From the firm- to the application-level: A new typology of AI use in organisations

arXiv:2611.15371 · econ.GN, q-fin.EC · Submitted 2026-11-11 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: We're looking at the paper: "From the firm- to the application-level".

Warren: Here's the gist of the paper. The paper addresses a limitation in firm-level literature on artificial intelligence by arguing that a single framework is insufficient to describe its variety, as a firm may simultaneously hold numerous AI applications, each possessing unique characteristics regarding novelty, sourcing mode, objective, and technology type.

Ruth: The core contribution is the proposal to shift the unit of analysis from the firm to the individual AI application.

Warren: This new approach develops a comprehensive typology by combining four key conceptual dimensions: novelty of the application (new only to the firm or new to its market), sourcing mode (in-house development or external acquisition), type of AI technology (analytical, perceptual, or generative), and objectives of AI use (efficiency, market, or search).

Ruth: This process involves a four-step construction: crossing novelty with sourcing mode; then crossing the resulting types with objectives; subsequently crossing those types with technologies and their required assets; and finally synthesizing these into a combined typology encompassing thirty-six distinct AI application types.

Paper summary: Ruth: So, we're talking about a paper called "From the firm- to the application-level: A new typology of AI use in organisations." This work tackles a problem where looking at AI adoption just by the whole company doesn't capture how messy it really is.

Warren: Right. The core idea here is that firms don't just adopt one single AI thing; they can have dozens of different applications running at once, and those things are all fundamentally different—different in how new they are, where they came from, what goal they serve, and what kind of technology they actually use.

Iris: Exactly. The paper argues that the unit of analysis needs to move away from the whole firm and focus on each individual AI application because those four dimensions you mentioned—novelty, sourcing mode, objective, and technology type—are defined by the application itself rather than just what's happening inside the firm >

Ruth: That's right. The authors propose creating a typology that combines these four specific things to map out all the possibilities for AI use within an organization. They build this by looking at novelty combined with sourcing mode first, then crossing those types with objectives, and finally linking those types to the actual technologies and assets required >

Warren: And this framework ends up generating thirty-six distinct types of AI applications that you wouldn't see if you just used a simple adoption or innovation label >

Iris: It’s important because it moves past the old idea that we're just looking at whether something is adopted or innovative, and instead shows these intermediate categories like co-invention and internalisation >

Ruth: So, co-invention means they build new things using technology that the firm itself hasn't developed yet, which is different from internalisation where they build in-house AI that someone else already made >

Warren: And the paper also shows how adoption and innovation are just endpoints on a spectrum, highlighting these transitional types between them >

Iris: It also clarifies what kind of innovation an application is actually leading to, like whether it's a product or process innovation, and it shows how the cost of adoption scales with complexity >

Ruth: I was particularly interested in how the paper describes how organizational adjustment increases with novelty, and that the assets needed change depending on whether you're building something in-house versus buying something externally >

Warren: And they point out that if you shift the objective from just efficiency to market use, or from market use to search, the difficulty of proving a payoff goes up too >

Iris: Empirical evidence supports this; they used three original data sources: the World Economic Forum two thousand twenty-five MINDS cohort, the United States federal AI use case inventory, and the United Kingdom Algorithmic Transparency Recording Standard > <ref:2611.15371#pg1>

Ruth: What those initial results show is that efficiency objectives are pretty common and analytical technology is what you see most often in practice >

Warren: And internalisation actually accounts for a significant chunk of those applications they found >

Iris: So, if we take all this together, the paper suggests that firm-level measures shouldn't be simple adoption or innovation metrics; instead, they should be seen as compositions—averages taken over these diverse portfolios of AI applications >

Ruth: The final implication is that what you measure about a country’s AI success depends on the makeup of its applications rather than just how many things it has adopted overall >

Warren: It suggests that policy should focus on shifting the composition towards co-invention, because that seems to be where the real complexity and potential lie >

Iris: So, to wrap up, this paper is about giving us a better way to categorize AI use by looking at the application level instead of just the firm level.

Ruth: And it does that by creating this detailed typology of thirty-six distinct types based on four key dimensions >

Conclusion: Ruth: So we've looked at how this paper shifts our focus from just looking at an entire company to looking at every single piece of AI software they use.

Warren: Yeah, this work by authors puts a lot of effort into creating this new way to categorize what firms are actually doing with AI.

Iris: The core idea is moving away from simple adoption or innovation labels and building this detailed structure based on four specific dimensions: novelty, sourcing mode, technology type, and objective.

Ruth: It’s not just a new way to label things though; it creates this whole framework that yields thirty-six different types of AI applications.

Warren: Thirty-six types sounds like a lot to pin down. What does that actually mean for a company running twenty different AI tools?

Iris: It means you can't just treat them as the same thing, because one might be an in-house analytical tool for efficiency, while another is an external generative tool for market search.

Ruth: And these types reveal some things about the transition between just adopting something and actually innovating something new.

Warren: I’m hearing they call it intermediate categories like co-invention, where they build new applications using technology that someone else already made, which is different from internalisation.

