Sixty-one percent of service firms in the New York Fed's August 2026 panel use AI. Among the firms that do, the median share of workers actually using it is 17% (New York Fed, 2026). For manufacturers the two numbers are 51% and 7%.
One of those numbers goes on the board slide. The other one is your rollout.
Adoption is a company-level checkbox: somebody, somewhere, is using something. Penetration is the share of the workforce whose day now routes through the tool. Almost every 2026 AI plan I read is underwritten by the first number and delivered by the second, and the gap between them is where the business case quietly dies.
The Panel Worth Reading Before Your Next Renewal
The figures come from the Federal Reserve Bank of New York's August 2026 Empire State Manufacturing Survey and Business Leaders Survey โ the third annual wave of AI questions put to firms across New York and northern New Jersey, published 1 September by Jaison R. Abel, Richard Deitz, Natalia Emanuel and Nick Montalbano (New York Fed, 2026).
The adoption curve is the part that gets quoted:
- Service firms using AI: 25% (2024) โ 40% (2025) โ 61% (2026)
- Manufacturers using AI: 16% (2024) โ 26% (2025) โ 51% (2026)
Adoption roughly doubled in two years for services and tripled for manufacturing. On that evidence alone, the rollout is finished.
Two other findings from the same panel say it has barely started.
Penetration. Among adopters, the median service firm has 17% of its workers on AI tools. The median manufacturer has 7%.
Intensity. Three-quarters of service firms and more than 90% of manufacturers describe their AI investment as minimal to modest. About 5% of service firms call it a major strategic investment.
Hold all three at once. Adoption is near-universal, spend is still minimal-to-modest, and roughly one worker in six is on the tools. Those facts only look contradictory if you have been treating adoption as a synonym for deployment.
What 17% Does to Your Business Case
Run it at mid-market scale. A 200-FTE service company sits comfortably inside the adopter column. If the median holds, that is about 34 people whose work touches AI โ not 200.
Now take the standard ROI construction: X hours saved per knowledge worker, multiplied by headcount. Applied across 200 FTEs when 34 are in scope, the participation term alone overstates the return by roughly six times. That is before anyone argues about whether the hours saved are real, and before you subtract the time your team spends correcting output.
The correction is not pessimism. It is a denominator. A business case built on penetration you have measured is a smaller number that survives contact with the finance team; one built on headcount is a larger number that does not.
Cost Is Not Your Constraint โ the Non-Adopters Prove It
The most useful paragraph in the Fed release is the one about firms that have not adopted. Asked why, roughly half said their work does not lend itself to AI. About a quarter said the technology is not good enough yet. More than a third cited data privacy and security. Around a third said they lack technically skilled staff.
Cost was among the least-cited deterrents (New York Fed, 2026).
That finding should end a particular kind of internal conversation. If price were the binding constraint, the fix would be a bigger line item and adoption would be the whole game. It is not. The stated constraints are task fit, output reliability, data governance and internal capability โ four problems a licence does not touch.
Which also explains the intensity number. Firms are not buying minimal-to-modest because they are broke. They are buying minimal-to-modest because they have not yet found the work to point it at.
A Second Instrument Measures AI Penetration, Not Adoption
Firm self-report is a weak instrument, and the Fed panel is entirely self-report. So it matters that a method which does not ask anyone anything lands in the same place.
ActivTrak's Productivity Lab classified AI adoption maturity from behavioural telemetry across 120,620 workers โ observed activity, not licences, logins or surveys. It found 27% of workers at research assistance, 14% at task execution, and 2% at workflow integration, the stage where AI is embedded in how work actually gets done (ActivTrak, 2026).
The instruments disagree on magnitude and agree on shape: access is scaling, integration is not. Worth noting that Gallup published an "organizational AI adoption jumps six points" headline the day before ActivTrak published the 2%. Same week, same economy, two stories โ because one measured distribution and the other measured changed work.
Retraining Is the Only Workforce Lever Actually Moving
Here is where the Fed data breaks the other dominant narrative.
