Scaling AI Ventures: How to Navigate Tensions between Automation and Augmentation

Abstract

AI ventures promise to automate and augment ever more human tasks. This provides rich opportunities for growth. Yet, digital and human resources that involve AI are oftentimes task-specific and hard to scale. Furthermore, clients remain skeptical to be fully automated by external services. Thus, it remains unclear how AI ventures achieve growth. We adopt a grounded theory approach on an interview study with founders, product managers and investors to inquire how resources afford or constrain scaling in AI ventures. For this, we blend the notion of (non-)scale free resources with the layered architecture of digital technologies. Our study suggests that AI ventures scale by organizing digital and human resources for replicability in that they keep AI-specific resources distant from clients while simultaneously externalizing human-intensive tasks to their clients. As we inquire the roles of human and digital resources, our study suggests that ventures seek to quickly find an optimal degree on the continuum between augmentation and automation when bundling resources.

Full text

Scaling AI Ventures: How to Navigate Tensions between Automation and
Augmentation
Jonathan Zebhauser
Freie Universität Berlin
j.zebhauser@fu-berlin.de
Hannes Rothe
ICN Business School
hannes.rothe@icn-artem.com
Janina Sundermeier
Freie Universität Berlin
janina.sundermeier@fu-berlin.de
Abstract
AI ventures promise to automate and augment ever
more human tasks. This provides rich opportunities for
growth. Yet, digital and human resources that involve
AI are oftentimes task-specific and hard to scale.
Furthermore, clients remain skeptical to be fully
automated by external services. Thus, it remains
unclear how AI ventures achieve growth. We adopt a
grounded theory approach on an interview study with
founders, product managers and investors to inquire
how resources afford or constrain scaling in AI
ventures. For this, we blend the notion of (non-)scale
free resources with the layered architecture of digital
technologies. Our study suggests that AI ventures scale
by organizing digital and human resources for
replicability in that they keep AI-specific resources
distant from clients while simultaneously externalizing
human-intensive tasks to their clients. As we inquire the
roles of human and digital resources, our study suggests
that ventures seek to quickly find an optimal degree on
the continuum between augmentation and automation
when bundling resources.
Keywords: AI Venture, Scaling, Grounded Theory
Method, AI Startup, Scale Free Resources.
1. Introduction
Artificial intelligence (AI) marks a new era of
information systems management and implies
interacting with an ever-evolving frontier of
technological advancements in the context of decision
making (Berente et al., 2021). AI seemingly gets more
and more integrated into our societies (Rahwan et al.,
2019) and greater numbers of ventures consider AI
essential for their products and services (Weber et al.,
2021). We refer to these types of ventures as AI
ventures. AI ventures produce market offerings that
change the value is being created and captured (Bughin
et al., 2018; Chui et al., 2018; Fontana, 2021; Iansiti &
Lakhani, 2020), because they promise to automate ever
more tasks done by humans or at least augment humans
in making better decisions (Raisch & Krakowski,
2021). Together, automation and augmentation provide
ample opportunities for venture growth if being able to
scale their available resources. We refer to growth as the
change in a relevant measure of firm size and scaling as
the relationship between multiple measures of size, e.g.,
available financial, human or digital resources (Schulte-
Althoff et al., 2021; West, 2017). Similar to other digital
ventures, AI ventures draw from the advantages of
increased modularity, flexibility, and malleability of
digital infrastructures (Henfridsson, 2020; Henfridsson
et al., 2014; Yoo et al., 2012) for growth (Huang et al.
2017). However, AI ventures seem not to scale like
other digital ventures as they leverage AI-specific
resources, that differs how specific their digital
resources are or entails people with different expertise
(Casado & Bornstein, 2020; Chui & Malhotra, 2018;
Giustiziero et al., 2021; Linde et al., 2020; Schulte-
Althoff et al., 2021).
To understand the problem that AI ventures face
when scaling, we adopt the notion of non scale-free
resources, including human resources. That is, each
additional unit implies an equal increase in costs
(Burström et al., 2021; Khan et al., 2020; Sjödin et al.,
2021). In fact, AI ventures are unique in how they
combine human with digital resources, e.g., digital
technology and data annotated with human input. This
dependency on non scale-free resources hamper
repeated value creation (Jöhnk et al., 2021) and hence,
it has a decisive impact on the venture’s scaling
ambitions (Levinthal & Wu, 2010) as it poses a limit
to growth (Penrose, 2009). For AI ventures it can take
much longer to repeatedly capture value from the same
bundles of digital resources, because new clients
require investment of additional human resources for
producing, annotating, integrating new data, or
updating machine learning (ML) models (Casado &
Bornstein, 2020). In addition, testing and monitoring
AI applications requires more human oversight than
testing and monitoring rule-based software, as it is
difficult to specify data and ML model behavior a
priori (Breck et al., 2017). As a consequence, recent
findings indicate that the average AI venture shows a
similar demand for human resources as service

ventures when growing (Schulte-Althoff et al., 2021).
This is surprising in that it counters the intuition that
augmentation and automation with AI should support
scaling. We therefore ask: How do AI ventures
organize digital and human resources for scaling?
How do they repeatedly and at greater pace create and
capture value from these resources? To answer our
research questions and to support building theory
around this phenomenon, we conduct a qualitative
research study. We use a modified grounded theory
method (GTM) approach study following Gioia et al.
(2013) to build theory from practice. For this, we draw
from experiences of experts in scaling AI ventures.