Iris: Exactly. Internalisation is when a firm builds in-house AI technology that others have created themselves, which has a totally different cost structure than buying something off the shelf.

Ruth: And this framework also helps map out what kind of actual innovation an application might be driving, like if it's leading to a product change or just improving the internal process.

Warren: It also shows how the cost of adoption scales with complexity; for example, organizational adjustment gets bigger when the novelty is higher.

Iris: And you get a different set of assets needed depending on whether you are building that technology internally or acquiring it externally.

Ruth: The paper uses real data from things like the World Economic Forum cohort and federal AI use cases to show these patterns aren't just theoretical.

Warren: What those results actually show is that efficiency objectives are pretty common, and analytical technology shows up a lot in the real world.

Iris: And internalisation ends up being a significant part of what firms are doing with these applications.

Ruth: So the authors conclude that firm-level measures should be seen as combinations of these diverse applications rather than just simple adoption counts.

Warren: It suggests that the actual outcome for a country’s AI success depends on the mix of these different application types, not just how many tools they have in total.

Iris: The final implication is that policy should focus on shifting this composition toward co-invention to get the most valuable results.

Fulvio Castellacci

University of Oslo

econ.GN, q-fin.EC

Submitted: 2026-11-11

Updated: 2026-11-11

Code: https://github.com/ombegov/2025-Federal-Agency-AIUse-Case-Inventory

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 92/100

The gist: The paper addresses a limitation in firm-level literature on artificial intelligence by arguing that a single framework is insufficient to describe its variety…

Key concepts

Novelty of the Application
This measures whether an AI application is new only to a specific firm or if it is novel to the entire market. It helps categorize applications based on how unique their existence is relative to the existing landscape, influencing adoption costs.
Sourcing Mode
This dimension distinguishes how an organization acquires its AI capabilities: whether they develop the technology internally (in-house) or acquire it from an external source. This affects the required assets and development complexity.
Type of AI Technology
AI technologies are categorized into analytical, perceptual, or generative types. These classifications help describe the fundamental nature of the AI being used, which dictates its potential applications and required resources.
Objectives of AI Use
This refers to the goal behind using an AI application: whether it is primarily aimed at achieving efficiency gains, capturing market share, or performing information search. Different objectives influence how adoption costs escalate with complexity.

Terminology

Summary

The paper addresses a limitation in firm-level literature on artificial intelligence by arguing that a single framework is insufficient to describe its variety, as a firm may simultaneously hold numerous AI applications, each possessing unique characteristics regarding novelty, sourcing mode, objective, and technology type. p. 0

Where the paper says it

This paper argues that the difficulty follows from the unit of analysis: a firm may hold many AI applications at once, each with its own novelty, sourcing mode, objective and technology type. These dimensions are defined at the level of the application rather than of the firm.

The core contribution is the proposal to shift the unit of analysis from the firm to the individual AI application. p. 3
Where the paper says it

The implication is that the firm-level literature finds it hard to conceptually define and empirically measure the complex variety of AI applications. The present paper argues therefore that current research on AI adoption and innovation in organisations should shift its unit of analysis from the firm to the individual AI application.

This new approach develops a comprehensive typology by combining four key conceptual dimensions: novelty of the application (new only to the firm or new to its market), sourcing mode (in-house development or external acquisition), type of AI technology (analytical, perceptual, or generative), and objectives of AI use (efficiency, market, or search). p. 4
Where the paper says it

5 assets are larger for in-house development than external purchase; and the difficulty of showing that an application has paid off is higher as the objective moves from efficiency to market and from market to search. Hence, it is reasonable to expect that AI applications do not spread evenly across all types of the new typology, but they will rather concentrate in specific directions, each with different expected economic impacts.

This process involves a four-step construction: crossing novelty with sourcing mode; then crossing the resulting types with objectives; subsequently crossing those types with technologies and their required assets; and finally synthesizing these into a combined typology encompassing thirty-six distinct AI application types. p. 5
Where the paper says it

Section 3 reviews the recent firm-level literature on AI, pointing out what this literature does and does not distinguish. Section 4 develops the new typology in four steps: novelty crossed with sourcing mode; the resulting types crossed with objectives; the same types crossed with technologies and the assets they require; and the combined typology that the three previous steps together define.

This typology moves beyond the traditional binary distinction between adoption and innovation, revealing intermediate categories such as co-invention—where new applications are built using technology not developed by the firm—and internalisation—where a firm builds in-house AI technology that others have already created. p. 3

Where the paper says it

One is co-invention, in which new AI applications are built from technology the firm did not create (Bresnahan and Greenstein, 1996; Bresnahan, 2024; Bresnahan et al. The other is internalisation, in which a firm builds in-house AI technology that others have already developed.

The framework demonstrates that adoption and innovation are endpoints of a spectrum, highlighting these crucial transitional types. p. 3
Where the paper says it

First, adoption and innovation are endpoints of a space, not a binary dichotomy as often assumed in the literature. In particular, the typology uncovers two important types of AI applications that are neither pure adoption nor innovation.

