The AI economy keeps producing facts that seem to contradict each other.
ChatGPT has reached a huge global user base. U.S. workers tell survey researchers that AI has entered their jobs. Claude conversations show real people using models for writing, coding, analysis, and practical guidance. Census data says firms are using AI across business functions. At the same time, measured productivity effects remain hard to see, labor-market signals are still noisy, and many executives say AI has not yet changed employment or productivity inside their own firms.
It is tempting to resolve this by choosing a side. Either AI is transforming everything and the official statistics are lagging, or the whole thing is hype and the usage numbers are a mirage.
That is the wrong move.
The better answer is that AI adoption is not one object. It is a ladder of different observables: technical exposure, product use, worker self-report, firm deployment, national readiness, social pressure, and measured economic effect. Each layer is real. Each layer has a different denominator. Each layer can move at a different speed.
That is why the current evidence can be both hot and cold. Individual contact with AI is broad. Organizational conversion is thinner. Aggregate economic effects are still hard to pin down. The central question is the gap between those facts: how access to AI becomes routine use, how routine use becomes workflow redesign, and how redesigned workflows become capability, income, productivity, or new forms of scarcity.
Call this the conversion gap.
It matters because almost every argument about AI and the economy crosses it. Claims about productivity, inequality, labor-market change, AGI, development, or relational work all depend on whether AI remains a tool people touch or becomes a capability people, firms, and countries can reliably build around.
Exposure, platform traces, worker surveys, firm data, readiness indices, social dynamics, and labor-market effects each observe a different layer of the transition.
1. Exposure tells us where AI could matter
The broadest adoption numbers are not really adoption numbers at all. They are exposure measures.
Task-based exposure work begins with a map of jobs and tasks, usually an occupational taxonomy such as O*NET. Researchers ask whether a model could perform, assist, or speed up those tasks. The influential early example is Eloundou, Manning, Mishkin, and Rock’s “GPTs are GPTs”, which estimated how much of the U.S. labor market is exposed to large language models. Related IMF work on generative AI and the future of work adds a second distinction: some exposed jobs may face substitution, while others may become more productive because AI complements human judgement.
Exposure measures are useful because they move before adoption is directly visible. They let us compare occupations, sectors, countries, and income groups while the technology is still diffusing. For policy, that is a real advantage. If advanced economies have more high-exposure cognitive work, and low-income economies lack the infrastructure to benefit from AI, then the distributional question appears before the macro data settles.
But exposure is a possibility measure. It tells us where AI might matter, not where AI has already changed work. It can overstate impact when a task is technically feasible but legally constrained, socially resisted, badly integrated, or economically unprofitable. It can understate impact when new workflows emerge that old task databases do not describe.
Exposure is the map of possible contact. It is not the record of use. 1
That difference is the first clue. If exposure is much larger than observed adoption, the interesting question is not whether one measure is “right.” It is what has to happen between a technically exposed task and a changed job.
2. Platform traces get closer to behavior, but not to the whole economy
The next layer is more concrete: what people actually do inside AI products.
Anthropic’s Economic Index uses privacy-preserving analysis of Claude conversations and API traffic, maps observed tasks to O*NET, and separates some uses into augmentation and automation patterns. Its March 2026 report sampled one million Claude.ai conversations and one million first-party API transcripts. Coding remains central, but Claude.ai use has become less concentrated: the top ten O*NET tasks fell from 24 percent of traffic in November 2025 to 19 percent in February 2026. Anthropic also reports that about 49 percent of jobs have seen at least a quarter of their tasks performed using Claude, while global per-capita usage remains concentrated in a small set of countries.
OpenAI’s consumer-facing evidence points in a related but broader direction. The NBER working paper How People Use ChatGPT, written by OpenAI researchers and David Deming, studies ChatGPT use through July 2025. It estimates that ChatGPT had reached around 10 percent of the world’s adult population. It also finds that non-work usage grew to more than 70 percent of all use, and that practical guidance, information seeking, and writing account for nearly 80 percent of conversations. Work use is more concentrated among educated users in higher-paid professional occupations.
