The Business of AI Startups

James Bessen, Boston University

Stephen Michael Impink, HEC Paris
Lydia Reichensperger, Boston University
Robert Seamans, New York University

July 25, 2023*

Abstract: Increased computing power and advanced machine learning (ML) techniques have
accelerated the development and use of “artificial intelligence” (AI). Despite the increased
commercial focus on AI, we have limited information about how AI products are developed by
resource constrained startups. We collect survey data from startups producing commercial AI
products to provide a glimpse in how these product impact labor across industries. This paper
chronicles the last five years of AI startup survey data, documenting information about their
business models, product, suppliers, partners, and customers. Our descriptive findings add to
conversations on whether AI products enhance or replace human capabilities; which algorithms,
data protections, and frameworks high-tech startups use to develop their AI products; and which
geographies, customer-types, and industries are sales targets.

JEL codes: O33, J21, L10

Keywords: artificial intelligence, entrepreneurship, automation, labor, technology

*This manuscript is a revised version of the working paper released on November 29, 2018 with
the same title.

Electronic copy available at: https://ssrn.com/abstract=3293275

 1. Introduction

Breakthroughs in machine learning have led to the acceleration of commercial “artificial

intelligence” (AI) development for a wide range of uses. The potential of these new technologies

has prompted numerous assessments that these products will drastically alter the economy

(Acemoglu and Restrepo, 2018; Frey and Osborne, 2017; Furman and Seamans, 2019; Felton et

al., 2021). According to some commentators, we are entering a “Race Against the Machine”

(Brynjolfsson and McAfee, 2011), a “jobless future” (Ford, 2015), a “Fourth Industrial

Revolution” (Schwab 2016) and the “Second Machine Age” (Brynjolfsson and McAfee, 2014).

Results from these studies differ in their estimation of AI’s impact and their tone1, yet there is

consensus that AI development and using big data to train AI not a passing trend.

Much of this new literature on the economic consequences of AI has been based on

estimated capabilities of the new technology, without accounting for possible difficulties in

commercialization and technology adoption. For instance, the Frey and Osborne study (2017) is

based on the assessments of a small panel of machine learning experts who were asked which of

70 occupations were “completely automatable” in 2013. Even if these assessments are correct, just

because a job is automatable does not mean that it will be automated. Moreover, both theoretical

models and historical evidence show that even when most of the labor of an industry is automated,

employment can grow rapidly (Bessen, 2019).

If AI is likely to drastically change the economy in a decade or so, then we should see

evidence of that at the actual businesses where AI is being applied today. That is, if 47% of jobs

will soon be at risk from AI, then surely jobs will be at risk already at those firms using AI

1
On the more pessimistic side estimates that 47% of jobs in the US will be “at risk of automation” in the next
decade or so (Frey and Osborne 2017). On the less-pessimistic side, Bessen (2019) and Furman and Seamans (2019)
discuss the potential for AI to augment labor’s capabilities.

Electronic copy available at: https://ssrn.com/abstract=3293275

applications. A look at the cutting-edge applications of the technology provides a window into

likely outcomes over the next decade or so.

To glimpse the impact of AI development, our paper reports on a global survey of startups

developing and licensing commercial AI products. The survey questionnaire covers a wide range

of topics, including questions about startups’ markets and customers, their AI technologies and

frameworks used in product development, their use and protection of training data, their product

users and use cases, and their customers’ demographics and industries. These findings paint a vivid

picture of how startups pull together the needed resources to develop and monetize their AI

applications. And that picture is at odds with alarmist predictions about the impact of AI. Rather

than drastically reducing jobs, current data commercial AI applications are much more often about

enhancing human capabilities than about reducing labor costs. While some occupations are

experiencing job losses, others are experiencing job creation. The actual effect of AI on jobs varies

by occupation (Felton et al., 2021), which our data reflects.

The survey provides a window into other topics in addition to how AI effects labor,

including how AI startups address barriers to entering this nascent industry, how they protect

customer data and respond to new data protection regulations such as the EU General Data

Protection Regime (GDPR), and the extent to which they develop their own machine learning

algorithms or rely on third party providers. While the survey establishes that there is a vibrant

startup community of AI application developers and many new startups enter each year, it is

possible that investments by more established firms may create barriers to entry in some markets.

Respondents answer questions on data, hardware, software and firm size, which have previously

been linked to entry barriers.

Electronic copy available at: https://ssrn.com/abstract=3293275

 Though startups are constrained by the type of data they collect, in many cases lacking the

diversity or scale of data collected by larger firms, 80 percent of survey respondents report using

their customers’ data to train AI. Customers proprietary data correlated with increased future

performance, suggesting that the inability for startups to access barriers to scale from having less

imitable data-resources. In the last several years, access to cheaper computing power (i.e., IT

hardware) has made it easier to enter the industry and develop their products, as introductory offers

from cloud provider and venture capital “spray and pray” investing have reduced the cost of IT

(Ewens et al., 2019). Moreover, access to standardized tools and other resources from cloud

suppliers may lead to time saving, enabling startup to focus on differentiating their product instead

of building core app infrastructure (Impink, 2023). These more standardized tools may benefit

performance without necessarily inhibiting innovation (Miric et al., 2022). Due to the availability

of cloud IT and basic development tools, labor (Rock 2022) and proprietary data (Bessen et al,

2022) may be a key source of firm differentiation.

2. Prior research on how automation impacts labor and occupations

In addition to basic facts about the business of AI, this study provides evidence that bears

on several major questions that have been raised in the literature, including the impact of AI on

jobs and occupations, and the role of data as a critical resource for AI-enabled startups. The first

question, noted above, concerns the impact of AI on jobs. Frey and Osborne (2017) base their

estimate on a technical evaluation of “automatability,” the technical feasibility of automation. The

literature highlights several economic and technical reasons this might not be a sufficient metric

for understanding the impact of AI on jobs. One reason is that automating some or even most tasks

an occupation performs does not necessarily eliminate the occupation. Historically, most

Electronic copy available at: https://ssrn.com/abstract=3293275

automation has been partial, such as the introduction of the assembly line in manufacturing

(Rosenberg, 1963). Even though automation increased productivity, workers may still be cogs in

the broader process (Chandler, 1962). Bessen (2016) finds that of 270 detailed occupations listed

in the 1950 Census, only one can be described as having been eliminated due to automation,

namely the job of elevator operator. Furthermore, partially automating a job can increase

employment in that occupation or industry as well as decrease employment (Acemoglu and

Restrepo, 2018). This is because automation tends to decrease prices, driving greater demand.