Our qualitative study is based on twelve expert
interviews that include diverging views of founders,
product and business developers and investors, all in
the context of AI ventures. During our analysis, we
learned that AI ventures organize digital and human
resources differently, depending whether these
resources relate to the content, service, network, or
device layer of the digital infrastructure (Yoo et al.,
2010). We present our findings in three aggregated
dimensions and four second-order themes. Our
research indicates that AI ventures scale by organizing
digital and human resources for replicability and far
away from the client while simultaneously
externalizing human-intensive tasks to their clients. In
order to become better in this repetition, ventures build
a supportive digital infrastructure around externalized
tasks, such as data annotation. At the same time, we
learn that clients shy away from being automated by
products and services of AI ventures. We thereby
show that AI ventures face the unique problem of
finding an optimal degree in the continuum between
automation and augmentation. AI ventures seek to
maximize the use scale-free resources, which
increases automation while also following their
client’s call for augmentation. The study contributes to
theoretical and practical discourse about
understanding mechanisms behind scalable AI
ventures as well as by blending the notion of scale free
resources and non-scale free resources with the
layered architecture of digital technology to offer
further perspectives on AI ventures and scaling.
2. Conceptual Background
2.1 AI Ventures
AI can be described as an ever-evolving frontier
of computational advancements in the context of
decision-making problems (Berente et al., 2021). It
enables ventures to create new products, business
models and services (Brynjolfsson & Mcafee, 2017;
Makridakis, 2017). Industry-agnostic, AI technology
could have an impact in all industries and sectors.
Therefor it can be understood as a general purpose
technology (Brynjolfsson & Mcafee, 2017;
Trajtenberg, 2018). Following Penrose’ understanding
of firms (Penrose, 2009), ventures are a set of resource
bundles that differ in their ability to produce scale. AI
ventures are built on digital and human resources.
Digital resources entail software code and digital data.
They can be AI-specific, e.g., annotated data, or
software code that enables machine learning
algorithms. Also human resources can have AI-
specific roles, e.g., data scientists, machine learning
engineers, data product managers, machine learning
researchers which can be summarized as AI experts.
Combining these resources is an important aspect of
applying AI. The “human in the loop”, for instance, is
supposed to evaluate, interfere or tune decisions
enshrined in software code of AI algorithms. Humans
manage date in that they annotate, integrate and
maintain digital data that is supposed to feed machine
learning. Besides these AI specific resources there are
complementary non-AI specific resources necessary
such as cloud infrastructure, software engineers, or
product designers (Fontana, 2021; Metelskaia et al.,
2018). Ventures differ in how they combine these
resources (Weber et al., 2021) in that they, for
instance, combine pre-trained machine learning
models with software code for a mobile application to
address demands of multiple customers. Similarly,
ventures may also produce custom data and software
code to create unique AI algorithms within custom
service relationships. These different combinations
might have different effects on a venture’s ability to
scale.
2.2. Scaling AI Ventures
Iansiti and Lakhani (2020) propose the notion that
AI allows ventures to create and capture value through
three mechanisms: first, through the repetition of value
creation within a domain, second, through the
repetition of value creation between domains and
third, through attaining the resources for repetitive
value creation and capture. To understand scaling in
our context, we first look at digital ventures, of which
AI ventures form a subset. Digital ventures draw on
digital infrastructure, which enables them to execute
their actions on a given structure (Henfridsson, 2020).
This setting facilitates two scaling modes: scale due to
design flexibility and scale due to design scalability
(Henfridsson et al., 2014). Digital ventures have a
flexible design that allows rapid reaction to changing
circumstances because they use a digital infrastructure
that is not pre-defined (Henfridsson et al., 2014).
Following the work done by Yoo et al. (2010), we

understand the architecture of the digital firm as a
layered modular architecture. It consists of four layers:
contents-, service-, device-, and network layer. The
contents layer contains the data, while the service layer
represents the application functionality that serves the
user. The network layer is divided into a physical
transport layer including hardware and a logical
transmission layer which includes protocols or
network standards. The device layer is also divided
into a physical machinery layer consisting of e.g.,
computer hardware, and a logical capability layer that
consists of e.g., an operating system. The layered
modular architecture combines the modular
architecture and the described layered architecture.
Modular architecture is structured through
standardized interfaces between components - the
highest degree of modularity makes these components
product-agnostic (Yoo et al., 2010). Digital ventures
have a scalable design because of low costs of
replication of digital resources (Giustiziero et al.,
2021; Shapiro et al., 1998). As scalability is achievable
for most digital ventures, scaling faster than
competitors is important, especially when it comes to
winner-take-all markets (Cohen & Levinthal, 1990;
Schilling, 2002). In conclusion, time moderates
scaling activities and their success in respect of
competing ventures. Studying a credit rating company
that was scaling on a rapid pace, Huang et al. (2017)
traced three mechanisms that enable rapid scaling,
which are data-driven operation, instant release, and
swift transformation. Data-driven operation enables
framing, hedging and monitoring of opportunities and
activities of innovation with data. Instant release
enables fast deployment of innovation ideas, only with
a short time-lag. Swift transformation enables the
effortless contextualization towards new value-in-use
aligned with an updated venturing identity (Huang et
al., 2017). Through these mechanisms the speed of
scaling operations can be increased, drawing on
productive techniques at a high pace (Henfridsson,
2020). While they highlight opportunities, challenges
for other ventures adopting these strategies remain
unclear.
