October 2026 gave us more official counts of business AI use than any month I can remember, and the counts disagree with each other. On October 8 the Federal Reserve Bank of New York said 46% of small employer firms use AI. On October 29 the UK government said 32% of UK businesses do. The US Census Bureau's own running survey has sat close to 20% all year. Then on October 19 Gartner reported that only 13% of CIOs say their company has gotten significant or transformational value from AI. So depending on who you ask, adoption is a fifth of the economy, a third, or nearly half, and value is a rounding error on top of it. These are careful institutions using decent samples, so sloppy work doesn't explain the spread. The spread comes from what they chose to count. Adoption, as everyone measures it, has no unit.
That should worry any executive whose board deck has an "AI adoption" slide on it. Most of those slides repeat the survey mistake inside one company. They count logins, seats, prompts, copilots turned on. They count presence. The AI adoption metrics that matter count something else, and the October numbers show what that something is.
Four surveys, four definitions
Start with the Census Bureau, because it is the most careful and because it changed its own ruler halfway through. Its Business Trends and Outlook Survey used to ask firms whether they used AI "in producing goods or services." In November 2025 the Bureau changed the question to ask about AI use "in any business function," which, as its own analysts write, now includes simple tasks such as drafting emails. Under the new wording, overall use ran between 17% and 20% from December 2025 to May 2026, and the national rate was 19.8% as of May 3, 2026. That comes from a May 26 Census story by Adam Grundy, Cory Breaux and Dhanapati Khatiwoda. Firms with 250 or more employees reported 37%. Firms with four or fewer reported under 20%.
Look at what the redefinition did. A company that writes one email a week with ChatGPT now counts the same as a company that runs underwriting through a model. Both are "adopters." The Bureau had good reasons to widen the question, since the narrow one was missing real use. Widening it still meant the headline number started measuring the bottom of the distribution instead of the middle.
The New York Fed figure is higher because it asks a different population a different thing. Will Aarons and Asani Sarkar, writing on Liberty Street Economics on October 8, used the 2025 Small Business Credit Survey, which ran from September to November 2025 and covered 5,248 employer firms in its AI module. 46% of firms with at least one employee reported AI use. Another 15% planned to adopt within twelve months. On paper, that is a near-majority of American small business.
Look a little further down the same post, though. 77% of firms reported no change in labor costs from AI. Only 31% reported increased sales from it. So close to half the firms use it, and roughly three quarters of them see nothing in the one cost line AI is supposed to hit hardest.
The UK survey, published October 29, is the most honest of the four about this gap, partly because it says so in plain English. 32% of UK businesses used at least one AI technology in 2026, up from 16% in 2025. That doubling will be the headline in every trade paper. The same executive summary says 24% of adopters saw operating costs fall, 64% saw no immediate financial effect, and 9% saw costs rise. Only 14% considered AI of high or critical importance to their operations. The authors' own summary sentence is the one I would pin to the wall: "AI delivered measurable but largely incremental improvements."
Then Gartner, which asked the people who write the checks. Its 2027 CIO and Technology Executive Agenda survey covered 2,002 CIOs and technology executives, fielded May through July 2026, and was released at Gartner IT Symposium/Xpo in Orlando on October 19. 13% said their enterprise had realized significant or transformational value.
Line the four up and you get a funnel. Somewhere between a fifth and a half of firms are "using AI." Around a quarter of UK adopters see a cost line move. About an eighth of big-company CIOs see something they would call significant. Each survey measures a different layer of the same funnel and reports it as if it were the whole thing.
Presence is the cheapest thing to measure
There is a reason adoption surveys count presence. Presence is binary, it is cheap to ask about, and respondents can answer it without looking anything up. "Do you use AI in any business function?" takes a second to answer. "What share of your accounts payable volume no longer touches a human?" takes a finance team a week.
Companies inherit the same bias for the same reason. The easiest dashboard to build is the one your vendor already ships: active users, weekly prompts, seats assigned. Microsoft, Google and OpenAI all give enterprise admins usage reports, because usage is what they bill for and what they need you to renew. A usage report tells you how much of the product you consumed. It tells you nothing about how much of the business changed.
I used to think the fix was better value measurement. Build an ROI model per use case, track hours saved, convert hours to dollars. Most companies that try this end up with a spreadsheet of self-reported time savings that adds up to more hours than the staff actually work. The UK survey points at the same issue in its own way: its benefit figures are what businesses said about themselves, not outcomes anyone measured independently. When 71% of adopters report a productivity benefit and 64% report no immediate financial effect, both numbers can be true, because people feel faster without anything on the ledger moving.