When demand is elastic enough, greater demand will offset the labor-saving effect of automation.

For example, during the 19th century, the textile industry was heavily automated, yet employment

rose (Bessen, 2019). More recently, the automatic teller machine (ATM) automated some of the

work of bank tellers, yet their employment grew as well (Bessen, 2016). Also, the effect of AI is

not just to automate tasks which replace humans, but also to enhance human capabilities at both

performing new tasks and old tasks more effectively.

In the recent empirical literature, (Arntz et al., 2017) modify the basic estimates of Frey

and Osborne to account for the partial nature of automation. However, they do not consider the

adoption of the technology or the labor demand effects that might cause automation to increase

employment. The McKinsey Global Institute (Manyika, Chui et al., 2017; Manyika, Lund et al.,

2017) attempts to take a range of hypothetical adoption rates into account in estimating impacts,

and they also consider hypothetical labor demand effects. Accenture (2018) also estimates how

the impact of AI on employment will vary depending on the role of different skills, and also predict

where there will be deficits and surpluses of these skills on a geographic basis.

Several papers have looked at the impact of robots on employment, with differing results

(Acemoglu and Restrepo 2017, Graetz and Michaels 2015, Dauth et al. 2018). Graetz and Michaels

Electronic copy available at: https://ssrn.com/abstract=3293275

(2015) use robot shipment data at the country, industry and year level from the International

Federation of Robotics (IFR) and find that robots added an estimated 0.4 percentage points of

annual GDP growth between 1993 and 2007 on average for the 17 countries in their sample

(accounting for about one-tenth of GDP growth during this time period) while having a slightly

positive but statistically noisy effect on employment.

In contrast, Acemoglu and Restrepo (2017) use the IFR data to examine the impact of the

increase in industrial robot usage on regional U.S. labor markets between 1990 and 2007. These

authors find that industrial robot adoption in the United States was negatively correlated with

employment and wages during this time period—according to their estimates, each additional

robot reduced employment by six workers and that one new robot per thousand workers reduced

wages by 0.5 percent. The authors note that the effects are most pronounced in manufacturing,

particularly in routine manual and blue-collar occupations, and for workers without a college

degree.

Dauth et al. (2017) combines German labor market data with IFR robot shipment data and

finds that while each additional industrial robot leads to the loss of two manufacturing jobs, enough

new jobs are created in the service industry to offset and in some cases over-compensate for the

negative employment effect in manufacturing. This finding echoes a broader study by Autor and

Salomons (2018) which explores the general impact of productivity growth on employment,

including both own industry and cross industry effects, and finds that productivity growth does not

reduce employment in general, thanks in large part to positive spillovers into related sectors.

A second question we address is which occupations will be affected. Frey and Osborne

(2017) argue that lower wage occupations will experience greater job losses than higher wage

occupations. The McKinsey Global Institute projects, somewhat differently, that high wage

Electronic copy available at: https://ssrn.com/abstract=3293275

occupations will grow while mid-wage occupations shrink. Kaplan (2015) posits that “automation

is blind to the color of your collar” and many professions will be devastated. Susskind and

Susskind (2015) argue that new technology will lead to the decline of the professions. Our survey

provides some evidence about which occupations are growing and which are losing jobs in

response to AI. Other recent papers that take the task- or ability-based approach include

Brynjolfsson, Mitchell and Rock (2018) and Felten, Raj and Seamans (2021). Moreover, a large

literature has explored the differential impact of technology on different groups of workers. This

includes the literature on skill-biased technical change (Katz, 1999) and on job polarization (Goos

et al., 2009; Autor, Katz, and Kearney 2008).

Next, a third question our study addresses are the extent of entry barriers into AI markets.

Investment in AI is currently dominated by large firms, especially a few large tech firms with slack

cloud compute resources – Amazon, Google, and Microsoft. Moreover, these firms also control

generative AI models (BARD, ChatGPT, Dalle) and AI development frameworks (TensorFlow,

PyTorch), and they have tons of customers that provide data to use their services and products. In

particular, some observers, such as Stucke and Grunes (2016), argue that the combination of data

and network effects creates substantial entry barriers in online markets. Others point to scale

barriers stemming the scale, IT infrastructure, and data-resource wielded by the largest technology

firms (Benzina, 2019; Khan 2017). For example, because Google has an enormous amount of

search data, it might be hard for a new competitor to compete in the search engine market.

Others contend that data, by itself, is not likely to pose an entry barrier (Lambrecht and

Tucker 2015, Sokol and Comerford 2016). It may be that Google’s advantage comes not from the

amount of data, but from the results of the product-focused experiments. Varian, Google’s chief

economist, and co-authors (2010) report that Google conducted 6,000 experiments on its search

Electronic copy available at: https://ssrn.com/abstract=3293275

engine in 2008. Moreover, while large amounts of data are typically needed for machine learning,

there may be diminishing returns to the amount of data beyond a certain point. For instance, Bajari,

Amazon’s chief economist, and co-authors (2019) find that increasing the number of online

products that Amazon tracks does not significantly improve machine learning prediction accuracy

after a certain point, implying that data quantity provides only a low barrier to entry. Amazon and

other large technology companies benefit from the breadth and recency (Chiou and Tucker, 2017)

of the data captured. For example, startups are more likely to sell a subset of the broader product

offering of a larger company and to engage with small to medium size customers. In any case, our

survey provides basic information about startups’ access to data and circumstantial evidence on

relative entry barriers in different industries.