2.3. Scale free and non-scale free resources
Ventures, considered as bundles of resources
(Penrose, 2009), depend on characteristics of these
resources especially when it comes to scaling. Some
resources, like brand names, scale almost without
boundaries in that they can be replicated across many
domains (scale-free), other resources can not be easily
replicated and therefore produce limited scale (non
scale-free) (Levinthal & Wu, 2010). The application
of such non-scale free resources depends on the
opportunity costs of deploying them in another domain
(Levinthal & Wu, 2010). Thus, ventures have to assess
the utility of deploying resources in one domain
compared to another, especially when considering to
offer products across market segments (Levinthal &
Wu, 2010) or when testing different markets. This is
important, because ventures are initially unaware of
the market segments in that they are eventually able to
scale (Giustiziero et al., 2021). This implies that when
a resource bundle is scalable, the opportunity cost of
deploying these resources anywhere else than in the
ventures focal domain is high, as it contains
complementary resources in form of specialized
human and managerial resources. Digital resources are
supposedly scale-free as they can be replicated almost
error-free, being globally distributed at low costs and
steadily improved in performance and costs as more
they are used (Agrawal et al., 2018; Brynjolfsson &
McAfee, 2014; Giustiziero et al., 2021). More
specialized resources need to be managed with
opportunity costs in mind. Human resources cannot
simultaneously develop software code for a generic
product while also conducting service for a specific
client project. This is especially important the more
unique, rare, and highly regarded these human
resources are. AI experts fall into that category due to
their level of training in software development, data
analysis, statistics, and/or management training, which
might put a unique burden on AI ventures.
3. Methodology
To understand how AI ventures organize their
digital and human resources for scale and replicable
value creation and capture, we used a qualitative
research approach. Our research design follows a
modified grounded theory methodology (GTM)
approach (Corbin & Strauss, 2015; Gioia et al., 2013)
in the process of analysis, which is suited to
understand and explore an IS-related phenomenon in a
complex environment (Wiesche et al., 2017). Our
approach to data collection differs from GTM: We
prepared for an empirical exploration of our research
field using extant theory as Goldkuhl and Cronholm
(2019) propose.
Due to the evolving frontier of AI and its
implications on business we saw the need to inform
ourselves to be able to reach an insightful level in the
conversation with experts. Authors we draw from
regarding the topics AI, constraints of scaling and
current forms of value creation and capture in AI
ventures are, among others, Anderson and Tushman

(1990), Casado and Bornstein (2020), Fontana (2021),
Giustiziero et al. (2021), Iansiti and Lakhani (2020),
and Schulte-Althoff et al. (2021). For the analysis we
adopted the approach of Gioia et al. (2013). In the first-
order analysis we identified and used empirical codes
and terms that seemed central to the interviewees. In
the second-order analysis we identified theoretical
concepts related to our empirical observations, before
finally turning to further abstraction in aggregate
dimensions. Interviews are a common method for data
collection in GTM studies (Corbin & Strauss, 2015;
Gioia et al., 2013). To capture individuals’
experiences and perspectives framed by our research
focus, we chose to conduct semi-structured expert
interviews. To visualize our findings, we followed the
approach of Gioia et al. (2013).
3.1. Data Collection
Our sampling covers different levels of insights
into how resources afford or constrain scaling in AI
ventures. We expected founders to reflect on the impact
of initial resources in their ventures as well as their
strategic development. Product and business developers
were supposed to provide insights into operating digital
resources in grown ventures. Investors active in
multiple AI ventures were supposed to provide a
perspective of a well-informed third-party. Our sample
was supposed to include ventures incorporating value
creation and value capture from divergent resource
bundles. We used the business model typology from
Weber et al. (2021) to operationalize these divergency
in our sample. As researching AI ventures in general is
a high level view on a field with great differences
regarding sector and customer dynamics, we included
experiences with different industries, B2B and B2C
markets as well as different regional focuses (US, India,
EU). To keep the exploration space open-ended we used
the critical incidence technique (Flanagan, 1954) in our
semi-structured interviews, which is suited for
accessing practical knowledge (Langley & Meziani,
2020). We did not ask for mechanisms known for
scaling digital ventures in IS research or known
problems from theory directly, to avoid leading the
answers into a specific direction (Bogner et al., 2014).
3.2. Data Analysis
Figure 1 illustrates our data structure and
summarizes our findings. We iteratively build our
inductive theory using a modified grounded theory
approach (Gioia et al., 2013). We used the coding
software MAXQDA to analyze our interview
transcripts and performed open coding on all cases. This
led to an initial collection of codes on which we draw
our first collection of first-order categories. To derive
labels for our first-order categories, we used abstraction.
Following a first iteration of open coding, we switched
to axial coding. Here, we found that a conceptual
framework could help us structure the impact of
different digital resources on an AI ventures’ ability to
scale. We chose the concept of a layered modular
architecture (Yoo et al., 2010) and its corresponding
layers to link our findings to the digital infrastructure of
an AI venture, as described in the conceptual
background. This iteration helped us sort first-order
categories and organize them into second-order themes.
While the first-order concepts are still close to our raw
data, the second order themes have a strategical quality,
inspired by Huang et al. (2022). The aggregated
dimensions help us embed our findings in theory and
therefor have a theoretical quality. Using layered
modular architecture as a conceptual framework, we
saw that keeping it as a part in the aggregated
dimensions is useful and helps to transmit our
theoretical contribution.