So the fix has to come from somewhere other than better self-reporting. It comes from the Gartner number almost nobody led with.
The number hiding under the 13%
The 13% got the headlines. The more useful figures sat lower in the release, as reported by eWEEK on October 22. Only 31% of respondents have rules for stopping unsuccessful AI pilots. Only 17% have criteria for abandoning low-value AI initiatives that are already in production. Organizations with pilot-stopping rules were 2.3 times more likely to land in Gartner's high-performing "vanguard" group, and those with production exit criteria were 2.6 times more likely.
eWEEK is careful to say this is an association, not proof of cause, and that's right. Still, it is the most informative pair of numbers published about enterprise AI this year, because it measures something presence metrics can't see: whether a company can tell its AI projects apart.
A company with no exit criteria has one category of AI project, the kind that exists. A company with exit criteria has at least two, the kind that stays and the kind that goes, and to sort them it had to decide in advance what "working" means, in a unit someone could check. Exit criteria force a definition of value. The definition is the asset. The rule for killing things is just the receipt that the definition exists.
Alicia Mullery, a VP analyst at Gartner, said of these criteria that "Establishing and adhering to these criteria is a team sport." The phrase sounds soft. The content is hard. A pilot only gets stopped when the business owner who asked for it agrees, ahead of time, to a number that would embarrass them. Most organizations never write that number down, because writing it down makes failure legible.
AI adoption metrics that matter
If presence is the wrong unit, what is the right one? Here is the scorecard I now push clients toward. It has five measures, and none of them is a usage count.
The retirement rate
Count how many AI initiatives you shut down in the last two quarters, divided by how many you started. A retirement rate of zero means one of two things. Either every bet you placed paid off, which nobody believes, or you have no mechanism for telling winners from losers. Gartner's 17% figure suggests the second case is the common one at the production layer. For illustration: say a company launched twelve AI pilots in the first half of 2026 and stopped none. Its adoption slide shows twelve green boxes. Its real state is twelve unknowns, plus a growing bill.
Work removed, not work assisted
Assisted work is the email the model drafted that a person then edited. Removed work is the invoice match no person ever opened. The NY Fed's 77% no-labor-cost-change figure tells you most small firms live almost entirely in the assisted column. Assisted work is fine. It makes people feel better, and some of that feeling turns into speed. It almost never shows up as a reduction anywhere, because the human stays in the loop and the loop keeps its old shape. Count the steps in a process that a person no longer touches at all. That count is small at nearly every company I see, and it is the one that predicts whether costs move.
The cost line that moved
Pick the specific general-ledger line you expect each AI initiative to bend, and name it before launch. Claims processing labor. Outsourced translation. Tier-one support contractor spend. If you can't name a line, the initiative is a capability bet, which can be fine, but it has to be labeled as one and funded out of a different bucket. The UK survey's split (24% saw costs fall, 64% saw no immediate effect, 9% saw costs rise) is roughly what you get when a whole economy adopts without naming the line in advance. Some firms hit by accident, most hit nothing, and a few pay more.
Time to the first kill decision
Measure how many weeks pass between a pilot's launch and the first meeting where someone is allowed to end it. Many companies have no such meeting at all. Pilots die by neglect, which is slow and expensive, or they turn into production by default, which is worse. Suppose your answer is "never scheduled." Then the pilot-stopping rules Gartner asked about don't exist at your company, whatever the governance policy says.
Dependency depth
Count how many of your processes would stop, rather than slow down, if your main model provider went dark for a day. Usage dashboards treat this number as success. It is also exposure. A company where AI is truly embedded should have a nonzero answer and a plan for it. A company that answers zero after two years of "adoption" has learned something more useful than any usage report could tell it: the AI is decoration on top of processes that never changed.
Why the denominators keep moving
There is a deeper reason adoption numbers drift, and it explains why a company-level presence metric will keep flattering you.
Every time a vendor ships AI into a product you already own, your adoption rate rises without anyone deciding anything. When Microsoft puts Copilot into Outlook, or Google puts Gemini into Workspace, or Salesforce turns on an agent inside a tool your sales team already uses, an employee who clicks "summarize this thread" becomes an AI user. The Census Bureau's widened question will pick that up. So will your internal dashboard. The number goes up because the software around you changed, not because your operating model did.
That is why the US national rate can climb steadily while the NY Fed finds labor costs flat for most firms. Adoption is drifting upward with the defaults. It will keep doing so whether you have a strategy or not. A metric that rises on its own is a weather report, not a performance measure.
