AI access is expanding faster than many organizations are learning to distinguish access from advantage.

Companies can now acquire capable models through enterprise software, APIs, cloud platforms, open-source systems, and tools already embedded in everyday work. Tasks that once required specialist teams can be attempted by individual employees within minutes.

This expansion matters. But it also creates a strategic illusion. An organization may purchase AI products, run pilots, train employees, automate tasks, and substantially increase usage without developing any capability that competitors cannot reproduce.

The important question is therefore no longer simply:

New patent research provides a useful answer.

What the patent research found

Harvard Business School researchers examined AI-related patents granted to public companies between 2001 and 2023. During that period, the proportion of patents classified as AI-related increased from 9% to 30%. The researchers estimated patent value by examining stock-market responses when patents were granted, treating those reactions as indicators of investors' expectations about future commercial value.

The headline findings were substantial:

The July 2026 HBS Working Knowledge summary reports these figures. The findings do not establish that every AI patent creates these outcomes, nor that companies can reproduce them by purchasing a chatbot or increasing employee usage. The evidence supports a narrower conclusion: AI creates greater firm-specific value when an organization can translate technical progress into an application that fits its knowledge, operations, and market position.

The companies capturing value were not only building AI

The research examined companies in sectors such as finance, manufacturing, energy, and building systems that developed applications using advances originating elsewhere in the AI ecosystem.

Examples documented in the HBS analysis include:

• an AI method used by Chevron to assess petroleum-reservoir production;
• machine learning used by Bank of America to identify potentially unacceptable electronic communications;
• an AI system patented by Johnson Controls to improve equipment-maintenance models.

These organizations were not attempting to create a general-purpose foundation model. They were applying AI to problems defined by their own assets, industry knowledge, operating constraints, and commercial priorities. The examples and research interpretation are available in the HBS Working Knowledge article.

The breakthrough may be public. The return is organizational.

The study's examination of AlexNet makes this especially clear.

AlexNet was a widely disseminated deep-learning breakthrough introduced in 2012. The underlying scientific advance was not privately available to only one downstream company. Yet its economic effect was not distributed uniformly.

In the accessible February 2025 version of the working paper, two model specifications estimated that AI patents granted after AlexNet were 32.3% and 37.5% more valuable in firms with greater AI potential than in less-exposed firms. The authors interpret this as evidence that public technical breakthroughs create more value where organizational activities are already better suited to AI integration.

That finding changes how leaders should interpret AI progress. A better model may increase the technical possibilities available to the entire market. But a technical possibility becomes a business result only after an organization can:

• recognize its relevance;
• connect it to a valuable problem;
• supply the necessary domain knowledge and data;
• redesign the surrounding workflow;
• validate its output;
• assign accountability;
• capture what the organization learns.

The breakthrough expands the opportunity frontier. The surrounding organization determines how much of that opportunity is captured.

The BBGK AI Conversion Capacity Model

The model builds on an established organizational principle: absorptive capacity.

Organizational scholars Wesley Cohen and Daniel Levinthal defined absorptive capacity as a firm's ability to recognize the value of external knowledge, assimilate it, and apply it commercially. They also argued that prior related knowledge affects how effectively an organization can understand and exploit new information.

The patent research brings that principle directly into the AI era. Its accessible working-paper version reports that prior software knowledge, complementary non-AI domain knowledge, and organizational capital help explain which non-technology firms develop valuable AI innovations.

AI Conversion Capacity is BBGK’s operational extension of that logic for AI deployment. It does not replace absorptive capacity; it specifies the additional work of contextualization, governed operationalization, and retained learning.

The model contains four operational capabilities.

1. Absorb

Can the organization identify which external AI advances matter to its actual business?

New models, agents, architectures, and tools appear continuously. Most are technically interesting. Far fewer are strategically relevant to a particular company.

Absorption requires enough technical understanding to evaluate the opportunity and enough domain understanding to recognize where it might matter. This is not the same as following AI news or testing the newest product. It involves determining:

• which capabilities have materially changed;
• which existing problems have become newly solvable;
• whether the required data and infrastructure exist;
• where the organization possesses relevant prior knowledge;
• whether the opportunity is economically meaningful.

An organization with weak absorption capacity reacts to announcements. A strong one translates external progress into a specific opportunity map.

2. Contextualize

Can the organization combine general AI capability with knowledge that is specific to its customers, operations, risks, and standards?

