emphasize how performing the task (here, using
51
ML) will increase the employee’s value to her job and to the firm, as opposed to focusing on the
efficiency gains that the employee may reap in performing the task that is set to be automated.
Incentivizing employees for innovation is distinct from incentivizing employees to use a
technological innovation in terms of how the mechanisms of uncertainty and learning interact.
When incentivizing innovation, focusing on the long-term tampers the uncertainty employees
may feel from unknown outcomes from the innovation process that may affect managers’
perceptions of their value. Effectively, the long-term perspective here rewards proactive
experimentation and reduces the uncertainty associated with how their performance will be
judged. When incentivizing use of a technology like ML, employee uncertainty does not come
from performance; indeed, performance on a task when using ML is expected to become more
efficient. Instead, the employee uncertainty when using ML stems from the effect of the
efficiency gain on their job, rather than on performance. Further, while learning is important to
employee innovation, the learning that occurs when using a new technology specifically reduces
the uncertainty employees feel about the role of the technology in potential job displacement. In
both scenarios, effective incentive schemes emphasize long-term value over short-term value,
but sources of uncertainty and the role of learning are different.
While incentivizing innovation and incentivizing use of a ML enabled product are
distinct, I propose that the most effective incentive schemes for motivating employee behavior to
engage in activities with unknown outcomes will be similar.
Preliminary Analysis Using Usage and Renewal Data
Subsequently, I performed preliminary analyses of these propositions using the
dependent variables of usage (number of contract uploads per customer firm) and customer
firms’ yearly renewal of their LawGeex license. These analyses are preliminary given the small
52
sample size: only thirty firms were involved in the topic modeling analysis, the time period over
which the usage data extends varies by firm, and LawGeex was founded in 2015. Additionally,
grouping the firms (e.g., firms that focused on short-term incentives in round one, and then long-
term incentives in subsequent rounds) further reduced the sample. However, despite these
limitations, the analyses serve as preliminary support for propositions herein.
To conduct these analyses, I looked to the topic modeling analysis and the raw call
transcripts for round one (kickoff calls). Using the topic modeling analyses, I reviewed the term-
frequency matrices and observed which customer firms were primarily driving certain topics. For
example, in the topic models for the kickoff calls (round one), I found that topics “technology
skepticism” and “anticipating task efficiency gains” were primarily driven by firms C, D, G, H,
M, X, AA, and BB (see Tables 2a, 2b, and 3). Based on my theorizing, firms that expressed
skepticism about adoption and interpreted the benefits of LawGeex as short-term efficiency
gains, should not see sustained usage over time.
Changes Over Time
Because the onboarding process takes place over a number of weeks, I used the topic
modeling data and the raw transcripts to observe changes to how firms discuss incentives in
subsequent onboarding calls. Firms could fall into one of four categories, depending on whether
they start in a category of “short-term” or “long-term” incentives, and then whether they change
over time. I measured a change between short and long term incentives as occurring at any time
between calls, whether it was between the kickoff call and an onboarding call, between two
subsequent onboarding calls, or between a kickoff call and a second or third onboarding call.
Using the topic modeling weights to determine which firms affected which topics in each call, I
was able to measure a firm’s changing attitudes. Using firm pseudonyms to maintain anonymity,
53
I chart each firm’s categorization over each call and measure changes over time in how
employees perceive the incentives to use ML. See Table 3.
Then, using data on employee level usage rates (how many contracts were uploaded per
firm per month) and firm level renewal rates (yearly), I analyze whether firms that changed from
short- to long-term incentive schemes demonstrated their commitment to long-term use of ML
by (1) continuing to upload contracts over time; and (2) renewing their LawGeex license after
the first year. Given the small sample size, it is difficult to draw robust conclusions about the
findings. However, the findings are presented herein as a preliminary analysis that serves to
provide some initial support for the propositions. See Tables 4a and 4b.
First, I examine whether firms that focused on short-term incentives in period one use the
product less (fewer uploads) and whether firms that focused on long-term incentives in period
one use the product more. I find support for the proposition that when employees at a firm have a
more short-term framing of the benefits that the ML will provide, they are ultimately less likely
to use the ML over an extended period of time. While this effect does not seem to be present
after the first call, after call two and subsequently after call three, the amount of contracts
uploaded between firms with short-term versus long-term orientations towards ML and jobs are
significantly different. Additionally, in terms of renewals, after call three in the onboarding
process (n=25 firms), there was a significant difference in renewals for firms categorized as
short-term versus long-term. See Tables 4a and 4b.
I also examined whether a firm changing from short-term to long-term (and vice versa)
had an effect on the likelihood of renewal, as compared to firms that stayed either short- or long-
term throughout the entire onboarding process. Firms that go from a short-term frame to a long-
term frame during the course of onboarding upload more contracts than those firms that do not
54
change frames and those that go from long-term to short-term. Firms that end on a short-term
frame at the end of the onboarding calls upload significantly fewer contracts than firms that end
up with a long-term frame at the end of onboarding. The same holds true for renewals: firms that
changed from short to long-term frames for understanding ML’s impact renewed at significantly
higher rates than those that did not change from short- to long-term. Firms that ended up as long-
term, whether they stayed consistently long-term or whether they changed to long-term, renewed
at significantly higher rates than those firms that ended up as focusing on short-term incentives.
The renewal rate for the LawGeex customer firms in this sample (the 30 firms used) was 67%
after one year and 80% after two years. These numbers are roughly equivalent to LawGeex’s
renewal rates for all customers. This suggests that while the sample of 30 firms used in this
project is considerably smaller than the overall number of LawGeex customers, the sample is
roughly representative.
Limitations
Given the small sample size of firms used herein (30), it is difficult to draw robust
conclusions about the behavior across a larger population of firms. In the future, I hope to
expand the sample and examine these questions from a more quantitative perspective. Further,
given that this paper only examines one firm’s technology and is limited to the legal industry,
there may be opportunities to broaden the scope and draw additional conclusions.
This is a descriptive paper, not a regression paper; thus I examine data from CI and its
customer firms in order to describe the phenomenon of ML technology adoption and the value
chain of adoption within the firm. I juxtapose findings from the detailed empirical examination
of the adoption and use of ML legal technology with existing theories that explain firm
dynamics, especially intra-firm dynamics, in technology commercialization and adoption. This
55
approach allows me to identify instances where existing theory either fails to explain the
observed patterns, or where its predictions are inconsistent with what is observed. I hope that this
work encourages future scholarly work that develops theories about the nuances of ML
technology adoption, particularly in terms of the dynamics of firm and industry structure.
Discussion
In this paper, I seek to elucidate the mechanisms, particularly those at the individual
level, that are involved in the use – and sustained use – of an innovative technology at the firm
level. ML is unique in comparison to other technologies in its ability to learn as it receives more
data inputs and in that its value increases relative to use. Firms thus may not immediately realize
the value, such as time savings, from use of a product with ML. Over time, however, the benefits
of using ML will become more apparent to the firm. Additionally, due to this unique
characteristic of ML and due to employees’ preconceived ideas about ML and job security,
employees may not be motivated to use ML. Employees are often fearful that ML will replace
them in their jobs. Evidence from LawGeex onboarding call transcripts and the topic modeling
analysis performed on these transcripts largely confirms this.
However, when the long-term benefits and value associated with using ML are
emphasized, either by managers or by a third party (in this case, the technology provider), the
employees become more willing to use the ML, as compared to when the incentives focus on the
short-term task efficiency associated with ML use. The difference in motivating use between
short-term and long-term frames aligns with the employees’ uncertainty about their jobs.
Focusing on how the elimination of one task through ML use may lead to more opportunities to
perform high value tasks or expand the employee’s job role into performing high value tasks will
motivate employees to be more willing to use the ML, and this effect seems to last.
56
Motivating Sustained Use of Innovation
Many aspects of new technology adoption and proliferation have been fully addressed by
the management literature, both from the perspectives of the users and the companies
manufacturing the technology, and at the firm and individual levels. Additionally, there is an
extensive literature on how managers can motivate employee innovation. However, there is little
work on how firms can motivate the sustained use of an innovative technology such as ML.
In examining how to motivate employees’ sustained use of ML, and thus how the
sustained use is integrated into firm processes and ultimately affects the firm, this paper also
seeks to address how individuals and agency theory dynamics interact with firm heterogeneities
and, in turn, affect firm capabilities. Firm capabilities are, broadly, what an organization can
actually do (Jacobides and Winter 2012, Pisano and Teece 2007, Teece et al. 1997), and are
distinct from concepts such as an organization’s intentions, incentives, motivations, and
production function. However, in the literature, capabilities are often conflated with, for
example, a firm’s motivations for learning or how a firm incentivizes innovation diffusion. The
conflation of capabilities with these other concepts can hinder understanding of what motivates
knowledge transfer in organizations.
I propose that the individual dynamics between managers and employees play an
important role in determining how firms develop capabilities and how firms incentivize sustained
use of innovation. Specifically, the managers’ framing of the ML technology’s applications as
either providing short-term or long-term benefits to the employees critically affects the
employees’ likelihood of sustained use. If managers focus on the short-term efficiency gains of
the ML, as in how the ML will automate one task that the employee performs, then the
employees are unlikely to use the ML technology. If the managers frame the ML technology as
57
providing a means to reshape the employees’ job role to focus on tasks that may be of higher
value to the firm (now that a routine task has been largely automated by ML), the employees will
be more likely to use the ML and to continue to use it over time.
As can be seen from the onboarding calls, which take place over a period of time, the
employees discuss both their managers’ framing of ML and their own experiences in using the
ML. Over time, it is possible to see from the onboarding calls that the employees learn about ML
and become more accustomed to the role that it plays in their job. Thus, this paper demonstrates
how managers can motivate employees’ use of a new technological innovation, and how
employees learn about that technological innovation over time, with the help of information
provided by the technology provider (here, LawGeex). Managers’ focus on the long-term value
of ML, rather than the short-term efficiency gains, is reinforced by the employees’ learning about
the technology over time. When employees do not understand what ML is and what its
capabilities are, they are unlikely to be motivated to use it over the long-term. When managers
frame the technology as a benefit to the employees job in the long term then employees are more
motivated to use the technology and they also have the opportunity to learn more about it too.
This work builds on existing literature that discusses how insights from agency theory
can inform how firm capabilities are developed in the context of technological innovation
through the intervention and framing of individual level incentives (Cabral et al. 2020, Chatterji
et al. 2019, Fu et al. 2019). Indeed, here the long-term use of ML relies not only on firm
learning, which begins with individual employees, but also on how managers frame use of ML
and the subsequent impact on the employees’ jobs.