Lastly, the fourth question this study addresses is how startups use data protections. Data

protection and data privacy has become a flashpoint in the media, due in part to high profile data

breaches such as at Equifax in 2017 and in part to high profile exposure of Facebook user’s

personal data to Cambridge Analytica in 2016 and 2017. Partially in response to these events,

government regulators have instituted tighter rules on data protection. This has most notably

manifested in Europe in the form of General Data Protection Regime (GDPR). There has been

some concern that the increased data protection required under GDPR may constitute a barrier to

entry for startups, and early research suggests that GDPR may in part be contributing to a

slowdown in VC investment in European based startups (Jia et al., 2021). Additional evidence

suggests that GDPR more negatively effects smaller web technologies firms than larger ones. For

instance, Google gains significant share after GDPR was passed at the detriment of smaller firms

(Peukert et al., 2022).

Electronic copy available at: https://ssrn.com/abstract=3293275

 Our survey adds to this nascent literature by providing basic information about the data

protections that AI startups have in place and allow us to compare the use of data protections across

small vs large startups as well as European vs non-European based startups.

3. Survey and Sample

Sample selection

The main portion of our sample comes from the Crunchbase database of firms. Each year

we select startups tagged as AI firms or described with the term ‘artificial intelligence’ or machine

learning’. At the time of selection, these firms are listed as in-business and operational. We focus

on English-language countries, where Crunchbase has greater coverage, and omit China. The list

of startups selected are primarily based in developed countries, which the vast majority of firms

are headquartered in the USA, Canada, UK, or the EU. In an attempt to have a more homogenous

sample, we only include responses from startups that are less than ten years old and less than 500

employees. This enables us to focus on high growth startups, while weeding out mom and pop

ventures and different business models.

In the first round of the survey, we also included startups from three additional sources:

alumni of the Creative Destruction Lab, a startup incubator, who were identified as working with

AI; a list of machine learning companies obtained from Philipp Hartmann and Joachim Henkel

(Hartmann and Henkel, 2018); also, O’Reilly Media ran a notice of the survey in its AI newsletter,

providing a link to the online survey. Once a startup is added to our sample list, we send them a

survey every year. Moreover, we reach out to multiple email addresses at the firm. For instance,

we would email the individual who responded from the firm in the prior year and any other email

aliases listed on Crunchbase (i.e., ‘[email protected]’ or ‘[email protected]’).

Electronic copy available at: https://ssrn.com/abstract=3293275

 Over the last five years, we have reached out to 5,254 startups. For each individual survey,

our response rate is around 8%. In total, firms responding at least once our response rate is 22%.

Survey rounds

Survey 1 was administered online using Qualtrics from in the summer of 2018. Surveys 2-

4 occurred the first few months of 2020, 2021, and 2023. Survey 5 occurred in the early spring of

2023. They survey consists of around 36 questions in a typical year. These questions (excluding

name and email questions) are pre-tested on academics and on roughly a half dozen firms with

interviews. In survey 2 we added questions on GDPR; in survey 3 we added questions on ethical

AI development; and in survey 4 we added question on the governance of AI. Survey 5 asks the

same questions as survey 4, with the goal of building a panel of responses throughout time. After

each round of the survey, we receive feedback from startups that helps us adapt the survey in the

next round. However, the core questions on firm demographics, customers, and core product

development that we report on have remained identical in all five survey rounds.

Across all surveys, we have received 1,178 responses from 917 unique startups. Figure 1

shows the number of responses by survey round. More than 160 startups have responded to the

survey more than one time in the last five years; two startups have responded all five times.2 Figure

2 reports then number of times that each startup has responded to the survey. This will be

particularly important in future projects as we build out a time-series of responses to determine

how those responses change over time.

2
Expect your gold star to be delivered via mail, if you responded all five times.

10

Electronic copy available at: https://ssrn.com/abstract=3293275

 Figure 1 – Responding startups (count) by survey round

298
300

263
254

225
count of respondents

200

138

100

0
Survey 1 Survey 2 Survey 3 Survey 4 Survey 5

Figure 2 – Startups responding in multiple survey rounds

814
800

600
count of respondents

400

200
133

24
4 2
0
1 2 3 4 5
Number of annual survey responses

11

Electronic copy available at: https://ssrn.com/abstract=3293275

 Characteristics of respondents

We show the representativeness of subsets of our data in several papers (Bessen et al.,

2022, Bessen et al, 2023). Generally, we find firms based in California are less likely to respond.

In certain subsets of the data, we find that firms that are less than two years old are less likely to

respond, potentially suggestion that more nascent ventures may not have full time employees.

Table 3 reports the basic summary statistics for responding firms in all surveys. The

startups in our sample are quite small with around half of the firms having less than 10 employees

(Figure 4) and existing for three or fewer years (Figure 5). More than half of the responding

startups are located in North America, for multiple reasons. First, there is a higher concentration

of startups in California, New York, Massachusetts, and Toronto. However, we are no blind to the

fact that Crunchbase has coverage issues in the developing and non-English language countries

and the survey was sponsored and run by two US universities. Upcoming survey will be sponsored

by HEC Paris in addition to NYU Stern and Boston University to increase response rates withing

Europe. Figure 6 shows the count of survey responses by region.