Their advantage is proximity to behavior. They show what users ask models to do: write, code, summarize, search for guidance, analyze, and troubleshoot. They can reveal changes that surveys miss, such as coding moving from chat interfaces into APIs or personal use growing faster than workplace use. 2
The trap is that platform traces feel more complete than they are. Claude data is not the AI economy. ChatGPT data is not the AI economy. A prompt is not a completed task. An API call may be part of a production workflow, or it may be a prototype nobody relies on. A conversation can be useful, ignored, checked, rejected, copied, or quietly forgotten.
Platform data gives us revealed contact with AI. It does not give us a denominator for all possible users, all work, or all economic output.
So the conversion gap remains. We can see people trying the tools. We cannot yet see, from product traces alone, how much work has actually been reorganized.
3. Worker surveys show breadth before depth
Worker surveys add the denominator that platform traces lack. They ask a population sample who uses AI, how often, and sometimes for what kind of work.
Bick, Blandin, and Deming’s “The Rapid Adoption of Generative AI” is the key worker-side reference. Using the August 2024 Real-Time Population Survey, the authors find rapid diffusion among U.S. adults age 18 to 64: 39.4 percent had used generative AI, 28 percent of employed respondents had used it for their job, 24.2 percent had used it at least once in the previous week, and 10.6 percent used it every workday.
Those are fast adoption numbers. But the same paper contains the more important sentence for this article: only 1 to 5 percent of all work hours were currently assisted by generative AI, with reported time savings equivalent to about 1.4 percent of total work hours. 3
That is the conversion gap in miniature. Many workers have touched AI. Much less work has been rebuilt around it.
Gallup’s Q1 2026 workplace survey gives the same pattern with fresher frequency data. Half of employed U.S. adults reported using AI in their role at least a few times a year; 28 percent used it a few times a week or more; 13 percent used it daily. Yet only about one in ten employees in AI-adopting organizations strongly agreed that AI had transformed how work gets done in their organization. 4
This is why worker surveys should be read as breadth measures before they are read as transformation measures. They can tell us which workers use AI by age, education, occupation, sector, and work arrangement. They can capture informal use that firms may not see. But self-report has familiar weaknesses: vague definitions, recall bias, social desirability, and the difference between “I used AI” and “AI changed the way my job works.”
The useful conclusion is not that worker surveys are unreliable. It is that they answer a specific question. They tell us how widely AI has entered people’s routines. They are less decisive about depth, quality, governance, and organizational consequence.
4. Firm data is where the wording becomes the result
If worker data shows breadth before depth, firm data shows how hard depth is to measure.
The U.S. Census Bureau’s Business Trends and Outlook Survey is one of the strongest real-time sources. It is also a warning label for anyone who wants a single AI-adoption rate. The early BTOS question asked whether firms used AI “in producing goods or services.” Using September 2023 to February 2024 data, the NBER paper Tracking Firm Use of AI in Real Time found that biweekly estimates of business AI use rose from 3.7 percent to 5.4 percent, with expected use around 6.6 percent by early fall 2024. 5
That was a deliberately hard bar: AI in production.
In November 2025, Census revised the core wording toward AI use “in any business function.” Under that broader concept, BTOS data from December 14, 2025 to May 3, 2026 showed overall AI use hovering between 17 and 20 percent of U.S. businesses, with 20 to 23 percent expecting to use AI in the next six months. The accompanying Census/NBER microstructure paper reports that during the November 2025 to January 2026 supplement period, 18 percent of firms used AI in at least one function, or 32 percent on an employment-weighted basis. Among adopting firms, scope was still limited: 57 percent used AI in three or fewer business functions, and AI-related employment decreases appeared in only 2 percent of firms. 6
That shift is not a technical footnote. It is the result.
Ask about AI in producing goods or services and adoption looks low. Ask about AI in any business function and adoption looks much higher. Weight by employment and the number changes again. Ask executives whether their firms “actively use AI” and still another picture appears.
