18% or 78%? Both Numbers Are Real. Only One Describes Your Company.

The Federal Reserve published two AI adoption rates this spring. One says 18%. One says 78%. They cover the same quarter, they come from federal statistical sources, and they appear in the same document, on the same page, in the same table. Neither is wrong. The 60-point spread between them is the most useful number in the AI conversation right now, because it tells you which one is about your company.

The document is a FEDS Note published April 3, 2026 by Jeffrey S. Allen at the Board of Governors. It sets three high-quality surveys side by side and then does the thing almost no vendor deck does: it explains why they disagree.

Here is the table, reproduced honestly.

The Census Bureau's Business Trends and Outlook Survey — biweekly, firm-level — puts AI adoption at about 18% of U.S. firms as of December 2025. The Atlanta Fed's Survey of Business Uncertainty, which targets business executives, puts it at about 78% for November 2025. The Real-Time Population Survey, which asks individuals, puts work-related generative AI use at about 41% of the labor force, with a further 54% figure from the SBU for firms using LLMs specifically.

The denominator test

The Fed's note is unambiguous about the cause. "The most important" source of variation, Allen writes, "relates to the survey goal, unit of analysis, and corresponding sample distribution."

Translated: the 18% counts companies. The 78% counts payroll.

The note publishes the distribution that makes this work, and it is worth reading slowly. 95% of U.S. firms have 1 to 49 employees — but those firms hold only 26.6% of employment. Firms with 250 or more employees are 0.9% of firms and 56.2% of employment. About 57% of all firms have fewer than five people.

Now run the arithmetic in your head. Weight every company equally and the small firms — who adopt least — dominate the average, and you get 18. Weight by headcount and the 0.9% of firms holding a majority of the workforce dominate the average, and you get 78. Adoption did not change between those two sentences. The denominator did.

The Fed states the practical conclusion directly: the BTOS "is the best source for an estimate of the percentage of U.S. businesses that have adopted AI," while the SBU "is a good upper bound on the scope of access to AI tools at work."

One number describes the companies. The other describes the employees. If you have 60 people, you are in the first number and you are being sold the second.

Call it the denominator test, and apply it before any AI statistic reaches a board deck. What is being counted — firms, employees, or individuals? Who is over-represented in the sample? And is the resulting figure a description of adoption, or a ceiling on access? Three questions. They disqualify most of what circulates.

What the 18% is actually measuring

There is a second reason the firm-weighted number runs low, and it makes that number more useful, not less.

The BTOS was designed conservatively. Until November 2025 it asked whether a firm used AI "in producing goods or services" — a production test, not a tool-usage test. The Census Bureau then broadened it to "any of its business functions." Allen notes that even after the revision, "it is possible that BTOS respondents are conditioned to understand the survey questions as referring to more material or extensive AI usage than what is covered in other surveys."

So the 18% is closer to a measure of AI that has reached the operating model. That is precisely the thing a CEO is trying to decide about. Somebody in accounting using a chatbot is not an operating change. AI inside the process that produces the thing you sell is.

The revision itself produced a useful natural experiment. When the question widened, reported adoption jumped by wildly different amounts across industries: 47% (10.6 points) in professional services, and 159% (7.5 points) in manufacturing. Manufacturing had roughly two and a half times as much AI sitting outside production as inside it. Professional services had far less hidden in the gap, because for a consulting firm or an accounting practice the tool and the product are nearly the same thing.

If you run a manufacturer, that ratio is your diagnostic in one line: most of your AI is in the office, not on the floor. We have written about what a fractional CAIO actually does inside a manufacturer, and this is the gap the role exists to close.

Adoption levels are the wrong clock. Planned adoption is the right one.

Here is the number in the note that should change a decision rather than a slide.

The BTOS also asks firms what they expect to be doing in six months. At the start of 2025, about 9% of firms said they planned to use AI within six months. By the end of June, actual adoption had caught up to that expectation. The forecast held. And in the most recent reading, more than 20% of firms expected to be using AI in the first half of 2026.

That is a leading indicator with a track record, and it reframes the question. The competitive gap is not between you and the 78%. The 78% is a payroll-weighted description of Fortune 500 employment. The gap that matters is between you and the fifth of firms — your size, your denominator — who told a federal surveyor they intend to be operating differently within two quarters.

Two decisions fall out of that, and they are the reason the distinction is worth this much space.

First, size the ambition to the denominator you actually live in. Enterprise AI programs are built for organizations where 56.2% of the workforce sits. Importing that shape into a 60-person or 400-person company is how pilots pile up without reaching production — the failure mode we priced out in The AI ROI Gap.

Second, buy judgment at the scale you need it. The adoption data shows professional services and financial firms furthest along at 33% and 30%, which means the expertise is concentrated exactly where billing rates are highest. A company in the 18% cohort rarely needs a full-time executive to cross into the 20% that plans to move. It needs someone who has done it before, for a defined window. That is the entire argument for a fractional Chief AI Officer, and we have published what the role costs and what you are actually buying.

The honest caveat: Allen lists other contributors to the spread beyond weighting — question framing, the materiality of what respondents count as usage, information asymmetry between a survey respondent and the executives who actually know, and social desirability bias, which he notes "may face pressure to report AI usage as an efficiency initiative" among senior leaders. Any reconciliation between these surveys is directional, not exact. The note itself declines to name a single true number.

But the operating lesson does not require precision. It requires knowing which population a statistic describes before you spend against it. Most companies reading that 78% figure are not in it. They are in the 18% — and, more importantly, they are one decision away from the 20% who said they are moving. That decision is not a technology decision. It is a question of who owns it, what phase the company is actually in, and whether the person you bring in has done this at your size rather than at fifty times your size. Which is the same question we ask about every seat, and the same reason we have written about what happens when AI expectations outrun AI governance.

Ten companies understood properly will beat a hundred benchmarked against the wrong denominator.