Table 3 – Responding startup demographics (all survey)

Mean SD Min Max
Age 4.55 2.25 1 10
Emp. (Count) 26.78 40.97 1 500
Small (< 11 emp.) 0.52 0.5 0 1
Healthcare Ind. 0.12 0.32 0 1
Financial Ind. 0.06 0.23 0 1
US 0.44 0.5 0 1
UK 0.05 0.21 0 1
France 0.03 0.17 0 1
Germany 0.04 0.19 0 1
Canada 0.05 0.21 0 1
Founder MBA 0.2 0.4 0 1
Founder Big Tech 0.04 0.19 0 1
Founder Female 0.08 0.27 0 1
Responses 1178

12

Electronic copy available at: https://ssrn.com/abstract=3293275

 Figure 4 – Age distribution

250
227

196
200

163
count of respondents

153
150
132

96
100
72
63

50 41
27

0
1 2 3 4 5 6 7 8 9 10

Figure 5 – Employee count distribution

600
541
518
count of respondents

400

200

73

32
6
0
5 25 50 150 250

13

Electronic copy available at: https://ssrn.com/abstract=3293275

 Figure 6 – Region

612
600
count of respondents

400
339

188
200

39

0
Americas Asia Europe MEA

Survey respondents disproportionately sell to mid-sized firms. Figure 7 shows the share

of customers in each size class from the survey. Almost half of the startups sell to firms in the

middle category with 51-1000 employees, but these firms make up only 26% of the market as

measured by employment size from the US Census Longitudinal Business Database (LBD). The

responding startups also sell to both smaller and larger customers, but proportionately less than

would be expected given the distribution of firms in developed countries. Most of the firms are

shipping product (71%), but 20% are in beta testing, and 9% are pre-beta (Figure 9).

14

Electronic copy available at: https://ssrn.com/abstract=3293275

 Figure 7 – Customers’ mean employment size

.5
.4
% of respondents
.2 .1
0 .3

Individuals 2-10 11-50 51-100 101-500 500-1000 1000+

Figure 8 – Product stage by region

80

Shipping
60
% of respondents
40

Beta
20

Pre-beta
0

Americas Asia Europe

Shipping Beta Pre-beta

15

Electronic copy available at: https://ssrn.com/abstract=3293275

 Table 4 reports summary statistics by survey round. Startups in the nascent industry are

become a bit older, from 3.7 years old on average in 2018 to 5.3 years on in 2023. The mix of

responses by geography shows a decline in the percent of respondents from Canada and an increase

in the percent of responses in Europe. Our initial list of startups included data from Creative

Destruction Labs (University of Toronto) in Canada, which may explain of this change. However,

we also had specific data from the TUM accelerator program in Germany in the first round of the

survey but show increases in Germany instead of declines.

The number of respondents with an MBA is also declining over time, suggesting earlier AI

startups had a higher proportion of management education. This potentially suggests entry barriers

were higher initially, required greater credentialling to kickstart the venture. Lastly, about 10% of

respondents are female, and this remains stable over time.

Table 9 – Responding startup demographics (by survey)

Mean
Survey 1 Survey 2 Survey 3 Survey 4 Survey 5
Age 3.69 4.18 4.14 5 5.26
Emp. (Count) 36.32 26.1 23.88 26.92 24.69
Small (< 11 emp.) 0.46 0.49 0.61 0.51 0.54
Healthcare Ind. 0.09 0.14 0.15 0.1 0.09
Financial Ind. 0.02 0.08 0.07 0.05 0.05
US 0.46 0.46 0.41 0.46 0.42
UK 0.08 0.06 0.1 0.01 0.01
France 0.03 0.02 0.02 0.03 0.04
Germany 0.02 0.02 0.05 0.05 0.04
Canada 0.12 0.04 0.06 0.03 0.03
Founder MBA 0.25 0.27 0.21 0.17 0.12
Founder Big Tech 0.04 0.05 0.06 0.04 0.01
Founder Female 0.01 0.09 0.07 0.1 0.1
Responses 138 263 225 298 254

16

Electronic copy available at: https://ssrn.com/abstract=3293275

 4. Findings

Heterogenous impact by industry

The survey respondents sell to customers in all major industry sectors as well as to

individual consumers (Figure 10). In this sense, machine learning is a general-purpose technology

that can be used in a variety of applications across a variety of industries (Cockburn et al 2018).

This broad distribution across industries suggests that the startup environment is healthy

and many opportunities for entry exist. However, some industries may have relatively higher entry

barriers than others. If there are no entry barriers, then both large and small firms invest based on

the size of economic potential in an industry. This means that without entry barriers, the

distribution of AI development spending should look the same for smaller firms as it does for

larger ones. If, on the other hand, startups face significant entry barriers in an industry, then large

firms would spend proportionately more than smaller firms in industries with entry barriers,

holding the anticipated economic potential across industries constant. Comparing startup funding

in different industries relative to total investment in AI in those industries helps identify those

industries with possible entry barriers.

Although our survey only includes startups and not large established firms, the McKinsey

Global Institute conducted a survey of top managers at large and small firms (Bughin et al. 2017).

By comparing estimates from the two studies for different industries, we can identify differences

in relative funding that might signal entry barriers. The McKinsey study reports the current

adoption of AI for select industries, presented as a weighted mean by firm size (Bughin et al. 2017:

Exhibit 4). Using this as a proxy for total AI investment in each industry for 2018, Figure 11 shows

the share of startup participating in each industry from our survey compared with the share of total

investment from the McKinsey survey.

17

Electronic copy available at: https://ssrn.com/abstract=3293275

 The chart shows that the share of investment going to startups and to all firms is roughly

the same in many industries, but in three industries total investment—likely reflecting large firm

investment—is disproportionately large: consumer packaged goods, transportation and logistics,

and high tech and telecommunications. On the other hand, it appears that startup investments in

AI applications serving retail and finance are disproportionately larger, likely reflecting venture

capital priority areas.

Figure 10 – Industry
15 10
5 % of respondents
0

Retail Fin Health IT Manuf Comm Util Serv Govt Edu Agr Law

18

Electronic copy available at: https://ssrn.com/abstract=3293275

Figure 11 – Industry Share of Funds: Survey respondents vs. McKinsey projections (2018 only)

Do AI products augment productivity or replace workers?

Finally, to understand the nature of the marketing appeal—the “unique selling

proposition”—the survey asked firms to rate various benefits that their products provided to

customers. Each benefit is rated on a 5-point scale from “strongly disagree” to “strongly agree”

(or “not applicable”). Figure 12 shows the share of responses ranked “strongly agree.”