Furthermore, the typology elucidates what specific type of innovation (e. p. 8
Where the paper says it

The fourth dimension that it is important to consider is the type of technology that we are looking at. A large literature in innovation studies has convincingly shown that opportunities, constraints and results of innovative activities are typically technology-specific, and it is therefore crucial to consider the characteristics of a given technological trajectory when studying its patterns, determinants and effects.

g., product, process, organizational, or marketing) an application may lead to. p. 3
Where the paper says it

Second, the typology shows what type of innovation (e. product, process, organisational or marketing) a given AI application may lead to.

It also establishes how adoption cost escalates with complexity—organizational adjustment increases with novelty; required complementary assets are larger for in-house development versus external purchase; and the difficulty in demonstrating payoff rises as objectives shift from efficiency to market use, and from market use to search. p. 4
Where the paper says it

5 assets are larger for in-house development than external purchase; and the difficulty of showing that an application has paid off is higher as the objective moves from efficiency to market and from market to search. Hence, it is reasonable to expect that AI applications do not spread evenly across all types of the new typology, but they will rather concentrate in specific directions, each with different expected economic impacts.

Empirical illustration utilizes three original data sources: the World Economic Forum 2025 MINDS cohort, the United States federal AI use case inventory, and the United Kingdom Algorithmic Transparency Recording Standard. p. 4

Where the paper says it

Hence, it is reasonable to expect that AI applications do not spread evenly across all types of the new typology, but they will rather concentrate in specific directions, each with different expected economic impacts. To examine this aspect, we provide an empirical illustration of the typology using three large sources of documented AI applications: the United States Federal AI use case inventory, a near-census of 3,611 reported applications; the United Kingdom Algorithmic Transparency Recording Standard, a register of records published by public bodies; and the World Economic Forum’s 2025 MINDS cohort, which documents a large number of applications in private firms.

These analyses reveal that efficiency objectives are prevalent, analytical technology is frequent, and internalisation accounts for a significant portion of applications. p. 7
Where the paper says it

The third dimension focuses on the objectives that a firm has when it decides to adopt or develop a new technology. We distinguish three distinct types of objectives: (1) efficiency: the new technology is meant to improve efficiency of the production process (e.

The paper concludes that firm-level measures should be viewed as compositions—averages taken over portfolios of heterogeneous applications—rather than simple adoption or innovation metrics. p. 4
Where the paper says it

Taken together, these results point to one general contribution. Firm-level measures of AI should best be seen as compositions: they are averages taken over portfolios of heterogeneous applications, and the main puzzles in this literature (the wide dispersion of adoption and productivity estimates, and the persistence of the adoption-innovation binary) are consistent with this aggregation.

The final implication is that the aggregate outcome depends on the composition of a country’s AI applications rather than just its level of adoption, suggesting policy should focus on shifting composition towards co-invention. p. 3
Where the paper says it

4 measures of AI use should be built up from application-level information rather than recorded directly. The implication is that the firm-level literature finds it hard to conceptually define and empirically measure the complex variety of AI applications.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Co-invention improvements can be achieved by building new AI applications built from technology that did not create by applying substantial inventive effort to adapt general-purpose technologies to specific organizational circumstances, such as turning a standard technology into a novel use for which there is no precedent.

  2. Internalisation improvements involve in-house building of AI that could have been bought instead, specifically developing in-house AI technology that others have already developed, often done when data cannot leave the organisation due to sensitivity or regulatory concerns.

  3. Search objective improvements focus on increasing the firm’s capacity to search the technological space by using AI used to develop further AI, such as employing model-assisted engineering or synthetic data generation in R&D projects.

  4. Analytical AI improvements involve leveraging ICT infrastructure (pipelines, storage, data quality) and specialised human capital to build systems that predict or classify from structured data, fitting them into existing decision processes.

  5. Perceptual AI improvements require integrating physical capital assets and the engineering human capital to interpret imagery or sensor streams through applications like computer vision in a working operation.

  6. Generative AI improvements can involve using a document archive with commodity language models to create services, such as building a case-law summarisation service for a specific regulatory area.

  7. Market objectives can be addressed by assembling AI into a new customer-facing offering that leads to product or service innovation, such as creating an automated claims triage service using commercial language models.

  8. AI Innovation in the Market objective involves training your own foundation model on vessel sensor time-series to build a predictive maintenance service, resulting in AI-native products and ventures.

Abstract

Artificial intelligence is applied in organisations in many different ways. However, the firm-level literature on AI lacks a conceptual framework able to describe that variety. Empirical studies typically examine one dimension of AI use at a time. This paper argues that the difficulty follows from the unit of analysis: a firm may hold many AI applications at once, each with its own novelty, sourcing mode, objective and technology type. These dimensions are defined at the level of the application rather than of the firm. We therefore propose shifting the unit of analysis from the firm to the individual AI application, and develop a typology combining these four conceptual dimensions. The new framework locates any application in a common space, and it points out in particular two types that the prevailing adoption-innovation distinction neglects: co-invention, novel uses of technology the firm did not develop, and internalisation, in-house building of AI that could have been bought instead. Documented use cases illustrate the framework, which is then used to discuss implications for firms strategies and the diffusion of AI in economic systems.

Sources

Related papers