In the future, I hope to gather additional data, both qualitative and quantitative, to further
explore the interactions between the dual mechanisms of learning and incentive framing in
58
motivating long-term use of machine learning. Additionally, this paper primarily uses qualitative
data to suggest propositions that are preliminary tested using a small sample size of firms. There
are opportunities to study these propositions using large-scale quantitative data that applies to
multiple firms across a variety of industries.
Additionally, I hope to use data to address questions that follow from this line of
research. For example, subsequent research could examine whether this effect of motivating ML
use is static over time, what incentive contracts for employees may look like if firms chose to
utilize this governance option and how these contracts may impact use, how incentivizing
managers for risk-taking may or may not incentivize employees to use ML, and whether these
incentive schemes, especially the role of information and learning in motivating use, may impact
firm boundaries. Given that this paper is specific to ML, it would be interesting to examine
whether this theorizing holds for other technological innovations, such as blockchain, as well.
Conclusion
Firms license and use technological innovations to add value. However, employees’
uncertainty related to the innovation may inhibit use of the innovation and, in turn, result in the
firm failing to capture value from the innovation. In the case of ML, this risk is especially salient
as ML, by its nature, improves – and adds value – relative to increased usage. Thus this paper
takes a dynamic view of incentive issues and attempts to understand the employees’ uncertainty
in using ML and, in doing so, provide a framework for how managers can reduce employees’
uncertainty, leading to increased employee usage of ML.
The academic literature on automation clearly finds that automation displaces and
reinstates labor (Acemoglu and Restrepo 2018a). However, the popular, widespread notions
about artificial intelligence (AI), automation, and ML tend to be that these technologies are set to
59
only displace (Smith and Anderson 2017). While ML will automate some of the tasks that many
employees, even those in high-skilled jobs, do, the reinstatement effect means that there will be
other, often more complex and high value tasks available for the displaced employees to do, even
within the same job (Acemoglu and Restrepo 2018a, 2019). In the absence of a large-scale
education campaign about the economic impact of ML and similar technologies designed to
reduce employee uncertainty, however, firms should take steps to ensure that they are able to use
ML and thus retain a competitive advantage.
The broader discussion about ML use also has implications for the commercialization,
growth, and diffusion of ML as a technology. While the capabilities of ML are growing given
advances in computer processing and data storage, the commercial uses of ML remain limited.
Firms that could potentially use ML run into a number of setbacks when it comes to launching
and sustaining use of ML. In turn, the commercial ML products tend to be limited in
technological capabilities and tailored to a highly specific use to ensure a market and growth for
the producer. In turn, this may have a negative impact on the technology S-curve; whether
permanent or temporary is as yet unknown (see e.g., Gans et al. 2020).
The nature of the mixed methods approach of this paper and the dynamic view of the data
allows for understanding some of the micro-mechanisms that inhibit or promote a firm capturing
value from a new innovation. These findings are relevant to the literature on automation and the
future of work. While we still know relatively little about how automation will impact the labor
market and productivity, past innovations have increased demand for labor and wages. If tasks
that require low-skill labor will be automated (first), then highly skilled workers can add
substantial value by contributing more effort to high-skill tasks that require more training and
rely more on human behavior. Yet framing these issues for the employees remains a challenge;
60
here, I suggest that using the essentially economic argument of different value tasks as part of a
job can help employees understand that automation can actually increase their value to the firm.
Understanding the factors that affect the individual (employee) decisions to begin using a
new technology such as ML are essential to understand for both determinants of growth and for
the producers of these technologies. Here, in examining how manager incentives and information
about the technology can decrease employee uncertainty about the technology, we begin to
understand more about how firms can ensure that they are able to capture value from
technologies like machine learning.
Tables 1a. Descriptive Statistics of Adopters and Non-Adopters.
61
Table 1b. Comparing Adopters and Non-Adopters.
Table 1c. Adopters and Non-Adopters by Industry.
62
Figure 1. Adopters and Non-Adopters by Geography of Headquarters.
Non-Adopters
Adopters
63
Table 2a. Topics from onboarding calls: Customer responses
64
Table 2b. Topics from onboarding calls: LawGeex responses.
65
Table 3. Customer calls chart.
66
Table 4a. Descriptives.
Table 4b. Descriptives.
67
CONNECTING CHAPTERS TWO AND THREE
Chapter two examines dynamics within a firm that has licensed a new AI-enabled technology
with the aim of capturing value and creating value from that technology. Chapter two focuses on
the licensing firm and its internal dynamics between managers and employees, while holding
constant the actions of the innovator firm. In focusing on the licensing firm’s internal dynamics I
hope to isolate some of the variables that contribute to successful technology adoption and
commercialization. I find that alignment between employees and managers, and emphasizing
understanding the value that the technology adds in the long term at the individual level, helps a
firm successfully integrate AI-enabled technology in order to capture value from it.
In addition to understanding the factors within the licensing firm that contribute to continued
licensing and technology adoption, I propose that understanding the interdependencies between
the inventor firms’ commercialization strategies and the characteristics of the licensing firms
allows for understanding what technology commercialization strategies are likely to be the most
successful, especially for frontier technologies such as ML.
With this in mind, in chapter three, we develop a framework for understanding heterogeneity
in licensing firms’ readiness to adopt ML and how this may contribute to the inventor firms’
commercialization strategies. I hope that this chapter extends the work from chapter two by
taking a firm level perspective on technology commercialization strategies, given firm
heterogeneities. Chapter three intends to extend the work of chapter two by attempting to
categorize firms based on their readiness to adopt the ML technology. Based on findings from
Chapter two, a firm’s readiness to adopt and sophistication about the technology depends on
factors at the individual level, including alignment between employees and managers. Thus in
Chapter three, we attempt to use these individual factors to extrapolate the firm’s overall
68
readiness to adopt ML and how this affects usage and renewals. Firms vary in their readiness,
and in their knowledge about the technology, and thus we explore whether firms can fit into
different categories based on these dimensions. We find evidence to support the framework and
the four categories of firms, which in turn can help inventor firms strategically commercialize
their technology based on specific characteristics of the licensees.
By going from the individual level factors that influence firm technology
commercialization and adoption, to the firm level factors that may influence commercialization
and adoption, chapter three builds on chapter two. I hope that continued work on Chapter three
may yield additional insights into how inventor firms can adapt their strategies depending on the
readiness and sophistication of the target licensee firms.
Additionally, chapters two and three add to the literature on technology
commercialization and technology adoption by showing, first, how individual level interactions
can influence a firm’s ability to capture or create value from a technology; and second, how the
processes of technology commercialization and adoption are interdependent, especially in the
context of a technology like ML where its value relies on diverse and robust usage.
69
CHAPTER THREE: Adoption or Acquisition?
Examples from the Legal Technology Industry
Abstract
We develop a framework for understanding variation in successful AI product commercialization
using novel data related to AI-enabled software consisting of onboarding calls between customer
firms and the product developer firm and data on customers’ usage and renewals. We describe
four distinct categories of firms that differ according to a customer firm’s readiness to adopt the
product and the specificity of the firm’s knowledge about the product, technology, and
capabilities. We examine the degree to which these variables impact renewals, product
customization, and use, finding that high levels of specificity matters more to renewals and
continued use, whereas a high levels of readiness to adopt is more likely to impact product
customization. We contribute to literature on AI commercialization and its impact on firm
strategy.
Introduction
Firms continuously evaluate new technologies and business models to attempt to
maintain competitive advantage – and to avoid becoming the next cautionary tale, such as
Blockbuster or Kodak. Even industries that are not considered to be particularly dynamic, such
as book retailing or video rental, can be disrupted by innovations (see e.g., Christensen & Bower,
1996). As an innovation appears, firms consider to what extent they wish to incorporate the
innovation into their existing business model, pivot to a new business model based on the
innovation, or not adopt the innovation. While many incumbents fear becoming the next
Blockbuster – being disrupted by novel technologies – these incumbents do not all take the same
approach as they seek to protect and enhance their competitive positions.
The work on technology adoption (see e.g., Adner & Kapoor, 2010; Raffaelli, Glynn &
Tushman, 2017, Hitt & Brynjolfsson, 1995) generally considers adoption as a dichotomous
choice in which firms either adopt a new technology or they do not adopt. But in practice, many
firms take a more nuanced approach, especially when the technology is novel, disruptive,
or process-based and does not necessarily involve producing a new product. When a new
70
technology is embedded in a product or way of selling, such as Amazon selling books on the
internet, adoption can be easier to achieve. Indeed, Barnes & Noble responded to Amazon’s
move and were considered to have adopted the new technology when they opened B&N.com—
their own online bookstore. Barnes & Noble adopted the new technology and embraced the new
way of selling (though it was essentially too late).
Technology such as artificial intelligence (AI) however, considered to be a general
purpose technology (Cockburn et al. 2018, Goldfarb et al. 2019), is very different from
many other technologies as it has the potential to disrupt aspects of firm strategy such as how
firms make their products, decide what products to make, and perform other decision-making or
process-based tasks that are largely invisible to customers, whether firms or individuals. As a
GPT, AI is poised to bring about a wave of complementary innovations in a wide and expanding
range of applications sectors (Trajtenberg, 2018). Firms associated with the GPT sector and
those at the forefront of the deployment of the GPT in the main applications sectors are poised to
be the main beneficiaries of the economic disruption. At the customer firm level, new output -
- more standardized contracts, better analytics, etc., -- will be evident after adopting AI, but
the role that AI played in those changes is less clear, unless the customer has technical
understanding of the AI or the firms deploying the GPT conduct elaborate onboarding to convey
this knowledge. Thus, a key challenge for customer firms when considering the adoption of GPT
like AI pertains to lack of explicit understanding of immediate and long-term economic benefits.
When a GPT such as AI becomes more commercially available, firms must decide to
what extent they want to adopt it (e.g., Brynjolfsson, Rock, & Syverson, 2018). Rather than
adoption as an all or nothing choice, firms decide to leverage AI to varying degrees in the
organization. Whether a firm “fully” adopts GPT like AI or just experiments with it off to the
71
side is an important choice for the firm. Full adoption carries greater risks and potential rewards,
while experimenting is a safe and low-risk way to learn more about the technology without
committing major resources to implementation. Therefore, it can be difficult to tell whether a
firm that licenses AI technology-based products or services is really using the products or
services in a way that is poised to fundamentally alter some aspect of how the firm conducts its
business -- or whether the firm is just licensing it for other reasons, such as legitimacy or
experimentation.