The three most frequent benefits reported are capabilities to make predictions or decisions,

to manage and understand data, and to create new and improved products and services. These

answers provide a gauge for understanding how much AI enhances human capabilities and how

much, instead, it tends to replace them. Of the survey responses, 53% strongly agree that their

19

Electronic copy available at: https://ssrn.com/abstract=3293275

products automate routine tasks and 48% strongly agree that their products reduce labor costs.

That is, AI appears to be about enhancing human capabilities, not just replacing them. AI may very

well eliminate certain types of occupations or asymmetrically affect industries (something we

explore further below), but because facets of the technology appear to augment human capabilities,

it might well be associated with increased employment and wages for at least some occupations.

Moreover, the replacement of workers appears to be significantly concentrated in the

certain sectors of the economy. We find that, for firms that sell to customers in agriculture,

manufacturing, utilities and transportation industries, 74% strongly agree that their products

benefit customers by automating routine tasks and/or reducing labor costs. For firms that do not

sell to those industries, only 46% strongly agree. This difference across industry groups is highly

significant (t-test probability value of 0.000).

Figure 13 reports the share of firms reporting whether each occupational group at their

customer firms is a user or not. White collar occupations are listed most frequently, including

professionals (58%), managers (49%), and sales and marketing occupations (35%). Occupations

that include clerical, administrative, manual or service work are less likely to be listed. In this

regard, the distribution of users of AI across occupations is very similar to the distribution of

computer users across occupations (Bessen 2016).

20

Electronic copy available at: https://ssrn.com/abstract=3293275

Figure 12 – Strongly agree (%): AI product augments labor or reduced labor costs

50
40
% of respondents

30
20
10
0

Automate tasks Reduce labor costs

Figure 13 – Types of workers impacted by the startups AI product

.6
.4
% of respondents
.2
0

Managers Professional Sales Front-line Admin. Manual

21

Electronic copy available at: https://ssrn.com/abstract=3293275

AI products appear to require a level of skill to the extent that the occupations involved have

skill requirements. But does the use of these products require STEM skills or specialized

training? The survey responses suggest that in most cases specialized computer skills or specific

training demands are modest. Only 10% of firms require users to have expert coding or data

skills. 59% require general familiarity with computers and the remainder require no special skills

at all.

In terms of the training provided to users, 86% of firms report this training takes one

week or less. 12% report that 2 to 4 weeks of training is required and 2% report up to 3 months

of training is required. Fully 44% of the firms provide no training to their customers. Of those

firms that do provide training, 70% provide onsite training classes, 32% provide offsite training

classes, and 47% provide online courses.

Augment productivity or replace workers?

The analysis of customer benefits above suggests that most firms are oriented to enhancing

customer capabilities rather than reducing customer labor costs, especially in the broad service

sector. This suggests that AI might have some job-creating potential, especially for the

professional, managerial, sales and marketing occupations that tend to be the users of these

products. On the other hand, AI also reduces labor costs for some customers. Moreover, these

effects might differ across occupations because use of AI differs dramatically across occupations.

Figure 14 reports the share of firms that expect their products to create jobs or eliminate jobs in

different occupational groups at their customers, listed in the same order as in Figure 13.

22

Electronic copy available at: https://ssrn.com/abstract=3293275

 AI is not all about destroying jobs. These findings suggest that occupations using AI may

benefit form augmentation of their capabilities (i.e., increased productivity). At the same time,

many jobs will be eliminated, especially in the three occupational groups that use AI relatively

less. In many cases, jobs will be created in some occupations and jobs will be destroyed in other

occupations at the same affected firms. Of the firms that responded to this question, almost half

(46%) reported that their products both create jobs and destroy jobs at customer firms. 26% report

only creating jobs and 28% report only eliminating jobs. It is possible, of course, that survey

respondents, perhaps sensitive to publicity about job losses, shaded their answers to reflect better

job outcomes than is actually the case. Nevertheless, Figure 13 reveals dramatic relative

differences between occupational groups. Professionals, managers, and quantitative sales workers

may benefit from AI more than administrative, frontline, and manual workers.

Figure 13 – Types of workers augmented by AI products (versus being replaced)

.25
.2
% of respondents
.1 .15
.05
0

Managers Professional Sales Front-line Admin. Manual

23

Electronic copy available at: https://ssrn.com/abstract=3293275

 Technology, frameworks and relationships with Big Tech

In building their products, responding startups use many different types of algorithms, AI

development frameworks, and cloud tools to develop their products. The most commonly used

technology is natural language understanding and text analysis (63% of firms), followed by natural

language classification and decision management (both at 56%). These are followed by visual

recognition, including image, face and video (45% of firms) and sentiment/emotion analysis (43%

of firms). Other technologies are used by smaller percentages of firms.

Overall, many more firms develop their own software for the most commonly used

technologies rather than purchase them from an external vendor. Only in two areas do firms rely

more on outside vendors: speech recognition with 19% using external products while 13% develop

their own, and natural language translation, with 17% using external software and 13% develop

their own.

Regarding the type of algorithm used (Figure 14), 65% of reporting firms use neural

networks including recurrent, convolutional, and generative adversarial neural networks. The next

most common methods are clustering algorithms (46% of firms) and Bayesian or other methods

of probabilistic inference (41% of firms). Other methods are used less frequently. Most startups

use TensorFlow or PyTorch (Figure 15). However, the selection of the AI development framework

depends on the fit of that framework with the algorithm, training data, and product. For instance,

Figure 16 shows that startups using neural networks or ensemble learning, commonly discussed

as more sophisticated algorithms, are more likely to use PyTorch and Google ML than TensorFlow

and Keras.