The BFI/NBER working paper Firm Data on AI surveys nearly 6,000 senior executives in the U.S., U.K., Germany, and Australia. It reports high stated firm use, but limited intensity: executives who use AI average only 1.5 hours a week, and over 80 percent of firms report no impact on either employment or productivity over the last three years. Expectations are more optimistic than realized effects. 7
This is not an embarrassment for firm surveys. It is the point of firm surveys. They expose the difference between experimentation, embedded vendor features, pilots, business-function use, production workflows, and measurable organizational change.
The Federal Reserve’s 2026 comparison of U.S. AI adoption surveys makes the same point directly: workers, executives, and firms are asked different questions about production use, business functions, organizational integration, and frequency of use. 8
If one employee uses ChatGPT for a memo, is the firm an AI adopter? If a vendor product quietly adds an AI feature, has the firm adopted AI or merely bought software? If a firm runs a pilot, is that production? There is no neutral answer. The question wording decides what kind of adoption we are measuring.
Selected adoption signals point in different directions because they use different denominators: people reached, workers using AI, hours assisted, firms reporting use, and reported impact.
People have clearly touched the tools.
The denominator shifts from contact to work actually reorganized.
Reported use is ahead of realized productivity and job effects.
5. The macro evidence is still stubborn
The counterargument to all this measurement caution is simple: maybe the data is just early. General-purpose technologies often take time to show up in productivity statistics. Firms need to reorganize. Workers need to learn. Managers need to redesign processes. Complementary investments lag.
The lag argument is the strongest objection here, and the best reason not to wave away thin early numbers. But it cuts both ways. The same logic that lets optimists say “give it time” lets skeptics say “then show me the reorganization.” A lag you cannot yet observe is a forecast, not evidence. So the current data earns a close reading, not a side.
The strongest current conclusion is not “AI changes everything” or “AI changes nothing.” It is narrower: AI use is real, exposure is large, individual contact is broad, but aggregate labor-market and productivity effects are still modest, noisy, or hard to detect.
Anthropic’s Labor Market Impacts of AI paper is useful because it tries to bridge possibility and behavior. It combines O*NET tasks, theoretical exposure estimates, and real Claude usage into an “observed exposure” measure. Even there, the authors find no systematic increase in unemployment for highly exposed workers since late 2022, though they find suggestive evidence that hiring into exposed occupations has slowed for younger workers. 9
Apollo’s “zero evidence of AI-related job losses” frame is useful as market commentary, but it should not carry the publication-grade claim by itself. Yale Budget Lab’s 2026 review is a better cautionary anchor: exposure metrics are not job-risk metrics, and the labor-market evidence remains mixed, noisy, and statistically weak in many places. 10
The bigger AGI economics debate belongs one layer above this evidence. Acemoglu’s near-term macro frame, Autor’s complementarity argument, and Korinek/Suh’s AGI transition scenarios are asking different questions from current adoption datasets. They are theories of what could happen if capabilities keep improving and complementary institutions, firms, and workers respond. They are not direct measurements of current AI use. 11
This distinction is worth preserving. If the current data does not yet show a productivity boom, that does not prove AI will be small. If platform use is growing quickly, that does not prove transformation has arrived. The responsible position is less satisfying and more useful: the conversion process is underway, but uneven and only partly observed.
6. Readiness is the part no model ships with
The conversion gap also exists at country level.
The IMF’s AI Preparedness Index compares economies across digital infrastructure, human capital and labor-market policies, innovation and economic integration, and regulation. It does not count chatbot users. It asks which countries are structurally positioned to absorb AI, manage transition costs, and convert exposure into broad gains.
Microsoft’s Global AI Adoption in 2025 report measures a different object: direct diffusion, using aggregated and anonymized telemetry adjusted for device access, internet penetration, OS and device market share, and population. It estimates that global generative AI user share reached 16.3 percent in the second half of 2025, up from 15.1 percent in the first half. It also finds a divide in working-age population use: 24.7 percent in the Global North versus 14.1 percent in the Global South. 12
These two measures should not be fused into one country ranking. Microsoft is closer to usage. IMF is closer to capacity. One asks where people appear to be using generative AI products. The other asks where societies have the complementary systems to benefit.