Extant strategy research offers little insight about characteristics of firms that adopt – and
successfully adopt over time – AI-enabled technologies, as compared to those firms that
adopt more superficially (i.e., adopt and then fail to renew their licenses to these technologies),
why firms adopt disruptive GPT like AI superficially versus more fully, and the factors that
determine firm decisions in this area. While there is work on the macroeconomic impact of AI
(see e.g., Brynjolfsson, Rock & Syverson, 2018; Furman & Seamans, 2018), there is less work
on how AI enabled products impact firm strategy. Recent work on how AI impacts labor markets
suggests that adoption of an AI product by in-house legal departments reduces the number of
lawyers in the legal departments (Rich, 2021). Other work similarly suggests that AI adoption
will automate prediction aspects of jobs (Agrawal et al., 2019; Felten, Raj & Seamans,
2019). And other work identifies the distinction between adoption and retention when
developing a new technology (see e.g., Taylor 2010. We propose that it is necessary to
understand how, given the nuances of AI and its impact as a GPT, AI products integrate with and
subsequently impact firm strategy.
In this paper, we seek to explore when, why, and which types of firms tend to engage in
superficial acquisition versus more complete adoption of AI. Using proprietary data from a
72
leading producer (LawGeex) of an AI-enabled contract review product about firms that have
chosen to adopt their AI-enabled product used by in-house legal departments and some law
firms, we seek to unpack the issues and successes that arise during customers’ first year of
product use. LawGeex has a lengthy onboarding process for new customer firms, with frequent,
multiple touch points with customers during the customers’ first month (or even year) of use. We
have seventy-five call transcripts from 30 of CI’s customer-firms (“customers”), interviews with
representatives from Contract Intelligence, customer surveys, and quantitative data on usage
(contract uploads) and renewal rates for the software. Triangulating these data, we use a mixed
methods approach that incorporates text analysis and quantitative data on usage and renewal to
examine our research question.
Through these analyses, we find a number of reasons why customers license LawGeex’s
technology. We break down these reasons for licensing into four distinct categories: (1) a desire
to experiment to see how the technology works and to see if it might be useful for the firm at
some point (tinkerer), (2) a desire to appear innovative without real intent to use the technology
(status seeker), (3) a desire to understand the technology but with a goal to only use it once it has
been more generally accepted (fast follower) and (4) a desire to fully implement the technology
in the near-to-medium term (serious user). By classifying and categorizing firms based on
their acquisition and adoption behaviors, we offer insights about factors that propel firms to
engage in such behaviors, the effects of an acquisition versus adoption decision for customer
firms and AI product firms, and examine implications for firm strategy that can enable
successful AI adoption.
Through this effort, we seek to several theoretical contributions. First, we seek to build
theory about AI adoption that takes a nuanced approach to understanding when, why, and which
73
firms seek to acquire versus adopt new GPT like AI. In particular, our approach allows for
deeper understanding about adoption of GPT by differentiating AI adoption from other
technology adoption given the unique aspects of AI and how it is deployed in customer products
such as software. We highlight that treating AI adoption as a dichotomous variable misses the
complexity inherent in AI technology and that successful firm strategy around AI adoption
requires a long-term orientation emphasizing firms’ underlying readiness to adopt the technology
and the specificity of their knowledge and capabilities about the process and the technology.
Next, we also hope to contribute to the burgeoning literature on AI in strategic
management. Research on AI tends to focus on more broad economic trends. We hope that
by leveraging our proprietary data on AI product use in multiple companies we will be able to
contribute to this literature and extant literature on technology adoption.
From a practical standpoint, we offer insights that will help firms formulate strategy for
how to incorporate AI into their business plans – or not. We hope this is a first step in pursuing
research into AI. We also believe that these insights can be useful for firms developing AI
products, in terms of competitive positioning, marketing, and product developments.
Theoretical Foundations: AI and Technology Adoption
Artificial Intelligence: Opportunities and Challenges for Firm Strategy
Recent work has shown that AI (broadly; and machine learning as well, more
specifically) is a general purpose technology (see e.g., Cockburn et al. 2018, Goldfarb et al.
2019; Trajtenberg, 2018). GPTs are characterized by pervasiveness, meaning they are frequently
used as inputs downstream; inherent potential for technical improvements; and innovational
complementarities (Bresnahan & Trajtenberg, 1995). As GPTs improve and proliferate
throughout the economy, they bring about generalized productivity gains (Trajtenberg, 2018).
74
Use of GPTs in the economy implies an increasing return to scale, which also impacts the rate of
technical advance. However, this model of potential growth can make it difficult to fully exploit
the opportunities presented by GPTs as they evolve: GPTs relationship with users should not be
limited to arms-length market transactions as this can limit innovation (Breshahan &
Trajtenberg, 1995).
When considering the applications of AI as a GPT, then, it is important to ensure that AI
is deployed in a software or system to its maximum potential. However, because AI is a process
technology, its applications – unlike that of the internet, for example – are less obvious. AI
solves prediction problems (Agrawal, Gans & Goldfarb, 2018) and as such it can be a key
component of shaping firm strategy. However, managers’ willingness to rely on, or
even consider, AI as an input for firm strategy is inconsistent. The lack of knowledge about the
types of problems AI solves and how to best utilize its outputs lead to a general hesitancy or
unwillingness to fully utilize AI to its full potential within an organization.
AI requires large amounts of data in order to make predictions and, as such, the
infrastructure required for deploying an enterprise level AI, or even a more limited AI-enabled
product, requires the coordination and control between multiple divisions, units, and employees
across an organization. Additionally, the results, or output, from AI are not always immediate,
given the time required to train and fine tune the system. Further, the predictions that result from
an AI, depending on the enterprise level system or the particular product, may require some
additional interpretation from users. Users’ expectations surrounding AI may be at odds with
what the technology can actually produce. It is important to understand the capabilities of AI
and, especially, what it cannot do (see e.g., Agrawal et al., 2018). Therefore, successful AI
75
implementation requires knowledge about the underlying technology and how it works, even to a
surface level, non-technical degree.
Technology Adoption: Implications for AI
As noted, AI is a general purpose technology and it is likely that a firm’s successful
adoption of AI would similarly benefit from the complementarities of the AI being used more
largely in the market and, more narrowly, within the firm itself. If a firm can successfully adopt
AI within a specific product for a limited use, it is likely that the benefits will diffuse throughout
the firm and will not be limited to the specific purpose for which the AI was adopted.
However, there are several complexities involved in AI adoption specific to the nature of
the technology itself. Therefore, it is not surprising that many firms have difficulty successfully
adopting AI at either the enterprise or product level. The traditional theoretical notion of
technology adoption is as a dichotomous variable (see e.g., Anderson & Tushman,
1990; Cabral & Leiblein, 2001) – the technology is either adopted or not adopted.
The underlying assumption in this approach is that the results of the adoption will be immediate
or nearly immediate, allowing for a feedback loop that reinforces the benefits of the adoption to
the user (see e.g., Schilling, 2002). But, if the results or outputs from the adopted technology are
not immediate, such as with AI adoption, there is evidence to suggest that adoption may have
more nuance than being a mere a dichotomous, yes or no, factor (see e.g., Lanzolla & Suarez,
2012). That is to say that there may be initial adoption for a period that then results, soon after, in
a failure to fully embrace the technology.
Similarly, the challenges that firms face in innovating, especially when the potential
innovation can affect firm strategy, are numerous. Firms face challenges stemming from inertia
(see e.g., Schilling, 1998); complementarities (see e.g., Adner & Kapoor, 2010; Teece, 2007); a
76
lack of capabilities (see e.g., Barney 1991); and uncertainty about dominant design (see e.g.,
Suarez, 2005) that can hinder efforts to innovate. If we consider these organizational challenges
as applying to customer firms who are considering adopting an AI enabled product, this might
also inform our understanding around adoption challenges that are particular to AI.
In theory, these challenges also suggest that firms that produce AI tech must promote the
development of customer firm capabilities for knowledge about AI and about the firm’s own data
infrastructure. Additionally, producers should develop their customer firms’ absorptive capacity,
or ability to assimilate and utilize future information (Cohen & Levinthal, 1989). In our research,
we delve into specific firm-level characteristics that might allow for successful, long-term AI
adoption. Specifically, we propose that it is important to understand a customer firm’s level of
readiness and existing capabilities/their absorptive capacity – qualities that might affect use and
long-term adoption.
A firm’s ability to create and appropriate value is increasingly seen as dependent on
complementary products in the business ecosystem and the firms that produce these products
(Brandenburger & Nalebuff, 1996; Porter, 1998). Often these products are new, innovative
products (see e.g., Wu, Zan & Levinthal, 2014). However, research on how firms build this
interdependence and incorporate these products into their own strategy is relatively scarce (see
e.g., Kapoor & Lee, 2013) especially when considered in the context of a technology like AI that
has the potential to transform a number of aspects of a firm’s value appropriation strategy.
METHODS
Research Context
Contract Intelligence (CI) produces AI enabled legal technology to help facilitate faster
and more accurate contract review. The company was founded in 2015 and has its headquarters
77
overseas with a large office in New York as well. The company has grown extensively since
2015 and has received multiple rounds of venture capital funding. CI has aggressive plans for
more growth and as such strives to continually improve their product and customer
experience. In our first interview with CI’s chief marketing officer and the head of customer
experience, we discussed how they were concerned about customers licensing CI’s product and
then, after a few months, gradually stopping usage. CI’s CMO was not sure what was going on –
or why these customers who were seemingly so enthusiastic about the product ceased usage a
few months later and then failed to renew – but he did know that there was a lot of variation
amongst the customer firms. Indeed, CI’s customer firms range from large, Fortune500
companies to small, five person law firms and vary across a range of industries. See Table 1.
CI’s technology uses machine learning (ML) to predict what clauses belong – or do not
belong – in a routine contract. In order to do this, CI has compiled a database of routine contracts
from public and private sources to train its machine learning algorithm. Routine contracts
are generally defined as those with low variation in the contract even despite different scenarios
(Rich et al., 2020). Examples include non-disclosure agreements, master sales agreements, and
SaaS agreements. Indeed, CI’s technology can accommodate all these types of agreements, in
addition to other agreements such as product supply agreements, service agreements, software
license agreements, statements of work, purchase orders, and distribution agreements. Based on
their database of contracts, CI is able to predict what clauses might belong in an NDA, for
example. As part of the onboarding process, CI customers identify which contract types they
want to use CI for. Then, they work with CI to come up with a playbook. The playbook serves as
a key for CI’s technology. Essentially, the machine learning algorithm uses its learning of what
clauses normally appear in an NDA, in addition to the clauses identified in the customer’s NDA
78
playbook, to determine whether or not the clauses in a particular NDA from a third party (for
example), fit with the customer’s standards for what clauses an NDA should contain.