ID Function Stage of Venture Business model typology that
covers divergent resource bundles Market Interview
Duration AI Experience
ID1 Founder Grown Venture AI-charged Product/Service Provider B2B 31:12 min Over 10 years
ID2 Founder Grown Venture Data Analytics Provider B2B 76:57 min Over 20 years
ID3 Managing Director Early Stage Venture AI Development Facilitator B2B 49:00 min Over 5 years
ID4 Founder Mid Stage Venture Deep Tech Researcher B2B 22:55 min Over 5 years
ID5 Product Manager Grown Venture Data Analytics Provider B2B 73:27 min Over 10 years
ID6 Founder Early Stage Venture AI-charged Product/Service Provider B2C 39:29 min Over 2 years
ID7 Investor Series A Startups - All 63:56 min Over 5 years
ID8 Founder Mid Stage Venture Data Analytics Provider B2B 33:22 min Over 10 years
ID9 Business Developer
AI Venture Builder All Patterns All 46:29 min Over 2 years
ID10 Product Manager Mid Stage Venture AI Development Facilitator B2B 35:49 min Over 7 years
ID11 Investor Series A Startups - All 36:58 min Over 7 years
ID12 Product Manager Mid Stage Venture AI Development Facilitator B2B 18:36 min Over 7 years
Figure 1. Data Structure.

4. Results
Following Gioia et al. (2013) we present our first-
and second-order Figure 1. As mentioned before, the
layered modular architecture (Yoo et al., 2010)
revealed itself as a scaffolding to sort our findings as
we went back and forth between theory and qualitative
data. In the following, we explain each dimension
according to his scaffolding.
4.1. Contents layer organized for scale
An important means to create scale free resource
bundles for AI ventures involves relationships with
clients on the contents layer. The corresponding first
order concepts are displayed in figure 1.
Joint bundling of scale-free digital resources
from the venture with non-scale free human
resources of externals: Value creation through AI
depends heavily on data. Preparing, integrating and
monitoring new data tie resources. Scale at rapid pace
can only be achieved if these tasks are whether highly
efficient or externalized. Data sources vary depending
on the market segment the venture operates in. Our
findings show that the source of data, its accessibility
and availability play an important role if aiming for
scale as a venture. If data sources are diversified and
its management cannot be fulfilled internally without
loosing resources, an infrastructure for externalizing
the corresponding tasks has to be created.
First, to organize such scale free data sources,
three ways of acquiring data were highlighted by most
interviewed experts: buy data, partner for data or use
open-source data. Speaking about an AI venture
working in health-sector, expert 10 states: “We either
buy data, create partnerships or find ways that we can
collect data. [...] Sometimes the health systems have
let us build up a data set based on the product that we
were building in pilot with them”. In other market
segments, reliable open-source data may be the way to
go, as expert 11 explains: “But most either buy
datasets or use existing open-source resources that
are already there. We also see this very often,
especially for example in the Geo-Spatial area [...]”.
Second, most of the interviewed experts pointed
to the importance of rapid and long-term access to the
data which the product gets built on. Binding data
through contracts is an option, expert 8 recommends:
“You tell them: If you want to use this technology, you
have to give us the data […] And then they either do it
or they don’t, and that’s why the contract is concluded
or not”. As expert 11 states, that this is also important
when selecting clients: “It must be ensured that this is
not just a one-time customer, but someone who can
imagine working with us for many years, because this
is the only way to ensure financial sustainability and
data volume”. Thus, assuring rapid and long-term
availability of the data used by the AI venture is one
aspect to consider when building a scale free contents
layer.
Third, most interviewed ventures enable and
incentivize their clients to annotate data themselves in
that they produced tools to make these tasks easy, e.g.,
Figure 1. Data Structure.

building intuitive interfaces or creating “no-code”
applications that required little prerequisite knowledge
in data management. Also, this environment should be
incentivizing, so that clients annotate the data and
integrating it by themselves. Most interviewees stated
that simplifying the data integration process for the
users is important. Expert 5 states: “[...] you have to
think of getting to a self-serve kind of model [...] that
is the only way to scale your company [...] The trick is
in figuring out incentivizing your users.”. In B2B cases
some client data may already have been annotated,
e.g., stocktaking. Therefore, ventures found that data
integration could be automatized to a high degree - one
of the ventures (ID8) lowered the data integration of
new B2B clients to approx. two hours. Therefore, this
venture is able to deploy their own resources for the
process. The same venture uses CAD models of clients
to create synthetic visual data with an automatized
process, resulting in even higher annotation accuracy.
One venture (ID2) we spoke with gives clients access
to APIs which the clients then have to use. In that way,
clients are forced to handle data management on their
own - still, this venture offers support which earlier
was done by internal experts and now is handled
through partnerships with consultancies, another way
to keep the contents layer scale free.
Our conclusion is that ventures facilitate a scale
free contents layer in their digital infrastructure that
enables clients to annotate, integrate and manage data.
Thereby, clients remain inherently involved in the AI
application which, even if supposed to be highly
automatic, thereby still offers augmentation functions.
While in a lot of domains building a scale free resource
bundle for full automation may be difficult to achieve,
an alternative is figuring out a replicable process for
externalizing these tasks to the client. As data quality
and data distribution influence the scale and cost of the
value creation, our interviewees’ experience shows
that non-experts working on data had to be educated
and monitored by experts in order to ensure that the
venture can still replicate their use of digital resource.