24

Electronic copy available at: https://ssrn.com/abstract=3293275

 Figure 14 – Commonly used algorithms

.6
% of respondents
.2
0 .4

Neu. Net. Cluster Bayesian Reinforce. Ensemble Instance Reg. Genetic

Figure 15 – Commonly used AI frameworks

.25
.2
% of respondents
.1 .05
0 .15

Tensor PyTor. Keras Goog ML Sagemaker Caffe Apache Therano CNTK

25

Electronic copy available at: https://ssrn.com/abstract=3293275

 Figure 16 – Commonly used AI frameworks by most popular algorithms

40
Tensor TensorFlow
PyTorch
30 PyTorch

Keras
% of category

Google ML
Keras
20

Goog ML
10
0

NN or EL Not NN or EL

Almost all the startups report that their products are cloud-based (97%) although 33% of

the firms additionally provide software on premises. Only 3% of the firms provide software on

premises only. Most of the firms (68%) provide a commercial application using AI. Some use the

AI in their own products and services (43%) and 12% provide developer tools for AI applications

(these are not mutually exclusive categories).

Thus, while startups use a wide variety of technologies to perform a variety of functions, a

typical firm uses Amazon AWS cloud platform services, neural networks, and either TensorFlow

or PyTorch to develop their AI product

26

Electronic copy available at: https://ssrn.com/abstract=3293275

 Figure 17 – On Premise versus cloud IT by region

Cloud
.6
.4

Both
.2

On-prem
0

Americas Asia Europe

Cloud Only Both On-Prem only

Figure 18 – Cloud IT provider by region

80

Amazon
60
% of respondents
40

Goog
20

Msft
0

Americas Asia Europe

Amazon Google Microsoft

27

Electronic copy available at: https://ssrn.com/abstract=3293275

 Figure 19 – Cloud IT provider by firm size
80

Amazon
60
% of respondents
40

Goog
20

Msft
0

Smaller (<11) Larger

Amazon Google Microsoft

Figure 20 – Startup data collaboration with Big Tech by industry

Retail
Fin
Health
No Tech Data Collab.
IT
Manuf
Comm
Util
Serv

Tech Data Collab. Govt

Edu
Agr
Law

0 5 10 15
% of respondents

28

Electronic copy available at: https://ssrn.com/abstract=3293275

 Data and data protections

Startups also use a wide variety of data. 57% of the firms use unstructured text, 44% use

transaction data, 38% use images, 37% use administrative data or other structured records. A

smaller share of firms use audio, video, or other types of data. These data are mainly used to train

algorithms. Consequently, the algorithms are re-trained as more data accumulates. Roughly a

quarter of firms report refreshing their models daily, weekly, or monthly each. 13% of firms report

having models that are not refreshed with new data.

Startup firms generally use other people’s data. The most common source of data is from

customers. 80% of the startup firms report using customer data, including data about their

customer’s customers and users as well as other data. 63% use data from third parties, including

government data, data scraped from the Internet, and public benchmarking data. 51% of the firms

report using their own proprietary data. Most of this use is in combination with data from other

sources; only 6% of firms rely only on their own data.

To protect their access to data, startup firms who use customer data retain secondary reuse

rights 52% of the time. To control the use of proprietary data between the firm and its customers,

83% of the firms use legal contracts that specify data uses. Additionally, firms use a variety of

technical means to protect and control data access, including de-identification, encryption,

passwords, access logs, and application program interfaces.

Only 22% of firms report that the EU’s General Data Protection Regulation (GDPR) has

impacted sales and marketing to non-EU countries. That figure is 27% for firms headquartered in

Europe, excluding the United Kingdom. Given that the GDPR went into effect during our survey

period, these figures might change as firms have more experience with the regulation. The survey

respondents were asked to select all types of data protection used. As reported in Figure 22 , across

29

Electronic copy available at: https://ssrn.com/abstract=3293275

all types of data protection, startup firms with customers in the EU report using data protections

more intensively than startup firms without customers in the EU. This could reflect the impact of

the GDPR and also different customer sensitivities.

There are some differences in data protection across firm size. Startup firms with ten or

fewer employees represent close to 40% of the total survey respondents. These small startup firms

are less likely to use legal contracts, de-identification, data encryption and password protected

access, as reported in Figure 8. However small startup firms are equally likely to use logged access

and application program interface as large startups.

Figure 21 – Types of job augmented by AI products (versus being replaced)

.4 .3
Percent of respondents
.2 .1
0

Images Structured Text Video Transactions Audio

30

Electronic copy available at: https://ssrn.com/abstract=3293275

 Figure 22 – Types of job augmented by AI products (versus being replaced)

.6 Contracts

Encryption

De-indentification
% of respondents
.4

Passwords

Logged access
.2

API
0

Not 0EU 1
EU

The attached appendix includes data for the survey question on ethics. Bessen et al. (2023)

uses responses from the 2021 and 2022 to examine the relationship between AI ethics policies,

pro-ethics actions, and venture funding. This appendix also includes data from the 2023 survey.

5. Discussion

AI and Jobs

The evidence above in Figure 6 suggests that the benefits of AI are more about enhancing

human capabilities than about replacing humans. This suggests that the technology often makes

human workers more valuable rather than less, possibly raising labor demand for professionals,

managers and sales and marketing personnel. The evidence in Figure 10 suggests that this is

exactly what is happening. AI is associated with increasing labor demand for some occupations,

31

Electronic copy available at: https://ssrn.com/abstract=3293275

decreasing demand for others. For the customers of these startups, AI appears to be less about

eliminating labor overall and more about shifting work from some occupations to others.

This shift is illustrated by an account given by one of the firms interviewed in the survey

pre-test. The firm’s product automated the retrieval of contact information for prospective sales

contacts from text files. This eliminated clerical work for the people who had manually searched

for contact information, but because it was able to generate more sales prospects quickly, it also

created jobs for sales and marketing personnel. This pattern of shifting work appears quite similar

to the pattern observed for computer use where professional and managerial occupations appeared

to grow while clerical and administrative occupations in the same industries shrank (Bessen 2016).

Entry Barriers

Commentators have identified several factors that might make it difficult for startups or for

their customers to compete using AI. Yet the survey shows that these factors do not pose entry

barriers for many startups in many markets. 80% of startups use customer data and 63% use data

available from third parties, including publicly available data. While data might pose a barrier to

entry in some markets, like search, where large amounts of diverse data are needed, there are

clearly many markets where it does not.