CI’s product is primarily intended for any firm with a large volume of routine contracts
that need to be reviewed. Firms must devote resources to contract review before executing the
contracts and for firms with a large volume of contracts, even those that are fairly routine, such
as procurement contracts or NDAs, it can be incredibly burdensome to review all these contracts,
especially when there are service level agreements (SLAs) that dictate the timeframe within
which the contracts must be returned. While CI does have some customers who are law firms,
the bulk of CI’s customers are large to mid-size companies.
Data and Participants
Our data from CI includes the audio from seventy-five onboarding calls; usage and
renewal data for the corresponding firms; eighteen interviews with CI executives and employees,
as well as email correspondence to follow up on questions from interviews. We triangulated
between these data sources to gather insights about how each firm’s onboarding
process proceeded from start to finish, and also to get a picture of any issues that may have risen
between the completion of onboarding and renewal or non-renewal (see e.g., Davis &
Eisenhardt, 2011).
Onboarding Calls
CI provided data on a total of 30 of their customers. Unfortunately, because CI is a
startup, their internal processes for organizing and logging onboarding calls was not well
established initially. Therefore, CI only had records of onboarding calls for which the customer
success manager remembered to save the record. CI has more than 30 customers, however calls
from the remainder of their customers were either improperly logged, inadvertently deleted, or
79
never recorded at all. Because CI has such an involved onboarding process, CI has multiple calls
with each customer before launching CI’s product to review contracts. Therefore, CI shared with
us seventy five onboarding calls from thirty different customers. There was variation in the
number of calls per customer, ranging from one call per customer to five calls per customer. For
the customers for whom we have more than one call, it allowed us to see how attitudes towards
onboarding and towards CI’s products changed over time.
CI’s onboarding process starts, once the customer has signed the contract, with an
introduction or ‘kickoff’ call between CI and the customer’s representatives. For twenty eight of
the thirty customer firms, the customer’s representatives who were to use the product joined the
onboarding calls. In two of the firms in our sample the people on the onboarding call(s) were not
those who were to use the product. In one case, the person on the call was an attorney/manager
who was attempting to get the product off the ground for paralegals to use. In the second case,
the person on the call was an attorney who was tasked with onboarding the product so that other
attorneys (her colleagues) could use it.
After the kickoff call, the customer and CI schedule a series of ‘playbook
validation’ calls. Generally, the same CI customer service specialists join all the onboarding
calls, from kickoff to go-live checkpoint. In between onboarding calls, the customer has
‘homework’, which entails the users using CI’s product to validate the playbook and ensure that
the playbook captures all the parameters and standards particular to the customer’s needs. Once
the playbook has been validated and the customer is satisfied that the results from CI’s contract
review successfully flag all issues for a given contract type, CI and the customer have one final
call in which they check in to resolve any outstanding issues (called a ‘go live checkpoint’).
Typically, each of these calls – the kickoff, playbook validation calls, and the go-live checkpoint
80
– are all scheduled about a month apart, depending on the specific customer’s needs and
schedule. Additionally, these calls are very detailed. The average number of minutes per call was
72 and in interviews with CI we confirmed that each onboarding call was typically very in depth.
The onboarding call dialogue is relatively consistent from the CI side. That is, during kickoff
calls, CI asks their customers the same questions; during playbook validation calls, they stick to
roughly the same script for each customer; and during go live checkpoint calls, they also stick to
roughly the same script. This allowed for further identifying variation between the calls because
the CI script was held relatively constant.
Data from CI
CI also provided us with their customers’ usage and renewal data. CI provided data on
each customer’s individual users in terms of their uploads: what type of contract was uploaded,
when, the turnaround time on CI’s end for the review, number of words in the contract, and
which playbook it used. At the firm level, we also had data on whether or not each customer
renewed their contract and the amount of that contract. This data from CI allowed us to
determine usage trends for each customer, in addition to providing an outcome variable that
spoke to a longer-term adoption metric.
Data on Customer firms and Users
We also collected data on CI’s customers and the individual users at these customer
firms. We collected this data from Compustat and Linkedin, including firm size, SIC code,
founding date, sales volume, number of employees with a legal function (attorneys, paralegals,
contract managers, etc.), and the publicly available LinkedIn profiles of users we could identify
from the onboarding calls.
Interviews with CI Employees
81
We initially connected with CI to discuss their stated problem of why their customers
were adopting their software, seemed to be happy with it, but then stopped using it, and failed to
renew. After discussing this with other firms with contract review products that used AI and with
larger firms that attempt to sell enterprise level AI solutions, we realized that this is a common
problem amongst firms that develop and/or sell AI technology. Because the output from the AI
and the benefits of the AI take longer to see and are harder to measure than with
other technologies (i.e., the internet) it is more difficult to incite immediate adoption. However,
in the meantime, these producer firms have a viable product – and customers on hand. Yet, the
particular adoption challenges faced here often lead to this tension of initial adoption, then a
failure to integrate the product into the company’s business model. Recognizing that this
problem occurred with other companies outside of CI and outside the legal technology
space affirmed the study’s broader applicability.
We conducted initial interviews with CI employees and executives to discuss this tension.
Initially, we interviewed executives at Contract Intelligence about their process in onboarding
and their general product and company strategy. Specifically, we spoke with the Director of
Customer Success, Head of Product, VP of Customer Success, Chief Marketing Officer, and the
Director of Strategic Solutions multiple times at this stage.
Analysis
Stage 1: Open coding and iterative theory building
After the initial interviews with CI that provided a better understanding of the context and
the core tension, we began an iterative process of open coding. Through listening to the
onboarding calls, conversations with CI, and reading the call transcripts, we iterated between
theory and data and began to code variables and build theory about the customer firms. We
82
listened to all the calls once through, then listened again and took detailed notes. Then we spoke
to CI to discuss some of our findings and confirm some of what we were thinking about the
customer firms’ onboarding calls conformed to their understanding of their business and their
customers. We had the calls’ audio professionally transcribed, while maintaining the privacy of
the data. With the seventy-five transcripts, we were able to read and engage with the content
more thoroughly. We continued to iterate between reading each transcript multiple times,
taking an additional set of notes, and talking with CI. We also had an independent research
assistant read and take notes on the transcripts. We then compared each set of notes and re-read
parts of the transcript where the notes diverged in terms of content and coding. We continued to
discuss with CI during this process as well.
During this process, we began to group the customer firms into categories. We based
these categories on what we observed the customer firms discussing during
their onboarding calls. We found that while there was variation in how each customer discussed
these topics, the customers tended to consistently discuss two critical points: (1) readiness to
adopt the technology and, (2) specificity in the customer’s purpose for the technology. We define
readiness as how eager the firm is to use and integrate the product. We defined specificity as the
customer’s engagement with the product in a detailed rather than abstract manner. For
example, we noticed that many of the onboarding calls stuck to similar topics of conversation,
despite saying different things in those topics.
As we developed the four categories and grouped customers into the categories, we
held frequent discussions with CI to see if these categories made sense to them. We also talked
with CI to glean additional information about the customers that aided in clarifying additional
context for conversations or questions in the onboarding calls.
83
Readiness
During the onboarding calls, eagerness and willingness to use and integrate the product
varied between firms. We observed that firms talked about readiness differently. However, there
were some themes that were evident across all calls. Firms either clearly trusted the product or
did not trust it; sometimes this was expressed as a lack of trust in AI as well.
Firms addressed their need – or lack thereof -- for adopting the product in some respect or
another. Additionally, firms actively expressed either a degree of risk aversion or a lack of risk
aversion towards using the product.
For example, one firm stated: “We don’t think the playbook is ready yet and we don’t
trust [CI] yet. We are doing a lot of manual review still because there’s just not a lot of trust
there on the technology”
1
. Similarly, another stated, “We do not feel confident in AI review [of
our contracts]”
2
. On the other hand, some firms began to trust the product more: “I feel
more comfortable with the whole process. I'm now understanding it more. As always, well, you
feel a little uneasy when you're not used to change. And what I've done is, it's taking me a little
longer than usual because I mean, just trying to get ahead of what's going on and still trying to
copy paste instead of I guess, freestyling with typing. I'm trying to get out of that rhythm and I'm
still finding myself where I don't want to take the whole entire sample, I just want to put in these
few little keywords. So, it's just a matter of trying to get still used to it. What I have been doing,
I would, of course, use [CI], I'm uploading everything now. As far as the client’s template, what I
do is I go through it by myself to see if anything was missing so I could at least try to make sure
everything is captured before I give it to the client. So, it takes me a little bit longer than if I
would do it myself. But like I said it's just a process that I have to get used to. It's only a couple
of things that I've caught but it's not like a list like I had before. It's not that big of a deal. It's just
84
little things. And like I said I could be nitpicky right now but it's not that bad. It really isn't. It's
gotten easier but like I said, everything takes time with change.”
3
We observed that some firms had a specific need for CI’s product while others did not.
For example, a firm that we considered as having a specific purpose for the product described:
“We receive a lot of NDAs, in particular, and we’re keen to give this a whirl to see if it can help
reduce the amount of time and energy that we expend on NDA reviews.. We’d like to get this up
and running as soon as possible too because – yeah – there is a need.”
4
Another firm described:
“The hope is it [adoption of CI’s product] will alleviate or give us some more capacity to deal
with the higher level strategic matters where we feel like we could add more value. So in terms
of focusing on contracts when messaging up to the global general counsel and executive team, I
guess the focus isn’t there. It’s more this is how we deal with matters of this transactional nature
until we get more efficient at doing that. Then we can build in some capacity to deal with more
of the strategic stuff that we see that adds more of a value to the business.”
5
Customers had different perceived needs for the product. Some customers had a very
specific need, for example stating, “some of the challenges, obviously, there’s been an increased
volume of IT contracts and just an overall rising of the need for contract review and delays in
delivering key services, a perception as legal as a bottleneck, which we see that all the time
and something everyone’s really trying to kind of change around. And then sometimes key
clauses get missed. And so some of the success, more importantly, the success we’ve had to date,
is a reduction in turnaround time, an increase in consistency, just because everyone’s following
the same playbook, picking up concepts that could be missed in a manual review, and then
reducing the city attorney's time of final review.”