4.2. Service layer organized for scale
Joint bundling of resources is positioned as client
augmentation: Scaling value creation requires
repeatedly serving clients in similar ways. Most of the
interviewees learned that clients, however, do not want
to be “automated”, as, e.g., expert 4 describes: “[...]
there were always discrepancies between the
management, which had the pressure to become more
productive, and the actual researchers [users on
client-side], who said, ‘Well, if we think this through
to the end, then you don’t need me anymore.’”. Thus,
ventures need to walk the fine line, replicating as many
resources as they can in order to scale, while also not
losing the client’s trust to not automate tasks that are
typically done by human users. Most interviewees
mention that offering education to foster
understanding of the AI technology in use on client
side facilitates value capture at greater pace by
minimizing support interventions. Through delivering
products that clients perceive as “augmenting”, the
fear of being automated by AI can be addressed. We
use the term “service layer organized for scale” to
denote that the resource bundle involved in the service
layer should be organized to be scale free. This might
ask for specific strategies, as clients might want to be
“in the loop” and not fully automatized.
First, interviewees mention, that the fear of being
replaced by AI often accompanies stakeholders.
Therefor ventures may position their offer as
augmentative instead of replacing, as expert 5 states:
„The biggest challenge with AI-Systems has been that
it can replace the people who are potentially going to
adopt it.”. Remaining in the loop rather than relying
on full automation keeps the decision-making
authority on how the AI behaves at least partly in the
customer’s perceived sphere of influence.
Second, with the goal of reaching a high degree of
automation as well as a customer/product interaction
that requires as little support and intervention from the
venture as possible, some of the companies we
interviewed enable their clients to control grade of
automation. Expert 2 explains that they design their
products interface in a way that the customer has both
the ability to intervene in the automated result but also
learns over time to interfere less, as it generally leads
to less optimal results. Expert 12 reports that they
work with a continuum that allows the user to
determine the level of automation provided by the
application. In their experience, this has always led to
decreasing intervention by customers, as they
increasingly trust the automation and the motivation to
intervene themselves decreases as a result. Enabling
clients to control the grade of automation by deciding
when and how often to intervene, while educating
transparently about the efficiency of the automation
and comparing results of decisions made by users and
the product, ventures may reduce the resources tied to
client intervention and reach a higher degree of
automation on the client side.
Third, concepts such as statistical uncertainty are
hard to grasp. Lack of knowledge on how AI works
can lead to clients escalating more often, as expert 2
describes: “Escalation occurs and we realize that
nothing is going on. They just didn’t understand that
there are always statistical uncertainties.”. Expert 5
adds: “So like most of these technical products usually
need a good hand-holding with go to market teams and

educating your customers on how to use your
products.”. Client escalation ties up support resources
and reduces the pace of value capture. To allow for
uninterrupted value capture ventures educate their
clients in concepts like prediction and uncertainty.
Client education should be developed as a scale-free
resource which scales with its demand. As this can be
difficult to achieve within a venture, partnering with
external consultants can be a solution. Our theoretical
conclusion is that by offering augmentation while
minimizing client input long term by education on the
service layer can be organized for scale.
4.3. Network / device layer organized for scale
Prioritize production of scale-free digital resources
by keeping non-scale free resources of the venture
distant from clients: Ventures need an infrastructure
that enables the creation of replicable resource
bundles. Thus, the resource bundle involved in the
network and the device layer and its corresponding
activities have to be organized to be scale free. First,
we learned that in order to avoid solving unique
problems of individual clients with custom solutions,
AI experts were kept at distant from the client. This
enabled them to fully focus on building scale free
digital resources, as expert 1 describes: “the exciting
question is, do you get the people who are, I’ll call
them AI experts, do you get them decoupled enough
from the respective customers that they actually build,
I’ll say standard products?”. Also, expert 2, the most
senior AI expert we spoke to, explains: “As a scientist,
of course, I found it great when we had a lot of
different applications based on the same core idea.”.
Interestingly, we noticed that this problem was
perceived by our informants as both sided: clients
were considered to ask for custom solutions that meet
their particular needs, while AI experts seem to have a
strong emphasis on solving hard and particularly new
problems and therefor are keen to develop as much
custom solutions as they can. Thus, keeping AI experts
at some distance to client conversations and their
inquiries, lowered the chances of producing custom
solutions.
Second, to scale fast all interviewed ventures rely
on cloud technology at the network layer. While this
can be costly, it allows for rapid scale and
internationalization, even though data privacy
preferences might enforce a shift to a localized cloud
solution in some cases. Third, some of the experts
interviewed clarify that the internal goal is full
automation, which is often not equally perceived on
client-side, as, e.g., expert 11 states: “So, internally,
automation is very hot. Externally, however, I don’t
think it’s being sold quite as much yet.”. Expert 2
explains: “We are in favor of fully automating this, of
course, but many customers don’t want that.”.
Designing for a high grade of automation enables
strong replicability on the network and device layer.
Fourth, to assure scalable linkages to resources
provided by partners such as cloud providers, a
scalable price model has to be established. Expert 2
explains their use of: “[...] a model where we define a
price ladder or a range and also a scaling law for a
certain product based on our experience and value
and effort estimation[...]”. Also, expert 5 underlines
the importance of measuring costs continuously to find
the right margins to charge. Expert 11 explains the two
price models she observes in their portfolio of AI
ventures: Software as a Service, or pricing by
consumption, e.g., data processing fees. Our
conclusion is that ventures should create an
infrastructure that bundles non-scale free resources,
especially AI experts, with digital resources on the
network and device layer. This enables focal work on
replicable products which are then scale free.