That distinction matters because workplace use is only one layer. AI diffusion also depends on infrastructure, skills, regulation, language coverage, trust, and firm capability. The same model can produce very different economic effects depending on broadband access, education systems, cloud availability, managerial quality, data governance, and institutional capacity.
The weakness of country indices is aggregation. National averages hide regional, sectoral, linguistic, and class divides. Composite indices depend on weighting choices. But for macro analysis they provide a frame that product logs and worker surveys cannot. They remind us that access to a model is not the same as the capacity to use it well.
7. Adoption can move before proof
One reason adoption data can run ahead of productivity data is that people do not adopt only after clean evidence arrives.
Adoption is also social. It is shaped by peer behavior, fear of falling behind, managerial pressure, client expectations, professional status, and the awkward feeling that everyone else has secretly become more efficient.
In Social Dynamics of AI Adoption, Leonardo Bursztyn, Alex Imas, Rafael Jimenez-Duran, Aaron Leonard, and Christopher Roth study parental demand for unrestricted AI tools in education. The setting is not the workplace, which matters. But the mechanism travels: demand rises strongly when parents believe more peers are using the technology. Information about possible harms changes beliefs and increases support for bans, but does not necessarily reduce individual demand. The fear of being left behind can sustain adoption even under uncertainty. 13
This helps explain the strange emotional temperature of AI adoption. The evidence is mixed, but the social pressure can be intense. Workers experiment because colleagues do. Firms buy tools because clients ask. Managers adopt because not adopting looks negligent. Students use AI because they think everyone else is using it.
That does not make adoption irrational. It means adoption can be rational before productivity is proven. In a competitive setting, waiting for perfect evidence can itself feel risky.
This is the bridge to the next labs essay. Once the measurement problem is clear, the economic question shifts. If AI makes some cognitive outputs cheaper, where does value move? What remains scarce? That is where Alex Imas’s relational-sector argument belongs: human provenance, care, trust, judgement, and relationship may become more central precisely where other outputs become abundant. 14
8. What we can say now
The most useful summary is not a verdict. It is a reading guide.
AI exposure is large. That tells us where models could matter.
Platform usage is real and growing. That tells us what users are trying inside particular products.
Worker adoption is broad. That tells us AI has entered everyday routines.
Firm deployment is narrower and definitional. That tells us organizational conversion is harder than individual contact.
Country readiness is uneven. That tells us model access is only one part of capacity.
Labor-market and productivity effects are still hard to measure. That tells us not to confuse use with transformation.
The mistake is to force these layers into one curve. They are different instruments pointed at different parts of the transition.
So the next generation of AI adoption measurement should measure conversion, not just contact. It needs to ask whether AI is used once or every day; whether it supports a peripheral task or a core workflow; whether output quality improves; whether human verification remains central; whether firms redesign data models and operating procedures around agents; whether teams maintain the systems they build; and whether gains accrue to workers, firms, customers, experts, novices, or platform owners.
The first-order question was whether people would use AI. They do.
The harder question is whether people, firms, and countries can build the layer above the model: data structures, workflows, governance, maintenance routines, human judgement, and institutions that let AI work reliably.
That is the conversion gap. It is also the bridge from measurement to economics. If AI makes some tasks cheaper, the next question is not only what gets automated. It is what remains scarce: trust, judgement, accountability, taste, care, relationship, institutional competence, and the ability to turn model output into reliable systems.
The job of measurement is to see that difference before the slogan does.
Further reading:
- Anthropic Economic Index
- Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations
- How People Use ChatGPT
- The Rapid Adoption of Generative AI
- Tracking Firm Use of AI in Real Time
- AI Use at U.S. Businesses
- The Microstructure of AI Diffusion
- Firm Data on AI
- Monitoring AI Adoption in the US Economy
- Gen-AI: Artificial Intelligence and the Future of Work
- Global AI Adoption in 2025
- Social Dynamics of AI Adoption
- Labor Market Impacts of AI: A New Measure and Early Evidence
- What We Do and Don’t Know About How AI is Affecting the Labor Market
- The Simple Macroeconomics of AI
- Applying AI to Rebuild Middle Class Jobs
- Scenarios for the Transition to AGI
- What will be scarce?