6
While other customers did not have a specific
need: “The NDAs, we do a lot of them, but whenever we bring this up with the contracts team,
85
their response to me is ‘Well, we push our form NDA,’ in which case the sort of value
proposition is significantly diminished. So I think there’s still value in setting up and using an
NDA playbook. I just don’t know what volume of NDAs we would be able to go through but
because that is one that you guys sort of pushed and highlighted as a very sort of efficient
contract type to review, we should still do it. And then at some point, we could also revisit if
we’ve got the perfect playbook for NDAs in place, do we need to be using our – do we need to
push our form contract, right, because you’re basically achieving the same thing by having the
playbook reconfigured and likely shortcutting the review process.”
7
Other customers discussed need, yet failed to pinpoint a specific purpose for adopting the
technology, instead speaking generally to a number of reasons: “We're obviously a small team.
The work is growing. From what I can see, the lower value, more
administrative contracts are coming in increasing in number. The volume's not so much that we
can't handle it ourselves -- it feels like a couple of things. One, as an organization we want to be
early adopters into kind of legal tech. Secondly, like I say, [redacted name] and myself, we're
very experienced lawyers. It's not efficient for us to be spending time on kind of low-value, more
administrative contracts. So any tools that can assist us in really automating that process is good.
And that's really kind of where we're coming from. It's not that we can't really manage these
contracts right now. And I'm definitely seeing on the non-NDA side a lot more kind of
technology, services agreements, and there pretty much is no reason why we shouldn't be feeding
an engine against a playbook and automating that review because it's not that difficult. So really
that's what I--and I know we're in the kind of entry-level plan. We only have 10--I think we have
200 contracts a year. So we're starting slowly in terms of adopting this, but we could bump up
against 10 docs a month without too much trouble in the not too distant future. So we really just
86
want to get the--I don't want to say pain, but I think there is a certain amount of pain and
testing in just getting this up and running.”
8
Some customers did not discuss need but it was implied from their lack of discussion of it
that there was not a specific need: “We do have to check with the timeline because the problem
is, I'm going to have to-- and you're not privy to this, but my next piece of what we might want to
shift in our agenda is just because I have to go through semi-unwilling attorneys and hold their
hands and shepherd them through this sort of one at a time, to revise – I have a lot of revisions
that I’ve been storing up that once I see this in your playbook form, I can edit it there and polish
it up and make comments for our attorneys, and then I have to shepherd them through it over the
next month or two. And it’s just that I don’t know what the end dates are but those will have to
be flexible. Even if we just have some milestone dates that are inflexible, it will be very hard to
move all the pieces. So I just want to put that out there.”
Customers’ risk aversion could often be gleaned from the how they talked about their use
of the product and questions they asked. For example, if the customer talked about “As long as
[the contract] isn’t very one sided, we generally just skip through it. But again, all of this stuff is
– it’s all green thumbs, so as you said, it’s just a matter of us kind of just learning to trust the
AI, and so we can come to grips with that”
9
as compared to “Criteria number one [in terms of our
expectations] is that the product comes back to us and meets our expectations in terms of
accurately reflecting our playbook and not exposing us to any potential risk”
10
it was clear that
the customer was risk averse to using the product. In all cases, we confirmed these
categorizations with CI who agreed that this framework conformed to their understanding.
Some firms were more risk averse when considering using the product. For
example, some firms expressed reservations with regard to using the AI: “We feel – we feel a
87
little bit like the guinea pigs and that the software just maybe isn’t-- it just isn't completely
doing certain things”
11
Another firm expressed similar risk averse sentiments: “We keep hearing
from you guys that when in fact something that's wrong or doesn't flag something that it should,
that it gets sent to the technical group and it's going to get taken care of. Or we hear that that will
develop over time and it will get better at that. But I guess, that begs the question then, what are
we to do in the meantime? I feel like we're sort of guinea pigs and we're paying for the service,
but it's not fully baked yet. That isn't what we were told when we signed up. That's what's
leading to this lack of confidence when we do use it. I just don't know how it's saving us time if
we don't have any realistic expectation of having a high level of confidence in what it can do in
the near future. By the near future, I mean the next week or two.”
12
On the other hand, other firms were less risk averse and seemed to eagerly use the
product despite technical issues. For example, one firm explains a technical issue that occurred
as CI updated their user interface: “We went through a couple contracts in the system and they
didn’t seem to have any formatting issues. It looked like one perhaps was under the old system
but we still saw the thumb. And then we saw one that looked like perhaps it was uploaded in
the new system because it had no thumbs in the contract. Is that a new change?” and CI
responds, “Yes, that is how the old contracts were uploaded – with the thumbs. But currently
when you click on the name in the concept it will highlight the area in the contract without
having you thumbs up and thumbs down.” “Yes, that’s what we thought. That’s great.”
13
This
exchange demonstrates how some firms accepted and moved past technical issues during use.
Specificity
It was also clear that specificity, or the customer’s depth of knowledge about the
technical aspects of the product and the customer’s sophistication about the specific domain, in
88
terms of use, product, and technology, was frequently discussed. When considering
specificity, we considered the complexity and depth of the customer’s questions about the
product, both in terms of technical aspects of the product and in terms of how sophisticated the
customer was regarding its own capabilities for technology adoption and use, and its overall
sophistication in terms of AI adoption.
Some customers asked questions that clearly demonstrated their technical knowledge
about AI and CI’s product. For example, in this call, a playbook validation call between Psi and
CI, representatives of Psi discuss with CI how to approach a certain clause:
Psi: So I notice sometimes you’ve got required and acceptance criteria for something.
CI: Yeah. So for instance, in the effective termination, we have a specific of the concept,
which deals with payments and refund. And I didn’t have a chance to read through that wording
so fast, but if you have a certain revision regarding payments and refunds that have to do with
effective termination, then this will give you an AI-specific result of this concept.
Psi: Oh, yeah yeah yeah. So if you’ve got refund and payment, let’s just keep the
effective termination to things outside payment and refund. Is that right? What is the payment
legal concept about? What’s the AI looking for?
CI: For instance, right, you terminated your agreement with your contractor but you
terminated it, I don’t know, in the middle of the month, for instance, and you have a monthly
payment.
Psi: So this is payments in accordance with termination – it’s not payments general?
CI: Exactly. This is specific –
Psi: How do I know that? Because is that why the blue bar is at the very left? Does that
mean it’s part of termination?
CI: Exactly
Psi: Gotcha, okay.
This conversation excerpt demonstrates technical knowledge in that Psi is discussing how the AI
will review the contract. It also demonstrates Psi’s domain specific sophistication about the
product: these specific questions, which would only result from Psi’s actual use of the product,
show that Psi understands that the purpose of onboarding is to fine tune the playbook so that it
can best serve Psi’s needs. Indeed, many customers asked questions that conveyed their technical
knowledge and domain specific sophistication with regard to this product in particular: “If a
clause appears twice that relates to the same topic, does the AI flag it twice?”
14
This technical
89
question implies that not only has this user engaged with the product substantially in order to
identify this as a potential issue, but also that the user understands the underlying technology
enough such that she can ask a relatively technical question about how the product’s algorithm is
set up.
On the other hand, some customers asked basic questions about the product’s elemental
functions such as, “Sorry, could you kind of – I know we’ve already touched on this a couple of
times, but could you maybe talk through how – so when it determines that [the clause] is there
but it’s not being captured appropriately, why do I see a red thumbs down?”
15
. These questions,
while they may be specific, relate to much simpler aspects of the product. In this case, the firm
had some sense of how the AI worked – some technical knowledge – but very little domain
specific sophistication in terms of how the product actually functions. In some cases, a
customer’s question conveyed a lack of knowledge about AI and in other cases, customers
asked CI specifically for information about AI: “Can you explain to us a little more about how
the AI here actually works? Like what does it do?”
16
Additionally, some firms demonstrated a lack of technical knowledge and domain
specific sophistication. These firms had little idea what they had undertaken with regard to the
capabilities of the product or the expectations required for successful onboarding. For example,
when one firm encountered a formatting issue with converting their draft from a PDF to a word
document, the following exchange occurred:
CI: So this sometimes, during the OCR process – formatting may change a little, but this
definitely should not happen. I would like you to share with me the links to both these documents
and then I’ll forward them to our product team to see what’s causing the issue.
Mu: How do I do that? Share?
CI: You can just highlight the link, copy it, and email it. And you can also – when you
click on share, it has the same functionality.
Mu: I see this feedback thing over here, perhaps, is that something?
CI: Sure, yeah, you can send it directly there and I will receive that.
Mu: Okay
CI: You have to click send.
Mu: Sorry. . . I think I am a good test, to be honest, because I am a complete idiot when it
comes to IT.
90
Other firms were more aware of their own capabilities but lacked knowledge about AI
that would be required to have a more seamless onboarding process. For example: “We looked at
this as more of an opportunity for us to do something modern and be more efficient in terms of
our operations, stop doing things in an old fashioned way. But I’ll be honest, we are not
technical people, we do not know what really this app, this product can do, especially when it
comes to AI. We just have no idea – we’re just a bunch of lawyers, we just want to do this faster
and potentially for other contracts too.”
17
Categories
After iterating through multiple rounds with CI and the transcripts, we found that firms
were either high or low along each dimension (readiness and specificity). Given that need, risk
aversion, trust made up the readiness dimensions and domain specific sophistication and
technical knowledge made up the specificity dimension, we focused on what emerged with
regard to firms in these categories. See Figure 2.
Firms that had low need for the product, high risk aversion towards adopting the product,
and low sophistication with regards to the capabilities of the product were considered to be only
adopting the product in order to be perceived as innovative or technologically progressive. For
example, these firms tended to state explicitly “we want to be seen as tech forward” or “our
management wants to be perceived to be innovative”. In more subtle ways, firms that clearly had
little to no knowledge of the product or of AI generally, yet who seemed enthusiastic about the
idea of the product (“this seems like something we should do”) were also categorized as firms
that were adopting for reasons related to perception, either by customers or managers. We
categorized these firms as status seekers.
91
Firms that had a moderate need for the product – or in some cases did have a need and in
some cases did not have a need – and had high risk aversion, and low sophistication tended to be
those that wanted to experiment with the product rather than actually use it. They were interested
in the product’s capabilities but because of their low sophistication about the product’s
technology and, in some cases, their own lacking technical capabilities, were unable to
fully adopt the product. For example, firms that discussed a desire to “play around” with the
product or talked about how they “were not quite ready” to use the product in a serious way were
put in this category. Additionally, firms that explicitly stated “we don’t have a pressing need” for
the product or “we want to see if this is right for us first” were in this category. We categorized
these firms as tinkerers.