General-purpose models operate across broad patterns. Business value usually depends on narrower context: proprietary data, process history, industry terminology, customer behaviour, regulatory requirements, exceptional cases, internal quality standards, tacit employee knowledge, and economic and operational constraints.

Context is not merely information added to a prompt. It includes the deeper domain logic required to determine whether an output is correct, useful, safe, or strategically appropriate.

This is where AI Conversion Capacity connects with the BBGK Knowledge Distance framework. AI can extend work into adjacent domains, but the quality of the result still depends on access to domain-native judgment and the ability to evaluate what the system produces.

3. Operationalize

Can the organization move the application from demonstration into governed, repeatable work?

A functioning prototype is not yet an organizational capability. Operationalization requires the application to enter a real workflow with a defined problem, an accountable owner, reliable data access, output-validation standards, escalation procedures, human decision boundaries, security and compliance controls, performance measures, and a process for handling failure.

This stage often determines whether an AI initiative remains an impressive demonstration or begins to affect actual outcomes. It also explains why technical performance alone is insufficient. An application can work well in testing but fail in practice because responsibility is unclear, employees distrust it, exceptions are unmanaged, or the workflow surrounding it has not been redesigned.

The BBGK AI Leadership Readiness Stack addresses the governance conditions leaders must establish before AI spreads faster than accountability. The AI Decision Boundary Framework addresses the related question of which work should be automated, assisted, or retained under direct human ownership.

4. Accumulate

Can the organization retain what each deployment teaches it?

AI systems produce more than immediate outputs. Their use can reveal where company data is incomplete, which exceptions matter most, how experts evaluate quality, where customers experience friction, which decisions can be standardized, where human judgment remains indispensable, which workflows should be redesigned, and what the next application should solve.

But this learning compounds only when it is captured. Corrections must improve future instructions, data, evaluation sets, system design, operating procedures, and employee knowledge. Otherwise, the organization keeps paying for the same lessons.

The accessible working paper found that AI patents generated more follow-on citations than comparable non-AI patents, supporting the view that AI innovations can become foundations for subsequent innovation rather than isolated technical events.

Accumulation is therefore the difference between using AI repeatedly and becoming systematically better at using it.

The complete conversion path

This is a diagnostic sequence, not a mathematical formula. It does not imply that every organization must develop patents, proprietary models, or completely unique technology. Nor does it imply that completing the four stages guarantees a competitive moat. It identifies the organizational capabilities required for general AI progress to become increasingly specific, repeatable, measurable, and difficult to imitate.

Why wider access still produces unequal returns

Prior knowledge shapes what an organization can see

Two firms may observe the same technical breakthrough but identify different opportunities. Software knowledge helps assess the technology; domain knowledge helps identify where it can solve a valuable problem. The patent study associates both forms of prior knowledge with future AI innovation. Expertise is therefore not displaced by general-purpose AI; it becomes part of the conversion mechanism.

Workflow knowledge shapes what can be operationalized

A vendor may provide a capable model, but it does not automatically know where decisions stall, which edge cases create disproportionate risk, where poor data enters the process, or who should be accountable when the system fails. Organizations that can surface and structure this distributed operating knowledge have more material with which to build a useful application.

Retained learning shapes whether advantage compounds

A competitor can buy the same model but cannot immediately reproduce an organization’s historical corrections, evaluation data, exception cases, redesigned workflows, or experience of failure. Shared technology can therefore produce differentiated performance when disciplined use creates path-dependent organizational knowledge.

Cost reduction is valuable — but often easier to imitate

Many companies begin with efficiency: faster drafting, lower administrative burden, accelerated software development, or automated customer interactions. These benefits are legitimate, but gains produced mainly by broadly available tools are generally easier to reproduce than gains rooted in proprietary integration, differentiated data, or deeply redesigned operations.

Cost advantage can still be defensible when supported by scale, accumulated process knowledge, distribution, data, capital intensity, or a learning curve competitors cannot quickly match. The HBS interpretation of the patent research similarly argues that cost reduction is only one layer of AI’s value and that larger gains may emerge from products, services, and firm-specific innovation.

Diagnosing AI Conversion Capacity

Leaders should not assess AI maturity primarily through the number of licences purchased, employees trained, prompts submitted, or pilots announced. Those are activity measures. The more useful question is where conversion is failing.

A weak result should not automatically be interpreted as evidence that the model is inadequate. The failure may sit elsewhere in the conversion system.

AI Conversion Capacity and the AI Adaptation Gap

The BBGK AI Adaptation Gap describes the distance between rapidly improving AI capabilities and the slower evolution of human skills, workflows, institutions, governance, and judgment.

