Overview

Fast AI diffusion does not determine market structure by itself. This chapter synthesizes the series through three steps: access to an AI service, conversion into a reliable process through skills, data, workflows, and investment, and capture through prices, quality, markups, market share, imitation, ownership, or switching costs. It reconciles the early evidence: U.S. adoption is concentrated among larger firms and often shallow; task augmentation and capital change outpace net employment change; French and Portuguese user data through 2022 show no systematic markup-growth increase for mainly non-generative use; GenAI exposure, patent concentration, and start-up acquisitions identify possible routes to stronger concentration. These studies are descriptive or correlational and do not cover every AI layer, including chips, cloud, and frontier models. The calibrated conclusion is conditional: diffusion can increase contestability when access, skills, imitation, and independent entry broaden together, or increase concentration when durable capabilities and critical ownership remain concentrated.

Learning outcomes

AI access, conversion, and capture are distinct: availability does not show that a firm can build or retain a durable advantage.
The early evidence points in two directions: adoption and innovation are dynamic, but capabilities, inputs, and independent rivalry can remain concentrated.
The key test is whether functional integration, imitation, new entry, and independent growth keep pace with ownership and complementary assets.

Keywords

AICompetitionDiffusionCapabilitiesContestabilityConcentration

What Fast AI Diffusion Will Decide About Markets

Fast AI diffusion can still produce concentrated markets when access to tools spreads faster than the capabilities needed to use them, the inputs needed to improve them, and the independent firms able to challenge their owners. This series has examined that possibility through the organization of adoption, the distribution of skills and integration, the effects on work and capital, direct market indicators, and the lifecycle of AI start-ups. The final judgment is conditional: the evidence does not show a general AI-driven monopoly, but it does show mechanisms through which widely available technology can reinforce existing advantage.

That conclusion is more useful than choosing between “AI will democratize business” and “AI will inevitably concentrate it.” It tells us what to look for. If access leads to broad functional integration, imitation, new entry, and independent growth, diffusion can increase contestability. If the returns to integration accumulate among firms with superior data, talent, capital, distribution, or ownership of critical technology, diffusion can increase concentration even while more firms report using AI.

Access spreads faster than capability

The central distinction is between access and conversion. Access means that a firm can obtain an AI service, model, or application. Conversion means that the firm can turn that access into a reliable process: it can identify a useful task, connect the tool to appropriate data, train people, verify outputs, redesign workflows, and sustain the cost of learning. Capture is a third step. It asks whether the resulting advantage reaches customers through lower prices and better products, is retained as a higher markup or larger share, is copied by rivals, or is protected by ownership and switching costs.

This three-step model is a synthesis of the evidence, not a new statistical index. It explains why a low price or broad availability at the access stage does not settle the competition question. A general-purpose model can be available to a small company and a large incumbent at the same time. The incumbent may still convert it more effectively because it has structured data, specialized staff, established distribution, and managers who can finance experimentation. The smaller firm may still win a niche if its focused product needs fewer complements or if the model lowers an entry barrier that previously mattered more. The outcome depends on which constraint is binding in the relevant market.

The organizational evidence makes the conversion step visible. Firm-level adoption, functional deployment, and worker-task use are different observations. Formal adoption may show that an enterprise has authorized or purchased AI. Functional breadth shows whether the system has entered several parts of the business. Worker-task use shows whether employees are applying it in particular activities. These layers can align, but they can also separate. A background system can operate without direct worker interaction, while workers can experiment before a practice becomes formal. Adoption is therefore not a switch that moves an entire enterprise from “without AI” to “with AI.”

The economic meaning of those layers also differs. Worker-level assistance can improve output while leaving the firm's organizational design almost unchanged. Broader functional use may signal that management is coordinating the technology across revenue, operating, and administrative processes. Operational investment can signal an even deeper change in which data, software, equipment, and responsibilities are being reconfigured. A firm that has only one useful application may gain a local advantage; a firm that has built a cross-functional capability may accumulate learning that rivals cannot reproduce as quickly.

The distribution of adoption shows why access and conversion can diverge. In the U.S. supplement, 18% of firms reported firm-level AI use, but those firms represented 32% of employment when weighted by employment. Worker use in work-related tasks was reported by 23% of firms and represented 41% of employment. These are not interchangeable measures: they use different questions and reference periods, and worker use is reported by firms rather than directly by employees. Their divergence nevertheless shows that adoption is concentrated among enterprises that employ more people.

Sector and size reinforce the same point. Current firm-weighted adoption was highest in Information, Professional, Scientific, and Technical Services, Education, Finance and Insurance, and Real Estate. The Information sector's employment-weighted use reached 54% and was expected to approach 67% within six months. Firms with 250 or more employees reported 31% current use, compared with about 18% among firms with fewer than 20 employees. A minority of enterprises can therefore expose a large share of employment to AI without AI being common across the business population.

That concentration has two interpretations. It may reflect a productive race in which large or specialized firms adopt first because they have strong use cases and can spread fixed investments across many operations. It may also create a first-mover advantage.

Published by DeepMox

FAQ

What is the difference between AI access, conversion, and capture?

Access is obtaining an AI tool; conversion makes it a reliable process; capture determines who receives or controls the resulting advantage.

When can AI diffusion increase market concentration?

It can do so when skills, data, capital, distribution, critical technology, or learning remain concentrated and rivals cannot imitate effectively.

What evidence would show that AI is increasing competition?

Evidence would include broader entry, imitation, independent growth, or a wider distribution of innovation, not merely accessible tools or high adoption.

Previous12 minThe Entrant Lifecycle: From AI Start-Up to Acquisition

All chaptersBack to Topic