Hardware. Almost all startups provide their products over the cloud, and evidence shows

that small firms are able to access and utilize cloud computing effectively (Jin and McElheran

2017). Consequently, access to large computing hardware does not seem to be an issue for startups.

Most of the startups in our sample develop their own software for most applications,

suggesting that skilled developers and basic software tools are available to them. Anecdotally,

many software tools are available for free under Open Source licenses, such as TensorFlow.

32

Electronic copy available at: https://ssrn.com/abstract=3293275

 In many cases, large firms may have an advantage because of economies of scale or

network effects. In particular, if the fixed costs of developing a new AI application are large, then

big firms can develop their own, but medium and small sized firms may need to rely on commercial

application developers. These developers effectively spread the fixed costs over multiple

customers. The survey finds that, indeed, startups sell disproportionately to mid-sized firms,

suggesting that a) these mid-sized customers can compete using AI purchased from startups, and

b) the startups can provide a more affordable solution than if the customers developed their own

AI applications.

Data protection regulation has been described as a barrier that raises the cost of entry for

startups. We do find that data protections are higher for startups based in Europe, and large firms

appear more likely than small firms to have data protections in place. But on the other hand small

firms and non-European firms do have a range of data protections in place as well.

This evidence of the entry of startups into some markets is encouraging, however, it is also

possible that competition, entry, and innovation could all further be improved. Moreover, Figure

5 suggests that there are some industries that might have a substantial entry barriers. In particular,

in consumer packaged goods, transportation and logistics, and high tech and telecommunications,

large firms appear to invest disproportionately more than startups in AI development. The disparity

in transportation and logistics might be related to the apparently large investments needed to

credibly enter the market for self-driving cars or the competitive power of the largest firms in those

industries. The disparity in high tech might reflect the importance of network effects in markets

for online commerce and advertising or the disproportionate ownership of IP by the larger firms.

33

Electronic copy available at: https://ssrn.com/abstract=3293275

 6. Conclusion

Frey and Osborne (2017) predict that 47% of US jobs will be at risk of automation in the

next decade or so. If so many jobs were really at risk, we would see evidence of substantial job

loss associated with emerging AI applications today. This survey makes clear that is not the case.

The commercial AI applications offered by startups today are more about enhancing human

capabilities than they are about replacing humans. Only half of the firms strongly agreed that labor

cost reduction was a benefit to customers. And survey respondents replied that their customers are

using AI to create jobs in certain occupations about as often as they use it to eliminate jobs. As a

group, their applications more often prompt customers to increase employment in managerial,

professional and sales and marketing roles—contrary to some predictions—and to decrease

employment in service, clerical and manual jobs.

It is quite possible that AI might have different employment effects in the more distant

future. It is also possible that large firms might have a different impact on employment than the

startups in this survey. Nevertheless, the evidence tempers concern about mass unemployment or

disemployment of professionals. AI appears to be associated with a shift in work from some

occupation to others, meaning some people will lose jobs and other jobs will be created. However,

the new jobs may require major new skills, requiring workers to make major investments and

perhaps to endure difficult transitions. Thus, it is important to consider whether current labor

market policies are adequate to address these potential demands (Furman and Seamans, 2019). The

changing nature of demand for occupations and skills might well be highly disruptive even if there

is not mass unemployment.

Finally, the survey documents the broad scope of AI applications being offered by startups,

across a variety of industries. While the business for AI startups appears robust overall, there is

34

Electronic copy available at: https://ssrn.com/abstract=3293275

also some evidence that startups may face barriers to entry in some major industries. The extent of

these barriers, their nature, and possible policy remedies is a topic for future research. The survey

also suggests that startups use a variety of data protection mechanisms, and that startups with EU

customers are even more likely to engage in data protection. However, it is too early to assess how

GDPR is affecting AI startups, and this is another topic for future research.

35

Electronic copy available at: https://ssrn.com/abstract=3293275

 References

Accenture (2018). “It’s Learning, Just Not as We Know It.” https://www.accenture.com/us-

en/insights/future-workforce/transforming-learning.
Acemoglu, D., & Restrepo, P. (2018). Artificial intelligence, automation, and work. In The
economics of artificial intelligence: An agenda (pp. 197-236). University of Chicago
Press.
Arntz, M., et al. (2017). "Revisiting the Risk of Automation." Economics Letters 159: 157-160.
Autor, D., & Salomons, A. (2018). Is automation labor-displacing? Productivity growth,
employment, and the labor share (No. w24871). National Bureau of Economic Research.
Autor, D. H., Katz, L. F., & Kearney, M. S. (2008). Trends in US wage inequality: Revising the
revisionists. The Review of Economics and Statistics, 90(2), 300-323.
Bajari, P., Chernozhukov, V., Hortaçsu, A., & Suzuki, J. (2019). The impact of big data on firm
performance: An empirical investigation. In AEA Papers and Proceedings (Vol. 109, pp.
33-37).
Benzina, K. (2019). Cloud Infrastructure-as-a-Service as an Essential Facility: Market Structure,
Competition, and the Need for Industry and Regulatory Solutions. Berkeley Tech. LJ, 34,
119.
Bessen, J. (2016). "How Computer Automation Affects Occupations: Technology, Jobs, and
Skills." Boston University School of Law Working Paper No. 15-49.
Bessen, J. (2019). Automation and jobs: When technology boosts employment. Economic Policy,
34(100), 589-626.
Bessen, J. E., Impink, S. M., & Seamans, R. (2023). The Role of Ethical Principles in AI
Startups. Available at SSRN 4378280.
Bessen, J., Impink, S. M., Reichensperger, L., & Seamans, R. (2022). The role of data for AI
startup growth. Research Policy, 51(5), 104513.
Brynjolfsson, E., & McAfee, A. (2014). The second machine age: Work, progress, and
prosperity in a time of brilliant technologies. WW Norton & Company.
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).
Chandler, A. D. (1962). Strategy and structure: Chapters in the history of the industrial empire.
Cambridge Mass.
Chiou, L., & Tucker, C. (2017). Search engines and data retention: Implications for privacy and
antitrust (No. w23815). National Bureau of Economic Research.
Dauth, W., Findeisen, S., Suedekum, J., & Woessner, N. (2018). Adjusting to robots: Worker-
level evidence. Opportunity and Inclusive Growth Institute Working Papers, 13.