AI Conversion Capacity addresses a different question: once an organization begins adapting, can it translate that adaptation into meaningful value? The distinction matters.

Adaptation makes responsible integration possible. Conversion connects that integration to organizational performance. Accumulation determines whether the performance can compound.

What leaders should measure instead of AI activity

1. Problem significance

Is the application addressing a problem material enough to affect customers, risk, revenue, cost, quality, speed, or decision performance? AI does not make a trivial problem strategic.

2. Outcome change

Did the system improve decision quality, error rates, cycle time, customer outcomes, margins, risk exposure, service consistency, or innovation speed? Usage matters only when it contributes to an outcome.

3. Human correction burden

How much time do employees spend checking, repairing, or compensating for the system? Apparent speed can conceal transferred work.

4. Learning retention

Are corrections and exceptional cases improving the system, workflow, and organizational knowledge? Repeated correction without retained learning is operational waste.

5. Reproducibility

Could a competitor achieve substantially the same result by purchasing the same product? If so, the investment may still be necessary, but it is closer to competitive parity than differentiated advantage.

Three categories of AI investment

Parity investments

Common productivity tools, standard automation, and baseline functionality prevent the organization from falling behind. They may be necessary without being differentiating.

Performance investments

These improve an existing process, decision, product, or customer outcome. Their value is measurable, although competitors may reproduce it.

Advantage investments

These combine AI with proprietary knowledge, data, workflow design, accumulated learning, or market position in ways that become progressively harder to imitate.

Not every AI project should create advantage, but leaders should know which category they are funding. Otherwise, ordinary capability upgrades may be mislabeled as transformation.

What the evidence does not prove

Patent innovation is not ordinary AI adoption

The research examines patent-producing public companies. These firms are more likely than typical small businesses or service organizations to possess technical talent, capital, research capability, and established innovation systems. The findings cannot be applied directly to routine chatbot or copilot adoption.

Market-implied value is not realized profit

The reported patent values are inferred from stock-market reactions. They represent investor expectations about future value, not guaranteed revenue or cash flow. The associations with later margins and market share strengthen the commercial interpretation, but they do not establish that every AI innovation will improve performance.

Most of the sample predates widespread generative AI

The HBS patent dataset covers 2001 through 2023. It is therefore stronger evidence for AI-enabled innovation and proprietary applications than for the current wave of widespread general-purpose generative-AI use.

Patents exclude many forms of value

Some valuable AI systems are not patented. Process improvements, internal workflows, service innovations, organizational knowledge, and trade secrets may create value without appearing in patent data.

Firm-level results do not automatically become economy-wide productivity

A separate 2026 analysis from the Federal Reserve Bank of Chicago found that AI patents were associated with later increases in firm capital, revenue, employment, and total factor productivity. However, it also found that these gains were not yet clearly visible at the broader sector level. Its numerical patent-value estimates are based on a different sample and methodology and should not be compared directly with the HBS figures.

The evidence therefore supports meaningful firm-level potential — not universal or automatic transformation.

The better strategic question

Most AI-investment discussions begin with: which tool should we buy? A more disciplined sequence is:

1. What valuable problem do we understand unusually well?
2. Which external AI advance has changed what is technically possible?
3. What proprietary knowledge, data, workflow, or relationship can improve the application?
4. What must be redesigned for the system to influence real work safely?
5. How will corrections and usage improve the next version?
6. What will become more difficult for competitors to reproduce over time?

This shifts AI strategy away from technology accumulation. The objective becomes building an organization capable of repeatedly converting external progress into internal performance.

BBGK conclusion: the emerging divide

Most serious organizations will use AI. The more consequential distinction will be among those that merely consume common tools, those that improve existing operations, and those that convert external advances into accumulated, firm-specific capability.

AI’s diffusion does not make expertise, workflow design, governance, or learning less important. It makes them more decisive. Advantage lies not in completing one AI project, but in becoming consistently better at recognizing the next opportunity, operationalizing it responsibly, and retaining what the deployment teaches.

That is AI Conversion Capacity.

Frequently asked questions

Evidence note

The July 2026 HBS article summarizes a revised February 2026 working paper and reports a controlled AI-patent premium of 9.6%. A publicly accessible February 2025 conference version reports an earlier estimate and different sample totals. This article uses the later HBS summary for current headline figures and the accessible earlier paper only for detailed methodological and mechanism analysis.