The UK numbers show the same pattern from a different angle. Adoption doubled from 16% to 32% in a year, while 65% of non-adopters said they had no clear business case or proven return on investment. So the businesses that haven't adopted mostly name the value problem as the reason, and the businesses that have adopted mostly report value that's incremental. Doubling adoption didn't clear up the value question for anybody. It just moved more firms to the other side of it.
What the vanguard actually does differently
Gartner's vanguard group is small, and I won't pretend to know everything about it from a press summary. The correlation with exit rules is still worth taking seriously, because it lines up with what I see in practice.
The companies getting real value out of AI behave less like adopters and more like investors. They hold a portfolio of bets with stated theses. They review the theses on a schedule. They sell the losers. Their AI roadmap looks like a list of positions, each with an entry price, a target and a stop. Few of them talk about "AI adoption" at all. They talk about specific processes, such as month-end close or warranty claims or inbound triage, and what fraction of each now runs without a person.
The others behave like collectors. They acquire tools, distribute licenses, run enablement sessions and report rising usage. Every new capability gets added, nothing gets removed, and the AI line item grows by the quarter. Gartner's survey found 58% of CIOs under greater pressure to deliver cost improvements through AI, and 51% expect AI to raise the total cost of ownership of their technology. Both findings fit a collector's books: more pressure to save money, and a bill that keeps rising because nothing ever leaves.
Collectors aren't foolish. They are measuring exactly what their dashboards make easy to measure, and the dashboards were built by people who profit from consumption.
The board slide I would replace
Say your current board slide says "78% of employees are active weekly AI users." Here is roughly what I would put in its place.
The number of AI initiatives in production, and next to it, the number retired this half. The specific cost lines each initiative was meant to bend, with the before and after, in dollars, from the ledger and not from a survey of the people involved. The share of a few named processes that now complete without a human touch. The date of the next kill review, and who owns it.
This slide is uglier. Some quarters it will show retirements and flat lines. It will make at least one senior sponsor uncomfortable. It will also be the first AI slide your board can actually govern with, because every number on it is something someone decided rather than something that drifted in with a software update.
A second benefit took me a while to notice. When a company commits to the retirement rate, its vendors' behavior changes. A vendor whose product shows up as a line that has to bend on a named ledger account negotiates differently from a vendor whose product shows up as weekly active users. You stop paying for presence. You start paying for removal.
The trap in the survey season
Survey season breeds a particular anxiety. An executive reads that UK adoption doubled, or that 46% of US small firms are on board, and concludes that the company is behind and needs to catch up on usage. That reaction makes the problem worse, because it pushes the company to buy more presence just as the evidence says presence is the wrong thing to buy.
Read the October data again with the funnel in mind. The firms that pull ahead will have defined, in advance, what an AI project has to do to survive, and built the habit of ending the ones that don't. The firms with the highest adoption rate won't necessarily be those firms. Gartner's survey suggests fewer than a third of large enterprises have even the first half of that habit, and fewer than a fifth have the second half. At the moment, that gap is the most underpriced advantage in enterprise AI.
Build the scorecard before you buy the next tool
None of this comes in a box. No vendor will sell you a retirement rate, because a retirement rate is a commitment to buy less from somebody. No usage dashboard will tell you which ledger line should move, because the dashboard doesn't know your business. The scorecard has to be designed: which processes count, what "removed" means for each, which cost lines are in scope, who has authority to end a project and on what schedule, and how the numbers get pulled from systems of record instead of from the opinions of the people whose projects are being judged.
That design work is the strategy. It decides what your organization will learn from every AI dollar it spends over the next three years. Get it wrong, and you'll join the 87% of Gartner's respondents who spent heavily and couldn't point to significant value. Get it right, and every pilot you start comes with the means of judging it, which is the only condition under which spending more on AI makes sense.
Agor AI Advisory builds these scorecards with leadership teams. We start from your ledger and your processes rather than from a vendor's usage report. We help you write the exit criteria before the pilot launches, wire the measures into systems your finance team already trusts, and run the first kill reviews alongside you so the habit sticks once we leave. The October surveys made it obvious that adoption, by itself, measures nothing worth paying for, and your board will work that out soon enough. Be the executive who already has the better number.
Sources
- Gartner Survey Finds Only 13% of CIOs Say Their Enterprise Has Realized Significant or Transformational Value from AI, Gartner, October 19, 2026
- Only 13% of CIOs See Major AI Value, Gartner Finds, eWEEK, October 22, 2026
- AI adoption survey report 2026: executive summary, GOV.UK, October 29, 2026
- AI Adoption and Employment Expectations: Evidence from a Survey of Small Business Owners, Liberty Street Economics, October 8, 2026
- Large Firms With at Least 20 Employees Biggest AI Users, U.S. Census Bureau, May 26, 2026