What the AI noise is hiding. One brief, every Tuesday. We cut the hype and show you the numbers that actually matter, named for what they are.
Nobody was screened at the door, and 22.4% said yes. Tuesday, September 1, 2026.
Current issue · Issue 11 · September 22 · Issue 10 · September 15 · Issue 9 · September 8 · Issue 7 · August 26 · Issue 6 · August 19 · Issue 5 · August 11 · Issue 4 · August 5 · Issue 3 · July 28 · Issue 2 · July 21 · Issue 1 · July 14
Most AI adoption numbers are measured on people who already use AI. This week's strongest one was not: the US Census asked every employer business in the country, screened nobody, and 22.4% said yes — while 9.4% could not say either way, which is a governance reading wearing a survey artifact's clothes.
Read the first and the third together. 22.4% of businesses said yes, and 9.4% could not say either way. Nearly one business in ten has no line of sight into its own tool use, on a two-week recall window.
Issue 7 built its headline on this: “80% of respondents report that AI has improved their individual productivity.”
The sentence the survey actually supports, as carried by The Register's same-day report, is “80 percent of respondents who use AI in their roles.” The 80% is computed on AI users. The 37% is computed on all respondents.
Issue 7's headline set a screened subgroup against a full-sample figure and called the difference a finding. The gap between feeling faster and finding it in the accounts is still there. The arithmetic was wrong. Corrected in place on all four live surfaces on 28 August.
The rule this cost us, and it governs every figure below: a marker number must survive its own qualifier. If the qualifier will not fit, choose a different number. An audit of 103 claims across our own Issues 1 to 7 found the same defect every time — the body carried the qualifier and the headline dropped it, because the qualifier made the line long.
In the fortnight of 27 July to 9 August 2026, 22.4% of all US employer businesses used AI in any business function. 68.2% did not. 9.4% answered “do not know.”
Census asked everybody. Nail salons are in this denominator. So are two-truck plumbing firms. That is why Census says 22.4% while McKinsey's respondents, who use AI in their roles, say 80% of something else: the two instruments are not asking the same question of the same people.
Cite it as BTOS, reference period 27 July to 9 August 2026, release CB26-TPS.48. Never as the bare URL. That file is replaced on 10 September with the next vintage, and the URL will keep resolving to different numbers. Issue 7's correction was about exactly that failure mode.
Twenty consecutive fortnights are published. The series runs 17.3% to 22.4%, about 5.1 points in nine months, roughly 0.27 points per fortnight. Nothing in this data will support anyone describing it as a step change.
And the 9.4% is not noise. It has fallen from 11.1% last November, so it is being resolved slowly. But a firm that answers “do not know” is a firm with no line of sight into its own tool use. On a two-week recall window, that is a governance reading dressed as a survey artifact.
Salesforce, State of Agentic AI in the Enterprise, 27 August 2026. Vendor-sponsored research; Salesforce sells Agentforce. A double-blind survey of 2,025 agentic AI decision-makers across 20 countries, fielded 14 to 28 May 2026, all of whom influence agent purchasing. Respondents split into deployed (30%), piloting (47%) and evaluating (23%), and most findings reflect the 30% who deployed.
With that stated: lighter oversight reached positive ROI in 7.2 months against 9.3 for heavier governance, and below-average governance was nearly twice as likely to discover an agent operating outside its parameters only after a consequential error, 32% against 18%.
Both halves are self-reported by the buyers and neither is audited. It does not say agents reach ROI in eight months generally, and it does not say light governance causes faster returns. The report itself says being first to deploy does not cause faster returns.
The St. Louis Fed reads the share of US work hours saved rising from 1.6% to 2.2%, with hours assisted at 4.1% to 6.3%, on a nationally representative survey with no screen on the adoption question.
Issue 7 printed the ECB's 3.8%. Those two levels cannot be subtracted, and this brief will not subtract them. The ECB figure is whole-economy: it already divides by every worker, including the 51.2% who save no time at all, and its questionnaire allowed reports of up to fifteen hours. The Fed's asks employed respondents and caps reported savings at four hours. The cap difference alone can produce most of the distance.
Nothing here says the United States is behind Europe, and no reader should take it that way. What compares is direction, and both point the same way: up, and slowly. What both agree on is the shape — assisted hours run at roughly three times saved hours.
A reported McKinsey “94 percent have yet to create meaningful value.” If real, the invention-by-subtraction Issue 7 identified has been adopted by its own alleged author. That would be the best story here. It cannot be verified from this machine — mckinsey.com returns an Akamai 403 to every method tried, the Wayback capture has no usable text, and no non-McKinsey source quotes it.
GTIA's adoption study: members-only, no sample size, field dates or screen obtainable. Any BTOS figure from a third-party reprint, including a “21.5% as of 2 August” in circulation: BTOS numbers come from the data file or not at all.
“ECB, 56% adoption.” Two Spanish outlets printed 56%; the ECB text says 52%. Separately, two French outlets misnamed the authors. Both errors are circulating now.
Morgan Stanley's 8.2% to 12.3% productivity gains. A C-suite screen, ten industry groups chosen as those most likely to benefit, and a range splicing 935 interviews from October 2025 with 808 from April 2026 into one figure for “the past 12 months.” The report is honest about all of it; the range still cannot be quoted without four qualifiers.
No study published between 22 August and 1 September measured agentic AI in production from telemetry or logs. Every agent number in circulation this week is a human being's opinion of their own system, collected by someone selling into that system.
Nothing in window from NBER, the OECD, the IMF, the BIS, the Bank of England, Stanford HAI, MIT or the major lab economics teams. Searched and not found, which is not proof of absence.
U.S. Census Bureau, release CB26-TPS.48. Downloaded and parsed. Every lead figure, standard error and collection date read from National.xlsx and Sector.xlsx rather than the press summary.
Salesforce, 27 August 2026. Opened; methodology quoted verbatim from the page. Vendor-sponsored, and the outcome figures come from the 30% already in production.
Federal Reserve Bank of St. Louis, FRED Blog, 27 August 2026. Opened. Population, screen and both hour caps read from the post. The blog says of its own instrument that “the measurements are inherently approximate.”
European Central Bank, ECB Blog, 26 August 2026. Opened and verified against the live page: title, date and all three author names confirmed.
McKinsey & Company, 25 August 2026. NOT OPENED. mckinsey.com refuses connections from the machine this brief is written on. The 80% qualifier is confirmed in the best fetchable secondary, and this issue's own correction depends on it.
From this issue, every figure's source is named and linked. Issues 1 to 6 are deliberately not being retro-linked: a link added today to a 2025 citation would point at whatever now lives at that address.
Want a number on your own coordination tax? Measure it right now with the free Agentic Readiness Score, built on Align-ify™.