Firms where the need for the product varied – either high, medium, or low – there was
high risk aversion, and high sophistication for understanding the product’s capabilities were
likely to want to seriously adopt and use the product but only after it has been tested in the
market. Essentially these firms want to wait until the product has gained more traction amongst
competitors and in the market more broadly before adopting. For example, firms in this
category asked technical questions about the product but then balked when they were told by CI
that many of the issues would be resolved by using the product. Additionally, these firms often
had a need for the product but were “too busy” to use it now and wanted to postpone use. We
categorized these firms as fast follower.
And finally, firms where there was a high need for the product, a high risk aversion but
also a high level of sophistication are serious about using the product in that they recognize that
some risk aversion towards adoption is probably warranted, but given their need and their
understanding of the technology capabilities and their own capabilities, are planning to actually
92
use the product. From the start, these firms expressed optimism about the product’s capabilities
and were open to an involved onboarding process. For example, some firms in this category
explicitly stated, “we know that this [onboarding process] will take a long time and that’s ok”;
others asked detailed questions about the product that would only arise from heavy use and
anticipated future use. We categorize these firms as serious users. See Table 1.
Stage 2: Propositions and Preliminary Test of Categories Using Renewal Data
We then developed propositions about the long-term adoption prospects for each of the
four types of firms. Given the data that we have on renewal rates, we conduct a preliminary test
of the propositions. It is important to note that given a number of factors – in particular, small
sample size and potential for latent variables – we do not mean for this preliminary analysis to be
dispositive; it should merely serve as a starting point for future empirical or experimental
analysis using the framework developed herein. See Figure 3.
Tinkerer
Firms that are not ready to fully integrate the technology into their firm but that remain
excited about experimenting with the technology are classified as tinkerers. These firms are
likely to have a low need for the product, at least immediately, and also have a low level
of sophistication about what capabilities the product has and what capabilities they have as a
firm. These firms typically have little risk aversion to using an automation tool, yet they have
little motivation to actually use it because they do not have a pressing business need. These firms
do not know enough about the product and need to spend a significant amount of time
experimenting to understand how it works and its utility to the firm.
Proposition 1: A firm’s desire to experiment to see how the AI technology works and to
see if it may be useful for the firm at some point will result in a failed adoption.
Status seeker
93
Firms that want to adopt the AI tech purely for appearances and that do not care about the
capabilities of the AI product are categorized as status seekers. These firms may adopt the
technology because they want to impress – or have a mandate from – a manager, or they may
adopt because there is a perceived market value in being “tech savvy”. It follows that these firms
have a low need for the product in terms of its actual capabilities. Many of these firms do not
know quite what CI’s product actually does. Yet many of these firms still had a high risk
aversion to using the product, often alluding to a sense that automation would replace them.
Proposition 2: A firm’s desire to appear innovative without an intent to use
the technology (status seeker) will result in a failed adoption.
Fast follower
Firms that want to adopt the AI product due to competitive pressures but are unsure about
fully integrating it and want others to do so first to test risks are considered fast followers. These
firms generally have a low need for the product , except that they feel competitive pressure to
adopt. They have high risk aversion to the product based on uncertainty about the product’s
capabilities or, in some cases, the impact of AI on labor markets. Additionally, these firms have a
high level of sophistication regarding the product’s capabilities and their own capabilities –
this spurs their interest in the product and some knowledge of need, but they still remain
uncertain about the risks. Additionally, these firms generally fail to grasp that implementation is
harder than licensing a product; it requires effort in onboarding and integrating the product into
the firm.
Proposition 3: A firm’s desire to understand the technology but with the goal to only
use it once it is more generally accepted will result in failed adoption.
Serious user
Firms that understand the capabilities of the product and are ready to use the product long
term, integrating it into firm strategy and process, are considered serious users. These firms have
94
a high need for the product and low risk aversion towards using and integrating this technology.
Additionally, these firms have a high level of sophistication in that they have a command of their
own capabilities, the product’s capabilities, and the capabilities of AI as deployed in products
like this.
Proposition 4: A firm’s desire to implement the technology fully in the near term will
result in full adoption.
Analysis
We used data from CI on firm renewals of the product license to conduct preliminary
tests of these propositions. All thirty firms in our set initially adopted CI and signed
a yearlong contract to be renewed if desired. We considered that, given the price of licensing
CI’s product, firms that were not using the product in a meaningful way would not renew the
product for a subsequent year. Therefore we distinguish between full or comprehensive adoption
and adoption, the latter referring to merely licensing the product for a yearlong contract and
failing to integrate it into firm strategy and processes, ultimately resulting in a failed renewal. In
interviews with CI, we discussed with them what they considered to be a “good” renewal rate, or
one that would be acceptable to their investors. Generally, they said they expected to see a
renewal rate of around 60%, given that the technology – and the company – is new. Therefore
we used a renewal rate of 60% or above to indicate a successful adoption and anything below
60% was considered a failed adoption. Given that our sample was thirty firms, we used the
classification into the four categories to determine basic renewal rates. See Table 3. Firms
classified as tinkerers had an average renewal rate of 43%. Firms classified as status seekers had
an average renewal rate of 67%. Firms classified as fast followers had an average renewal rate of
50%. Firms classified as serious users had an average renewal rate of 90%. Overall this provides
support for Propositions 1, 3, and 4. There was no support for Proposition 2, regarding status
95
seekers as unlikely to renew, because status seekers had a renewal rate of 67% which is above
CI’s internal standard of a desired 60% renewal rate.
Discussion
The preliminary findings regarding renewal rates show that producer firms such as CI can
learn valuable lessons from categorizing and understanding their customer firms’ capabilities and
readiness to adopt AI. While our findings here may not be generalizable due to the small sample
size, it does provide support for increased analysis of how firms respond to the onboarding
process. For example, if a firm seems to be expressing a desire to experiment with the product,
the producer firm may want to devote more resources to ensuring that this customer actually
integrates the technology into its processes to ensure renewal.
We also looked at usage trends amongst the four categories of firms to see if a firm’s
continued usage of the product over the course of the first year of licensing correlates with the
likelihood of renewal. Overall, average usage of firms classified as tinkerers declined over the
first year. Average usage of firms classified as status seekers declined over the first year.
Average usage of firms classified as fast followers increased over the first year. And average
usage of firms categorized as serious users increased over the first year. It is interesting to
note that the firms classified as fast followers ultimately began using the product but then did not
renew. While we do not have the ability to identify or test a mechanism for this observation, we
hypothesize that this trend is due to these firms realizing a poor product fit with their
organization and needs, and possibly, equipped with the capabilities to properly evaluate AI,
finding another AI product.
Our data also includes two rounds of a customer survey about adoption, CI’s onboarding
process, and perceived utility of the product, distributed to all CI’s active users. One round was
96
completed in spring 2019. We hope to capture responses from firms in both rounds of the survey
in order to compare how attitudes towards the product change over time.
Going forward, we plan to incorporate this customer survey data and the data we have
collected on the customer firms’ individual users. We hope that these data will add additional
nuance to the categories developed and help us further develop the theoretical contribution.
Overall, this paper attempts to build theory around the central tension of why firms
adopt an innovative product – in this case AI contract review software – and then fail to fully
integrate it into their firm. In doing so, we classify the firms in our sample into four categories
based on an iterative theory building process that uses interviews with CI, onboarding call
transcripts, and data about the firms. Importantly, we identify two key dimensions, specificity
and readiness, that drive whether or not a firm is likely to fully adopt the AI product.
We draw on the unique nature of AI as a GPT and as an innovation that, while potentially
transformative, can take time to implement and benefit from. Given AI’s unique properties, we
discuss how firms looking to adopt an AI enabled product in even a limited context must have
the capabilities to do so. The firm’s capabilities, in this context, include its understanding of
AI, its understanding of its own technological knowledge, and its command of how to
strategically implement AI for a specific purpose. We distil these capabilities down into two
dimensions: readiness and specificity. Readiness refers to the firm’s trust in the
technology, ability to identify a specific organizational need for the product, as well as the firm’s
risk aversion to adopting a new technology. Specificity refers also to the firm’s technical
knowledge of the AI and domain specific sophistication about the product and also the firm’s
own capabilities in adoption and use.
97
Firms expand their core capabilities, meaning their knowledge and skill base, and
improve their ability to assimilate and use future information, or absorptive capacity, through
investment in technology development and its associated learning. A firm’s core capabilities
refer to its ability to process and deploy content or new knowledge, while a firm’s absorptive
capacity refers to a firm’s ability to incorporate new processes into firm strategy. Both of these
are considered components of firm learning. The literature on technology adoption generally
refers not only to a dichotomous choice to adopt or not, but also often focuses on the producers
or developers of new technology at the expense of the users.
We also recognize that the users of this technology, in adopting the product, provide
feedback to the producer, CI, about the quality of the product. Throughout the onboarding calls,
CI incorporated feedback from customers into product updates. CI updated the product based on
customer input to appease the customers but also to make the product better. The role of
internal versus external factors that influence technology adoption have been noted. External
factors include network externalities, customers’ choice of technology based on installed base
and complementary goods, barriers posed by patents, to name a few (see e.g., Schilling, 1998).
Internal factors include a firm’s attributes and strategy, for example. Here we suggest that it is
important to emphasize the value of customer technology adoption when thinking about adoption
of AI. Because AI requires a large amount of data in order to produce results for the
customer and also, often to optimize the functioning of the product, the internal and external
factors that impact technology adoption are more intertwined. The result of this, in the short
term, is that producer firms like CI are devoting substantial resources for onboarding in an
attempt to provide all customers with a baseline of understanding about the product. However,
producer firms could instead deploy their resources more strategically by understanding the
98
complementarities in the external and internal factors that impact technology adoption. For
example, CI’s customers’ choice of technology, based on complementary goods, will be
influenced by CI’s own strategy in designing the product and how they decide, for
example, to build in a feature to the product that enables the AI to learn and improve with
additional use (which CI eventually did, given customer requests). The overlap in external and
internal adoption factors implies that there is a significant interdependence between the producer
firms and the customer firms. The producer firms should spend more time strategically
evaluating and tailoring onboarding based on the particular characteristics and capabilities of
their customers, rather than trying to ensure all customers are at the same level.
Conclusion
Firms with an AI product, like CI, should emphasize different strategic approaches for
different customers given these interdependencies. For tinkerers, CI should emphasize the
specific value of the product for the customer and work towards full implementation. For status
seekers, CI should leverage external urgency and emphasize that status comes via full adoption.