Prioritize bundling of resources that can be
replicated for multiple clients: First, repeatedly
capturing value is important for ventures who seek to
grow quickly, which is why producing a resource
bundle on that can be replicated easily becomes
important for AI ventures. Producing replicable bundles
of digital resources from the network and device layer
becomes important for AI ventures to avoid becoming
an agency for clients that creates one-time payments
and project-based revenue. Consulting other AI
ventures, expert 1 explains this as a common hurdle that
has to be avoided: "[...] I’m really just sort of always
asking: What are you doing to not become an agency?”.
Expert 2 explains this with their own story: “At the very
beginning [...] this was still different for customer A
from customer B, but we really got to the point where
they were all exactly the same [...] before [...] we could
essentially only scale by hiring even more data
scientists”.
Second, in order to prioritize resource bundles that
can be replicated, the AI venture needs to find
commonalities between clients. Expert 7, an investor,
highlighted that she uses such commonalities between
clients as an important indicator for assessing potentials
for growth in AI ventures. In an early stage, finding
commonalities involves experimentation. These
experiments involve engagements with individual
clients that help produce digital resources that also serve
other clients. This may translate to less current revenue
for the venture, as interviewees underline.
Third, quitting clients that require digital resources
that cannot be replicated in other client engagements is
vital, this involves new sets of data, other expertise of

AI experts, or diverging ML models. Thus, keeping
various clients at the same time seems important.
Interviewees mention decisions might be tough at an
early stage, but is the only way to organize the venture
for scale in the long term. Expert 2 states: “We also had
to force a few customers: You have to take our product
now; we won’t develop the other one further”. Expert 8
explains how they solve inquiries for specifications that
do not match their offer: “[Sometimes] we say: You
guys, we can’t offer that, but we have a development
partner and he does it at good conditions”.
Fourth, while big orders by clients are an
opportunity for capturing value short, they may lead
into the “agency trap”: being dependent on only one or
a few clients positions the venture as an agency rather
than a product provider. Both expert 1 and 2 emphasizes
how a product focus protected them from getting
individual major customers for which they would have
become “an internal IT shop”.
5. Discussion
Our study identified three key dimensions that AI
ventures may address when striving for the creation of
scale free resource bundles, which is important because
these bundles allow AI ventures to attain growth. Our
work contributes to the ongoing discourse on AI and
scaling in three major ways.
First, our qualitative inquiry reveals how ventures
tackle a unique tension between automation and
augmentation, i.e., deciding to what extent they include
human and digital resources into their resources
bundles. Second, we found that scaling depends largely
on a venture’s ability to replicate resource bundling
which requires important decisions on securing rare
(human) resources on the service and network layer.
Third, we find that resources on the contents layer
is an important bottleneck for scaling which ventures
solve by introducing means to externalize data. In order
to scale, AI ventures are forced to organize their digital
and human resources in bundles. At the same time
human and managerial resources, which are generally
not scale free, are deployed on occasion only and not as
permanent actors in these processes. Figure 2
summarizes our core findings.
Tension between Automation and Augmentation:
Ventures aim to maximize replication of resource
bundles that could be sold to clients. Producing AI-
specific digital resources, such as ML models or data,
that allow to ventures to increase automation is a corner
stone of this endeavor. Clients, however, demand
continuous adjustments of resources bundles to their
specific needs and therefore would like to retain
“humans in the loop”. This reveals a tension between
automation and augmentation that is generally
acknowledged (Raisch & Krakowski, 2021), but seems
unique for AI ventures in that these firms need to scale.
AI ventures have to find an optimal point in the
continuum between automation and augmentation.
Larger companies can afford to develop AI produce AI-
based resource bundles that involve humans (i.e.,
augmentation) order to slowly produce more and more
data, expertise in creating algorithms, and build client
trust so that they later introduce resource bundles that
do not include human resources (i.e., automation)
(Raisch & Krakowski, 2021). Digital ventures,
however, are supposed to scale rapidly and early on
(Huang et al., 2017). Thus, while larger firms can
gradually produce these bundles (first augmenting, later
automating), AI ventures strive for replication early on.
Our study exemplifies that ventures learned how
include clients “in the loop” who effectively ask for
“augmentation”, while not falling into an “agency trap”.
This shows, how AI ventures face the inherent struggle
of creating AI-specific resource bundles that need to be
adjusted to client demands while avoiding to suffer
from reduced scale. One way how AI ventures solve
that issue is by creating tools that allow for the use of
scale-free resources on the contents layer, i.e.,
externalizing data annotation. At the same time, venture
keep their own non-scale free resources, especially AI
experts, distant from the client in order to avoid falling
into an “agency trap”, i.e., becoming an AI agency that
produces only custom AI applications. The tension is
depicted in Figure 2.
Replication of Resource Bundling: AI ventures
deploy their human resources to produce a core resource
bundle that serves a broad market segment and enforce
its adaption by denying adaptation of these resources
Figure 2. Layers Built for Scale in an AI Venture.

