You have probably seen both stories. One says nearly everyone is using AI at work. The other says barely any companies have adopted it. They run in the same week, sometimes on the same page, and they cannot both be the whole truth.
A new piece from the Federal Reserve Bank of St. Louis explains why they disagree. The short version: the question is doing the work, not the technology. Researchers found that how a survey asks about AI, not just whether companies use it, can move the measured adoption rate by nearly half. Same survey program. Same general period. Different wording. Very different answer.
For a leader trying to figure out where their own organization actually stands, that is the useful part. The number you get back depends almost entirely on what you ask.
Two surveys, two very different worlds
Start with the size of the disconnect.
Surveys that ask workers directly put AI use on the job at roughly 35% to 40%. People say they are using it. Then you look at the main U.S. survey of businesses, the Census Bureau's Business Trends and Outlook Survey, and the picture flips. It put AI adoption among firms at only about 5% to 7%.
Read those two numbers next to each other. Workers reported using AI at roughly five to eight times the rate captured by the narrow firm survey. Either a lot of people are imagining their AI use, or the survey was missing something.
The St. Louis Fed post, summarizing the authors' working paper on AI adoption in Europe and the U.S., makes the case that it was the survey.
A large part of the difference came down to one phrase
Here is where it gets practical for anyone who measures things for a living.
The Business Trends and Outlook Survey asked companies a single, general question: in the last two weeks, did the business use AI in producing goods or services? One question. One narrow target. From September 2023 through October 2025, that is how the U.S. measured firm-level AI adoption.
A European survey, the EU-ICT-Firm survey, took a different path. Instead of one broad question, it asks whether a company uses any of eight specific AI technologies. Things like machine learning, speech recognition, natural language generation, image recognition. Name the tools, and people recognize what they are doing.
That design choice has a measurable effect. Among European firms, the share using AI for any business purpose came out about five times larger than the share using AI specifically to produce goods and services. The narrow question was not wrong. It was just narrow. It counted the factory floor and skipped the rest of the building.
The natural experiment
The cleanest proof came when the Census Bureau changed the wording.
In November 2025, the Business Trends and Outlook Survey changed its AI question. It went from asking about "producing goods or services" to asking about "any of its business functions." Broader net. Same survey program, same general period.
The measured adoption rate almost doubled. It moved from about 10% to about 17%.
After the Census Bureau broadened its survey question in November 2025, measured AI adoption almost doubled, rising from about 10% to about 17% in the same period.
Source: Federal Reserve Bank of St. Louis, 2026The timing and size of the jump strongly suggest a measurement break, not a sudden overnight rollout of AI tools. The broader wording uncovered roughly seven points of AI use that the old question had been walking right past.
Separate BTOS AI supplements, designed to ask about more technologies and business functions, also produced higher adoption numbers than the single general question. The pattern held every time the net got wider.
Why the disconnect exists in the first place
There is a quiet assumption buried in the old question. Asking whether you use AI "in producing goods or services" treats AI as a production tool, like a machine on an assembly line.
But that is not where most of it lives.
Think about where AI actually shows up in a company. Marketing drafts. Customer support replies. Finance forecasts. Resume screening in hiring. Scheduling and reporting in operations. None of that is "producing goods or services" in the literal sense the question intended. So a manager filling out the survey reads the question honestly, thinks about the product, and answers no. Meanwhile, many workers may already be using AI in parts of the business the question never named.
The adoption was real. The survey just did not ask about the rooms where it was happening.
What this means for measuring your own company
This is the part worth taking with you, even if you never read another economics paper.
If you want to know how much AI your organization actually uses, the question you ask will shape the answer more than reality will. Ask only whether AI is being used to produce your product, and you will get a small, tidy number that feels almost reassuring. Ask instead whether AI shows up in any of your business functions, and you may find the real figure is closer to double that.
A leader who manages off the narrow number is managing off roughly half the truth. You would underinvest in training people who are already using the tools. You would miss the support rep quietly running every ticket through an assistant, or the analyst whose forecasts now start with a model. You cannot govern, secure, or improve what your own measurement refuses to see.
So the better instrument is simple. Measure by function, not by a single yes or no. Walk through marketing, support, finance, hiring, and operations one at a time and ask what is already in use. The honest count is almost always bigger than the headline count, and it is hiding in the functions you did not think to check.
That change, from asking "do we use AI" to asking "where, and how well," is also the move from counting tools to building real capability. It is the difference between a company that knows it has AI somewhere and one that can see exactly where its people sit and where they could go next. The 7 Levels of AI Proficiency exists to make that second view possible: a way to look past the single yes-or-no and see the actual range of skill across a team, function by function. The same lesson the researchers found in a government survey applies inside any company. The richer the question, the truer the answer.
Related reading: Level 6: Admiral.
Sources
- Measuring AI Adoption among Firms: How You Ask Matters (Federal Reserve Bank of St. Louis, On the Economy, June 2026)
- AI Use at U.S. Businesses (U.S. Census Bureau, May 2026, documents the BTOS AI question wording change)
- Mind the Gap: AI Adoption in Europe and the U.S. (Bick, Blandin, Deming, Fuchs-Schündeln, and Jessen, NBER Working Paper w34995)
- Mind the Gap: AI Adoption in Europe and the US (Brookings Papers on Economic Activity, Spring 2026 Conference Draft)
Frequently Asked Questions
Did AI adoption actually jump in late 2025, or just the measurement?
Just the measurement. The Business Trends and Outlook Survey changed its question in November 2025 from "producing goods or services" to "any of its business functions," and the reported rate moved from about 10% to about 17% in the same period. The underlying use did not change in that window. The question did.
Why do worker surveys show so much more AI use than firm surveys?
Workers get asked directly about their own daily tools, so they report the real thing. The main U.S. firm survey asked one narrow question tied to production, which skipped AI used in marketing, support, finance, hiring, and other functions. Worker surveys came in near 35% to 40%; the narrow firm question came in near 5% to 7%.
How should I measure AI use inside my own organization?
Ask by business function, not with a single general question. The European approach, naming specific technologies and specific functions, consistently surfaced more real adoption than one broad yes-or-no. In Europe, AI use for any purpose ran about five times the production-only figure. Expect the same shape in your own count.
Find your AI Proficiency level
The free 7 Levels assessment places you across seven stages of AI capability. Under ten minutes. Research-backed scoring.