Among AI-using service firms: 4% laid off workers because of AI (up from 1% in 2025), 15% hired fewer, 13% hired more. No manufacturer reported AI-driven layoffs in either 2025 or 2026 (New York Fed, 2026).
Hiring-more roughly cancels hiring-fewer. Layoffs are marginal and concentrated. Whatever AI is doing to these firms, restructuring is not the mechanism.
Retraining is. Just over a third of AI-using service firms retrained workers in response to AI; more than a fifth of manufacturers did. That makes retraining the single largest workforce adjustment in the panel โ and it means fewer than half of adopters are pulling the only lever that moved.
What the Retraining Teaches Is the Problem
The Fed authors characterise the content plainly: it is overwhelmingly about helping employees do their current jobs more effectively rather than preparing them for entirely new roles โ AI literacy, tool instruction, prompt technique, function-specific applications, plus responsible-use training on verifying output and handling data.
Useful. Also the cheaper half of the problem.
Deloitte's State of AI in the Enterprise 2026 โ 3,235 leaders across 24 countries โ found the same lopsidedness at global scale: 53% are educating the workforce to raise AI fluency, but only 30% are reimagining the organisation around new AI patterns and 33% are redesigning career paths (Deloitte, 2026).
And Microsoft's 2026 Work Trend Index, fielded across 20,000 workers in 10 countries, puts a ratio on which half pays. Organisational factors โ culture, manager support, talent practices โ drive more than twice the realised AI impact of individual factors, 67% against 32% (Microsoft, 2026).
Train an individual and you raise their fluency. Two-thirds of the realised impact sits in the work design around them. The Fed panel shows most firms buying fluency, and only a third even buying that.
The Honest Counter
Four limits, stated plainly.
The sample is regional and firm-reported. This is New York and northern New Jersey, answered by firm representatives, not a national probability sample of workers. Sector mix and the region's finance and business-services weighting will shape the adoption figure. Read the panel for direction and structure, not as a national point estimate.
17% may be correct rather than broken. Half of non-adopters said their work does not lend itself to AI โ a defensible position across large stretches of service and operational work. Low penetration is only a failure relative to a business case that assumed otherwise. If yours assumed 100% and you built for 17%, the error is in the assumption.
A median is not your company. The Fed reports the median adopter; the distribution is not published. Some firms in that panel are well above 17% and some are at 2%. The number to act on is your own, which is why the first recommendation below is a measurement, not a decision.
Aggregate calm can hide concentrated harm. The Fed post itself flags work from the Stanford Digital Economy Lab finding that entry-level workers may be affected significantly even where firm-level effects look small. "Only 4% laid anyone off" is a company-level statistic, not a reassurance to a 23-year-old analyst.
What to Decide This Quarter
Four moves. One of them is free.
- Compute your penetration number and print it next to your seat count. Not licences issued โ the share of your workforce whose weekly output demonstrably changed. Two numbers on one line reframes the entire renewal conversation, and you already have the data to produce them.
- Re-underwrite the business case on penetration, not headcount. Replace the headcount multiplier with your measured in-scope population. If the case still clears, you have a real one. If it only cleared at 200 FTEs, you never did.
- Move one workflow rather than one hundred people. ActivTrak's 2% is the stage that produces returns. Pick a single end-to-end workflow, redesign the handoffs, and gate further seat expansion on reaching integration there. Breadth is the cheap axis and it is the one everybody bought.
- Fund the retraining two-thirds of your peers have not โ and score it on redesign, not fluency. Fewer than half of adopters retrain at all, which makes this cheap ground to take. Judge the programme on whether a role or a process changed shape, not on how many people completed a prompting module.
Sixty-one percent adoption tells you the market moved. Seventeen percent penetration tells you nobody has arrived yet.
That gap is the only competitive advantage in this technology that is still available at mid-market prices โ and it closes the quarter someone in your sector stops counting licences and starts counting changed work.