that would serve a diversified portfolio of clients. This
way, AI ventures avoid becoming agencies for the
client. Instead, they let feedback and user research
iteratively shape a replicable product based on
commonalities, as depicted in figure 2, while not
sharing their AI expertise with the clients. By focusing
on a particular resource bundle, the AI ventures more
quickly produces resources, e.g., algorithms, human
knowledge, or data, that improves the quality of this
resource bundle. Human and managerial resources are
constantly being made available for this purpose. A
scale free tech stack based on state-of-the-art methods
such as modularity and cloud infrastructure help to
build a replicable resource bundle. While these findings
reiterate how digital ventures learn from client data in
order to scale (Huang et al., 2017), it also exemplifies
how AI ventures need to balance their willingness to
create a scalable resource bundle with the necessity to
hide its inner workings from the client. It might
therefore explain, why AI ventures scale more
comparable to service ventures than digital platforms
(Schulte-Althoff et al., 2021).
Contents Layer organized for Scale: We learned that
clients were hesitant to be fully automated by an AI
venture and rather ask for being augmented, because
they can retain some control over the impact of the AI.
Enabling human resources of a client to be bundled with
digital resources of the AI venture therefore plays two
parts. It allows augmentation but also plays into the
hands of AI ventures who seek to delimit the use of their
own human resources to avoid opportunity costs
(Penrose, 2009). This is especially important in cases
where ventures work with client data, because data
requires much human labor, e.g., for maintenance and
labeling. Data is important for many AI-based
resources, e.g., ML models (Fontana, 2021). Open data
oftentimes does not suffice for producing a purchasable
resource bundle. Instead, AI ventures may need to
create proprietary data that would be unique for every
client relationship. In order to avoid engaging in
forming unique resource bundles for every client
relationship, AI ventures strive to produce replicability
on the contents layer by producing software and
standardizing (meta) data, e.g., for labeling. If a venture,
that offers an API to their AI for their clients adds
hundred clients overnight, data annotation and
integration might pose a bottleneck. If the infrastructure
allows clients, however, to annotate and integrate the
data themselves and with little human support, hundreds
of new clients can be on-boarded overnight. This creates
synergies between the clients demand for being
augmented, their ability to monitor effects of the AI-
based resource bundle, and the ventures demand for
externalizing parts of the data management: both
activities call for educational measures and for
harmonization of content and service layer that
integrates well with a broad set of clients.
6. Conclusion
Following Gioia et al. (2013) we derived four
strategic implications for scaling AI ventures in form of
second order themes. Thus, our results provide valuable
information for organizing resources in AI ventures for
scale. Considering the limitations of our study,
opportunities for further research arise: Our findings are
not statistically generalizable, as a qualitative approach
as ours only aims to gain deep insight into
phenomenons (Flick, 2013). AI itself is not one
technology but consists of an ever-evolving frontier
(Berente et al., 2021), such research only grasps a tiny
part of the whole situated in a specific point of this
frontier’s evolution, which naturally poses a limitation
to our research. While we carefully sampled our
experts, future research could use a bigger sample size
and further validate our findings.
12. References
Agrawal, A., Gans, J., & Goldfarb, A. (2018). Prediction
machines: The simple economics of artificial
intelligence. Harvard Business Press.
Anderson, P., & Tushman, M. L. (1990). Technological
discontinuities and dominant designs: A cyclical model
of technological change. Administrative science
quarterly, 604–633.
Berente, N., Gu, B., Recker, J., & Santhanam, R. (2021).
Managing artificial intelligence. MIS Q, 45(3), 1433–
1450.
Bogner, A., Littig, B., & Menz, W. (2014). Interviews mit
experten: Eine praxisorientierte Einführung. Springer-
Verlag.
Breck, E., Cai, S., Nielsen, E., Salib, M., & Sculley, D.
(2017). The ML test score: A rubric for ML production
readiness and technical debt reduction. 2017 IEEE
International Conference on Big Data (Big Data),
1123–1132.
Brynjolfsson, E., & McAfee, A. (2014). The second machine
age: Work, progress, and prosperity in a time of
brilliant technologies. WW Norton & Company.
Brynjolfsson, E., & Mcafee, A. (2017). Artificial intelligence,
for real. Harvard Business Review, 1, 1–31.
Bughin, J., Seong, J., Manyika, J., Chui, M., & Joshi, R.
(2018). Notes from the AI frontier: Modeling the
impact of AI on the world economy. McKinsey Global
Institute.
Burström, T., Parida, V., Lahti, T., & Wincent, J. (2021). AI-
enabled business-model innovation and transformation
in industrial ecosystems: A framework, model and
outline for further research. Journal of Business
Research, 127, 85–95.

Casado, M., & Bornstein, M. (2020). The New Business of AI
(and How It’s Different From Traditional Software).
https://venturebeat. com/2020/02/22/why-ai-ompanies-
dont-always-scale-like-traditional-software-startups/
Chui, M., & Malhotra, S. (2018). AI adoption advances, but
foundational barriers remain. Mckinsey and Company.
Chui, M., Manyika, J., Miremadi, M., Henke, N., Chung, R.,
Nel, P., & Malhotra, S. (2018). Notes from the AI
frontier: Insights from hundreds of use cases. McKinsey
Global Institute.
Cohen,W. M., & Levinthal, D. A. (1990). Absorptive
capacity: A new perspective on learning and innovation.
Administrative Science Quarterly, 128–152.
Corbin, J., & Strauss, A. (2015). Basics of qualitative
research: Techniques and procedures for developing
grounded theory. Sage publications.
Flanagan, J. C. (1954). The critical incident technique.