For fast followers, CI should emphasize the product’s capabilities, including function and added
value. And for serious users, CI should develop a tailored technical roadmap for usage.
The variation in customers who are interested in adopting AI products, rather than being
seen as a disadvantage for producer firms, can be leveraged to the firm’s advantage by allowing
it to strategically deploy onboarding resources. By understanding customers’ capabilities
upfront, producer firms may be more successful in ensuring that adoption lasts beyond a year.
While this work develops grounded theory to develop a categorization of customer firms who
have adopted an AI product, we recognize its limitations. We hope that future work will continue
99
to explore this topic and, in doing so, will contribute to the burgeoning literature on AI adoption
and innovation within firms.
Figure 2. Conceptual Model
Figure 3. Firm Categories.
Readiness (Low) Readiness (High)
Specificity (Low) Status Seeker Tinkerer
Specificity (High) Fast Follower Serious User
100
Table 5. Customer Firms.
101
Table 5 (continued).
CONCLUSION
Firms that invent technology traditionally have either pursued a cooperative strategy, in
which they license the technology, or a competitive strategy, in which they take the technology
to market. An extensive literature discusses technology commercialization strategies, focusing on
the inventor firms and how the choice to cooperate or compete yields more value.
Similarly, when firms adopt a new technology, they rely on individual users to actually
interact with and use the technology, and in so doing, to ultimately create or capture value for the
firm. An extensive literature on technology adoption addresses how individual preferences can
drive technology adoption at the individual level. However, the mechanisms that support how
102
individual level technology adoption preferences translate to a firm’s integration and sustained
use of the technology are less clear.
Thus, in this dissertation I seek to first understand and reconcile the literatures on
technology commercialization and technology adoption. I explore where there may be overlap or
opportunities for additional research. For example, the literature on technology
commercialization primarily discusses commercialization strategies from the inventing firm
perspective. The literature discusses why a firm may license a technology, but thus far there is no
work that delves into the individual level mechanisms that motivate a firm to license a
technology. I propose that the licensing firm’s perspective is equally important, and can be best
understood by examining the micro-mechanisms that motivate a firm’s decision to license and
renew a new technology, particularly one that is ML-enabled. Further, I propose that despite
heterogeneities across licensing firms, a licensee firms can be categorized, based on their
readiness to adopt and their sophistication about the technology. Categorizing licensee firms can
help inventor firms adapt and hone their commercialization strategies for new technologies,
especially technologies that use ML. Because ML can be both a product and a process
technology where its value and capabilities increase with usage, it is critical to examine
mechanisms that support or inhibit individual level usage, and how this may affect firm adoption.
I also hope that this dissertation will open the door to future research on the technology
commercialization and adoption of ML, particularly in terms of how individual level
mechanisms can impact firm level strategies, and how employee and manager interactions can
affect technology commercialization and, potentially, how the technology develops.
103
References
Acemoglu D, Autor D (2010) Skills, Tasks and Technologies: Implications for Employment and
Earnings. NBER Work. Pap. Ser.
Acemoglu D, Restrepo P (2018a) Artificial Intelligence, Automation and Work
Acemoglu D, Restrepo P (2018b) Low-Skill and High-Skill Automation
Acemoglu D, Restrepo P (2019) Automation and New Tasks: How Technology Displaces and
Reinstates Labor. J. Econ. Perspect. 33(2):3–24.
Acs ZJ, Audretsch DB (1987) Innovation, Market Structure, and Firm Size. Rev. Econ. Stat.
6(4):567–574.
Adner R (2002) When are technologies disruptive? a demand-based view of the emergence of
competition. Strateg. Manag. J. 23(8):667–688.
Adner, R., & Kapoor, R. (2010). Value creation in innovation ecosystems: How the structure of
technological interdependence affects firm performance in new technology generations. Strategic
Management Journal, 31(3), 306-333.
Aghion P, Reenen J Van, Zingales L (2013) Innovation and Institutional Ownership. Am. Econ.
Rev.: 277–304.
Aghion P, Tirole J Formal and Real Authority in Organizations and European Bank for
Reconstruction and Development
Agrawal A, Gans J, Goldfarb A (2019) Artificial Intelligence, Automation, and Work. Econ.
Artif. Intell.:197–236.
Agrawal A, Gans JS, Goldfarb A (2018) Prediction machines: the simple economics of artificial
intelligence (Harvard Business Press).
Agrawal, A., Gans, J. S., & Goldfarb, A. (2019). Exploring the impact of artificial intelligence:
Prediction versus judgment. Information Economics and Policy, 47, 1-6.
Agrawal, A., J. McHale, and A. Oettl (2019). “Finding Needles in Haystacks:
Artificial Intelligence and Recombinant Growth,” in A. Agarwal, A. Goldfarb, and J.
Gans (eds.), The Economics of Artificial Intelligence. Chicago, IL: University of Chicago
Press.
Anderson EG, Joglekar NR (2005) A Hierarchical Product Development Planning Framework.
Prod. Oper. Manag. 14(3):344–361.
Anderson P, Tushman ML (1990) Technological Discontinuities and Dominant Designs: A
104
Cyclical Model of Technological Change. Adm. Sci. Q. 35(4):604–633.
Argyres NS, Silverman BS (2004) R&D, organization structure, and the development of
corporate technological knowledge. Strateg. Manag. J. 25(89):929–958.
Baker G, Robert Gibbons MIT N, Kevin Murphy NJ Informal Authority in Organizations
Barney JB (1991) Firm Resources and Sustained Competitive Advantage. J. Manage. 17(1):99–
120.
Blei DM, Ng AY, Edu JB (2003) Latent Dirichlet Allocation.
Brandenburger, A. M., & Nalebuff, B. J. (1995). The right game: Use game theory to shape
strategy (Vol. 76, pp. 57-71). Chicago: Harvard Business Review.
Bresnahan, T. F., & Trajtenberg, M. (1995). General purpose technologies ‘Engines of
growth’?. J. of econometrics, 65(1), 83-108.
Brynjolfsson, E., Mitchell, T., & Rock, D. (2018). What can machines learn, and what
does it mean for occupations and the economy?. In AEA Papers and Proceedings (Vol.
108, pp. 43-47).
Brynjolfsson, E., Rock, D., & Syverson, C. (2018). Artificial intelligence and the modern
productivity paradox: A clash of expectations and statistics. In The Economics of
Artificial Intelligence: An Agenda. University of Chicago Press.
Cabral JJ, Francis BB, Kumar MVS (2020) The impact of managerial job security on corporate
entrepreneurship: Evidence from corporate venture capital programs. Strateg. Entrep. J.
2(1)1357-1372.
Cabral, R., & Leiblein, M. J. (2001). Adoption of a process innovation with Learning‐by‐
Doing: Evidence from the semiconductor industry. J. of Indus. Econ., 49(3), 269-280.
Chao RO, Kavadias S (2008) A Theoretical Framework for Managing the New Product
Development Portfolio: When and How to Use Strategic Buckets. Mgmt. Sci. 54(5):907-
921.
Chao RO, Kavadias S, Gaimon C (2009) Revenue Driven Resource Allocation: Funding
Authority, Incentives, and New Product Development Portfolio Management. Manage. Sci.
55(9).
Chatterji A, Delecourt S, Hasan S, Koning R (2019) When does advice impact startup
performance? Strateg. Manag. J. 40(3):331–356.
Christensen C (1992) Exploring the Limits of the Technology S-Curve. Part 1: Component
Techologies. Prod. Oper. Manag. 1(4).
105
Christensen, C. M., & Bower, J. L. (1996). Customer power, strategic investment, and the
failure of leading firms. Strat. Manage. J., 17(3), 197-218.
Cockburn IM, Henderson R, Stern S, Professor H, Management E, Business H, Morgan S (2018)
The Impact of Artificial Intelligence on Innovation
Cohen, W. M., & Levinthal, D. A. (1990). Absorptive capacity: A new perspective on learning
and innovation. Admin. Sci. Q. 128-152.
Davis, J. P., & Eisenhardt, K. M. (2011). Rotating leadership and collaborative innovation:
Recombination processes in symbiotic relationships. Admin. Sci. Q., 56(2), 159-201.
Eisenhardt KM (1989) Agency theory: An assessment and review. Acad. Manag. Rev. 14(1):57–
74.
Eisenhardt KM, Martin JA (2000) Dynamic capabilities: What are they? Strateg. Manag. J.
21(10–11):1105–1121.
Felten, E. W., Raj, M., & Seamans, R. (2019). The variable impact of artificial intelligence on
labor: The role of complementary skills and technologies. Available at SSRN 3368605.
Feng A, Graetz G, Adermon A, Autor D, Bárány Z, Goos M, Jensen A, et al. (2020)
TTRAINING REQUIREMENTS, AUTOMATION, AND JOB POLARISATION *.
Foster RN (1986) Working The S-Curve: Assessing Technological Threats. Res. Manage.
29(4):17–20.
Fu R, Kraft AG, Tian X, Zhang H, Zuo L (2019) Financial Reporting Frequency and Corporate
Innovation. SSRN Electron. J.
Furman, J., & Seamans, R. (2019). AI and the Economy. Innovation Policy and the
Economy, 19(1), 161-191.
Galasso A, Simcoe TS (2011) CEO Overconfidence and Innovation. Manage. Sci. 57(8):1469.
Galbraith JK (1956) American Capitalism Revised. (Houghton Mifflin, Boston, MA).
Gans JS, Kearney M, Scott EL, Stern S (2020) Choosing Technology: An Entrepreneurial
Strategy Approach
Gans JS, Stern S (2003) The product market and the market for “ideas”: commercialization
strategies for technology entrepreneurs
Gibbons R, Murphy KJ Optimal Incentive Contracts in the Presence of Career Concerns:
Theory and Evidence
Goldfarb A, Taska B, Teodoridis F (2019) Could Machine Learning Be a General-Purpose
106
Technology? Evidence from Online Job Postings. SSRN Electron. J.
Grossman SJ, Hart OD (1983) Implicit Contracts Under Asymmetric Information. Q. J. Econ.
98:123–156.
Hall BH, Khan B (2002) Adoption of New Technology
Henderson RM, Clark KB (1990) The Reconfiguration of Existing Product Technologies and the
Failure of Established Firms. Adm. Sci. Q. 35(1):9–30.
Hill CWL, Rothaermel FT (2003) The Performance of Incumbent Firms in the Face of Radical
Technological Innovation. Acad. Manag. Rev. 28(2):257–274.