36

Electronic copy available at: https://ssrn.com/abstract=3293275

Ewens, M., Nanda, R., & Rhodes-Kropf, M. (2018). Cost of experimentation and the evolution
of venture capital. Journal of Financial Economics, 128(3), 422-442.
Felten, E., Raj, M., & Seamans, R. (2021). Occupational, industry, and geographic exposure to
artificial intelligence: A novel dataset and its potential uses. Strategic Management
Journal, 42(12), 2195-2217.
Ford, M. (2015). The rise of the robots: Technology and the threat of mass unemployment.
International Journal of HRD Practice Policy and Research, 111.
Frey, C. B. and M. A. Osborne (2017). "The future of employment: How susceptible are jobs to
computerisation?" Technological Forecasting and Social Change, 114: 254-280.
Furman, J., & Seamans, R. (2019). AI and the Economy. Innovation Policy and the Economy,
19(1), 161-191.
Goos, M., Manning, Salomans (2009). "Job Polarization in Europe." American Economic
Review, Vol. 99: 58-63.
Graetz, G., & Michaels, G. (2017). Is modern technology responsible for jobless recoveries?.
American Economic Review, 107(5), 168-173.
Impink, S. M. (2023). Both sides of cloud: More standardized tools and more diverse data
analytics. Available at SSRN 4200675.
Jia, J., Jin, G. Z., & Wagman, L. (2021). The short-run effects of the general data protection
regulation on technology venture investment. Marketing Science, 40(4), 661-684.
Kaplan, J. (2015). Humans Need Not Apply: A Guide to Wealth & Work in the Age of Artificial
Intelligence. Yale University Press.
Katz, L. F. (1999). Changes in the wage structure and earnings inequality. In Handbook of Labor
Economics, (Vol. 3, pp. 1463-1555). Elsevier.
Khan, L. M. (2017). Amazon’s antitrust paradox.
Lambrecht, N. and C. Tucker (2017). "Can Big Data Protect a Firm from Competition." CPI
Antitrust Chronicle.
Manyika, J., Chui, M., Miremadi, M., Bughin, J., George, K., Willmott, P., & Dewhurst, M.
(2017). A future that works: AI, automation, employment, and productivity. McKinsey
Global Institute Research, Tech. Rep, 60, 1-135.
Miric, M., Ozalp, H., & Yilmaz, E. D. (2023). Trade‐offs to using standardized tools: Innovation
enablers or creativity constraints?. Strategic Management Journal, 44(4), 909-942.
Manyika, J., Lund, S., Chui, M., Bughin, J., Woetzel, J., Batra, P., ... & Sanghvi, S. (2017). Jobs
lost, jobs gained: What the future of work will mean for jobs, skills, and wages.
Peukert, C., Bechtold, S., Batikas, M., & Kretschmer, T. (2022). Regulatory spillovers and data
governance: Evidence from the GDPR. Marketing Science, 41(4), 746-768.
Rock, D. (2019). Engineering value: The returns to technological talent and investments in
artificial intelligence. Available at SSRN 3427412.

37

Electronic copy available at: https://ssrn.com/abstract=3293275

Rosenberg, N. (1963). Technological change in the machine tool industry, 1840–1910. The
Journal of Economic History, 23(4), 414-443.
Stucke, M. E., & Grunes, A. P. (2016). Introduction: big data and competition policy. Big Data
and Competition Policy, Oxford University Press (2016).
Susskind, R. and D. Susskind (2015). The Future of the Professions. Oxford, Oxford University
Press.
Varian, H. R. (2010). "Computer Mediated Transactions." American Economic Review, 100(2):
1-10.

38

Electronic copy available at: https://ssrn.com/abstract=3293275

 Appendix

Appendix A.1. – Additional reporting on AI ethics responses

All data shown here comes from surveys 3-5 as these questions were not included in the

first tow rounds of the survey. These figures need to be viewed in color.

Figure A.1. – Types of job augmented by AI products (versus being replaced)

40

30
% of respondents

Minority Hiring
More Diverse Data
20 Expert Advice
Employee Trainings

10

0
Survey 1 Survey 2 Survey 3 Survey 4 Survey 5

39

Electronic copy available at: https://ssrn.com/abstract=3293275

 Figure A.2. – Types of job augmented by AI products (versus being replaced)

.8

.6
% of respondents

Minority Hiring
More Diverse Data
.4 Expert Advice
Employee Trainings

.2

0
No Policy Ethical AI Policy

Figure A.3. – Types of job augmented by AI products (versus being replaced)

50

40
% of respondents

30
Turned Down Business
Dropped Data
Dismiss Employee
20

10

40

Electronic copy available at: https://ssrn.com/abstract=3293275

 Figure A.4. – Types of job augmented by AI products (versus being replaced)

50

40
% of respondents

30
Turned Down Business
Dropped Data
Dismiss Employee
20

10

0
Survey 1 Survey 2 Survey 3 Survey 4 Survey 5

Figure A.5. – Types of job augmented by AI products (versus being replaced)

50

40
% of respondents

30
Turned Down Business
Dropped Data
Dismiss Employee
20

10

0
Smaller (<11) Larger

41

Electronic copy available at: https://ssrn.com/abstract=3293275

 Figure A.6. – Types of job augmented by AI products (versus being replaced)

50

40
% of respondents

30
Turned Down Business
Dropped Data
Dismiss Employee
20

10

0
Americas Asia Europe

Figure A.7. – Types of job augmented by AI products (versus being replaced)

.25

.2
% of respondents

Human-in-loop
.15 Int. Audits
A/B Testing
Oversight Board
Ext. Audits
.1

.05

42

Electronic copy available at: https://ssrn.com/abstract=3293275