Psychological bulletin, 51(4), 327.
Flick, U. (2013). The Sage Handbook of Qualitative Data
Analysis. Sage.
Fontana, A. (2021). The AI-first company: How to compete
and win with artificial intelligence. Portfolio/Penguin.
Gioia, D. A., Corley, K. G., & Hamilton, A. L. (2013).
Seeking qualitative rigor in inductive research: Notes
on the Gioia Methodology. Organizational research
methods, 16(1), 15–31.
Giustiziero, G., Kretschmer, T., Somaya, D., & Wu, B.
(2021). Hyperspecialization and hyperscaling: A
resource-based theory of the digital firm. Strategic
Management Journal.
Goldkuhl, G., & Cronholm, S. (2019). Grounded Theory in
Information Systems Research-from Themes in IS
Discourse to Possible Developments. In ICIS.
Henfridsson, O. (2020). Scaling digital enterprises.
Handbook of Digital Innovation. Edward Elgar
Publishing.
Henfridsson, O., Mathiassen, L., & Svahn, F. (2014).
Managing technological change in the digital age: The
role of architectural frames. Journal of Information
Technology, 29(1), 27–43.
Huang, J., Henfridsson, O., & Liu, M. J. (2022). Extending
digital ventures through templating. Information
Systems Research, 33(1), 285–310.
Huang, J., Henfridsson, O., Liu, M. J., & Newell, S. (2017).
Growing on steroids: Rapidly scaling the user base of
digital ventures through digital innovaton. MIS
Quarterly, 41(1).
Iansiti, M., & Lakhani, K. R. (2020). Competing in the age
of AI: Strategy and leadership when algorithms and
networks run the world. Harvard Business Press.
Jöhnk, J., Weißert, M., & Wyrtki, K. (2021). Ready or not,
ai comes—an interview study of organizational ai
readiness factors. Business & Information Systems
Engineering, 63(1), 5–20.
Khan, N., McCarthy, B., & Pradhan, A. (2020). Executive’s
guide to developing AI at scale. https://www.mckinsey
.com/business-functions/quantumblack/our-insights/
executives-guide-to-developing-ai-at-scale#intro
Langley, A., & Meziani, N. (2020). Making interviews
meaningful. The Journal of Applied Behavioral
Science, 56(3), 370–391.
Levinthal, D. A., & Wu, B. (2010). Opportunity costs and
non-scale free capabilities: Profit maximization,
corporate scope, and profit margins. Strategic
Management Journal, 31(7), 780–801.
Linde, L., Sjödin, D., Parida, V., & Gebauer, H. (2020).
Evaluation of digital business model opportunities: A
framework for avoiding digitalization traps. Research-
Technology Management, 64(1), 43–53.
Makridakis, S. (2017). The forthcoming artificial
intelligence (AI) revolution: Its impact on society and
firms. Futures, 90, 46–60.
Metelskaia, I., Ignatyeva, O., Denef, S., & Samsonowa, T.
(2018). A business model template for ai solutions.
Proceedings of the International Conference on
Intelligent Science and Technology, 35–41.
Penrose, E. (2009). The theory of the growth of the firm.
Oxford university press.
Rahwan, I., Cebrian, M., Obradovich, N., Bongard, J.,
Bonnefon, J.-F., Breazeal, C., Crandall, J. W., Christakis,
N. A., Couzin, I. D., Jackson, M. O., et al. (2019).
Machine behaviour. Nature, 568(7753), 477–486.
Raisch, S., & Krakowski, S. (2021). Artificial intelligence and
management: The automation–augmentation paradox.
Academy of Management Review, 46(1), 192–210.
Schilling, M. A. (2002). Technology success and failure in
winner-take-all markets: The impact of learning
orientation, timing, and network externalities. Academy
of Management Journal, 45(2), 387–398.
Schulte-Althoff, M., Fürstenau, D., & Lee, G. M. (2021). A
scaling perspective on ai startups. Proceedings of the
54th Hawaii International Conference on System
Sciences, 6515.
Shapiro, C., Varian, H. R., Carl, S., et al. (1998). Information
rules: A strategic guide to the network economy.
Harvard Business Press.
Sjödin, D., Parida, V., Palmié, M., & Wincent, J. (2021). How
ai capabilities enable business model innovation: Scaling
ai through co-evolutionary processes and feedback loops.
Journal of Business Research, 134, 574–587.
Trajtenberg, M. (2018). AI as the next GPT: A political-
economy perspective (tech. rep.). National Bureau of
Economic Research.
Weber, M., Beutter, M., Weking, J., Böhm, M., & Krcmar,
H. (2021). AI Startup Business Models. Business &
Information Systems Engineering, 1–19.
West, G. (2017). Scale: The universal laws of growth, inno-
vation, sustainability, and the pace of life in organisms,
cities, economies, and companies. Penguin Press.
Wiesche, M., Jurisch, M. C., Yetton, P. W., & Krcmar, H.
(2017). Grounded theory methodology in information
systems research. MIS quarterly, 41(3), 685–701.
Yoo, Y., Boland Jr, R. J., Lyytinen, K., & Majchrzak, A.
(2012). Organizing for innovation in the digitized
world. Organization Science, 23(5), 1398–1408.
Yoo, Y., Henfridsson, O., & Lyytinen, K. (2010). Research
commentary: The new organizing logic of digital
innovation: An agenda for information systems research.
Information Systems Research, 21(4), 724–735.