Jacobides MG, Winter SG (2012) Capabilities: Structure, Agency, and Evolution.
Kaarboe OM, Olsen TE (2006) Career Concerns, Monetary Incentives and Job Design. Scand. J.
Econ. 108(2):299–316.
Kapoor R, Adner R (2012) What They Know: How Firms’ Production and Knowledge
Boundaries Affect Competitive Advantage in the Face of Technological Change. Organ.
Sci. 23(5):1227.
Kapoor, R., & Lee, J. (2013). Coordinating and competing in ecosystems: How
organization forms shape new technology investments. Strat. manage. j., 34(3), 274-296.
Karp R (2020) Gaining organizational adoption: Discovering new uses for existing innovations.
Working paper.
Kirtley J, O’Mahony S (2020) What is a pivot? Explaining when and how entrepreneurial firms
decide to make strategic change and pivot. Strateg. Manag. J.: 41(2), 3116-3131.
Lanzolla, G., & Suarez, F. F. (2012). Closing the technology adoption–use divide: The role of
contiguous user bandwagon. J. Manage., 38(3), 836-859.
Leiponen A (2008) Control of intellectual assets in client relationships: implications for
innovation. Strateg. Manage. J. 29(13):1371–1394.
Lerner J, Merges RP (1998) The Control of Technology Alliances: An Empirical Analysis of the
Biotechnology Industry. J. Ind. Econ. 46(2):125–156.
Loch CH, Tapper UAS (2002) Implementing a strategy-driven performance measurement system
for an applied research group. J. Prod. Innov. Manag. 19(3):185–198.
Magelssen C (2020) Allocation of property rights and technological innovation within firms.
Strateg. Manag. J. 41(4):758–787.
Makri M, Lane PJ, Gomez-Mejia LR (2006) CEO incentives, innovation, and performance in
107
technology-intensive firms: a reconciliation of outcome and behavior-based incentive
schemes. Strateg. Manag. J. 27(11):1057–1080.
Manso G (2011) Motivating Innovation. J. Finance 66(5):1823–1860.
Mayer KJ, Argyres NS (2004) Learning to Contract: Evidence from the Personal Computer
Industry. Organ. Sci. 15(4):394.
McCa;;i,, A.K. (2002) MALLET:A Machine Learning for Language Toolkit.
http://mallet.cs.umass.edu.
McEvoy M (2019) Analysis: 2019 Legal Tech Investments Top $1B After Strong Q3. Bloom.
Law. Retrieved (October 7, 2020), https://news.bloomberglaw.com/bloomberg-law-
analysis/analysis-2019-legal-tech-investments-top-1b-after-strong-q3.
Milgrom P, Roberts J (1992) Economics, Organization and Management (Prentice-Hall,
Englewood Cliffs, NJ).
Moravec H (1988) Mind children: The future of robot and human intelligence (Harvard
University Press).
Murphy K (1992) Appraisal: Motivating Managers to Identify and Reward Performance.
Perform. Meas. Eval. Incent.
Pew Research Center (2017) Automation in Everyday Life
Pisano GP, Teece DJ (2007) California Management Review How to Capture Value from
Innovation: Shaping Intellectual Property and Industry Architecture
Pivovarov V (2019) 713% Growth: Legal Tech Set An Investment Record In 2018. Forbes
https://www.forbes.com/sites/valentinpivovarov/2019/01/15/legaltechinvestment2018/#650
0741c7c2b.
Raffaelli, R., Glynn, M. A., & Tushman, M. (2017). Frame flexibility: the role of cognitive and
emotional framing in innovation adoption by incumbent firms. Strategic Management
Journal.
Rich, B. (2020) The Impact of AI Adoption on Size and Composition of In-House Legal
Departments. Working paper.
Rich B (2018) How AI Is Changing Contracts. Harv. Bus. Rev. Retrieved (October 6, 2020),
https://hbr.org/2018/02/how-ai-is-changing-contracts.
Rothaermel FT, Hess AM (2007) Building Dynamic Capabilities: Innovation Driven by
Individual-, Firm-, and Network-Level Effects)
Schilling, M. A. (1998). Technological lockout: An integrative model of the economic and
108
strategic factors driving technology success and failure. Acad. manage. rev., 23(2), 267-
284.
Schilling, M. A. (2002). Technology success and failure in winner-take-all markets: The impact
of learning orientation, timing, and network externalities. Acad. Manage. J., 45(2), 387-
398.
Schumpeter JA “Creative Destruction” From Capitalism, Socialism and Democracy (1942)
Siemsen E (2008) The hidden perils of career concerns in R&D organizations. Manage. Sci.
54(5):863–877.
Suarez, F. F. (2005). Network effects revisited: the role of strong ties in technology
selection. Acad. Manage. J., 48(4), 710-720.
Taylor, A. (2010). The next generation: Technology adoption and integration through internal
competition in new product development. Org. Sci., 21(1), 23-41.
Teece, D. J. (2007). Explicating dynamic capabilities: the nature and microfoundations of
(sustainable) enterprise performance. Strat. manage. j., 28(13), 1319-1350.
Teece DJ (1986) Profiting from technological innovation: Implications for integration,
collaboration, licensing and public policy. Res. Policy 15:67–88.
Teece DJ, Pisano G, Shuen A (1997) Dynamic capabilities and strategic management. Strateg.
Manag. J. 18(7):509–533.
Terwiesch C, Xu Y (2008) Innovation Contests, Open Innovation, and Multiagent Problem
Solving. Manage. Sci. 54(9).
Trajtenberg, M. (2018). AI as the next GPT: a Political-Economy Perspective (No. w24245).
National Bureau of Economic Research.
W. A, J.M. U (1978) Patterns of Industrial Innovation. Technol. Rev. 80(7):40–47.
Weigelt C, Sarkar MB (2009) Learning from supply-side agents: The impact of technology
solution providers’ experiential diversity on clients’ innovation adoption. Acad. Manag. J.
52(1):37–60.
Williamson O (1975) Markets and Hierarchies. Am. Econ. Rev. 62(2):316–325.
Wu, B., Wan, Z., & Levinthal, D. A. (2014). Complementary assets as pipes and prisms:
Innovation incentives and trajectory choices. Strat. Manage. J., 35(9), 1257-1278.
Abstract (if available)
Abstract
In this dissertation, I examine firms’ technology commercialization strategies and firms’ technology adoption strategies in light of use and proliferation of AI-enabled technologies. In chapter one, I discuss the existing literature on technology commercialization and technology adoption, how AI is distinct from existing technologies, and why the growth of AI necessitates firms adapting their technology strategies to fully maximize benefits from this technology. In chapter two, I examine technology adoption of AI within firms. Specifically, I examine how manager and employee perceptions of the value of AI differ in firms that have adopted AI and how this is reflected in actual use of the AI-enabled technology at the firm. I find that manager and employee alignment about the value of the AI dictates actual usage of the product and, in turn, renewal of the product, suggesting long-term integration into the firm’s processes. In terms of firm strategy, these findings suggest that firms should devote significant resources to ensuring alignment between managers and employees when deploying a new AI-enabled technology, in order to ensure that the firm captures value from the technology. It also suggests that firm learning may be the key mechanism by which this alignment is realized. The third chapter discusses AI commercialization and adoption from the perspective of the innovating firm. I explore how start-ups with novel technology pursue different strategies for commercialization. In this third chapter, I develop a framework for understanding variation in successful AI product commercialization using novel data related to AI-enabled software. As firms begin to use AI in ways that extend beyond merely deploying a product in the organization, it is important to understand how the unique characteristics of AI may alter how firms strategize around technology commercialization and adoption.
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Exploring the role of learning in contract design: empirical evidence using buyer-supplier IT services contracts
PDF
Essays on the role of entry strategy and quality strategy in market and consumer response
PDF
Technology, behavior tracking, and the future of work
PDF
User lobbying amongst high-growth tech startups: the next evolution of constituency building
PDF
How do acquirers govern the deal-making process? Three essays on U.S. mergers and acquisitions 1994 – 2017
PDF
Empirical essays on relationships between alliance experience and firm capability development
PDF
Enchanted algorithms: the quantification of organizational decision-making
PDF
Theoretical and data-driven analysis of two-sided markets and strategic behavior in modern economic systems
PDF
Three essays on young entrepreneurial firms
PDF
The structure of strategic communication: theory, measurement, and effects
PDF
The morality of technology
PDF
Competing across and within platforms: antecedents and consequences of market entries by mobile app developers
PDF
Private disclosure contracts within supply chains and managerial learning from stock prices
PDF
Understanding the disclosure practices of firms affected by a natural disaster: the case of hurricanes
PDF
Changing fundamental analysis in the new economy: the case of DuPont analysis and STEM firms
PDF
Essays on the firm and stakeholders relationships: evidence from mergers & acquisitions and labor negotiations
PDF
Bounded technological rationality: the intersection between artificial intelligence, cognition, and environment and its effects on decision-making
PDF
The role of capabilities in new alliance creation and performance: a study of the biotechnology industry
PDF
Generating and utilizing machine explanations for trustworthy NLP
PDF
Common ground reasoning for communicative agents
Asset Metadata
Creator
Rich, Beverly
(author)
Core Title
Automated contracts and the lawyers who don't review them: adoption and use of machine learning technology
School
Marshall School of Business
Degree
Doctor of Philosophy
Degree Program
Business Administration
Degree Conferral Date
2021-12
Publication Date
10/14/2021
Defense Date
06/03/2021
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
artificial intelligence,capabilities,Contracts,incentives,innovation,machine learning,OAI-PMH Harvest,Technology,Technology Commercialization
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Mayer, Kyle J. (
committee chair
), Jia, Nan (
committee member
), Miric, Milan (
committee member
), Teodoridis, Florenta (
committee member
)
Creator Email
beverly.rich@eccles.utah.edu,brich@usc.edu
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-oUC16208268
Unique identifier
UC16208268
Legacy Identifier
etd-RichBeverl-10163
Document Type
Dissertation
Rights
Rich, Beverly
Type
texts
Source
University of Southern California
(contributing entity),
University of Southern California Dissertations and Theses
(collection)
Access Conditions
The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright. The original signature page accompanying the original submission of the work to the USC Libraries is retained by the USC Libraries and a copy of it may be obtained by authorized requesters contacting the repository e-mail address given.
Repository Name
University of Southern California Digital Library
Repository Location
USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA
Repository Email
cisadmin@lib.usc.edu
Tags
artificial intelligence
capabilities
incentives
innovation
machine learning