AI DevelopmentEvidence guide5 min readPublished September 21, 2026

4 evidence classes · 8 live figures · 3 questions · every number with its method beside it

Why AI Adoption Numbers Disagree So Much, and Which to Use

20% of US businesses use AI; nearly nine in ten survey respondents say theirs does. Both are real. Four evidence classes explain the gap, with eight numbers.

DA
Digital Applied Team
Research and practical guidance
Editorial dateSeptember 21, 2026
Sources readSeptember 22, 2026

The US Census Bureau says 19.8% of American businesses used AI in the two weeks to May 3, 2026. McKinsey says nearly nine in ten respondents report their organisation uses AI regularly in at least one business function. Both figures are real, dated and published by their authors, and someone will put each of them on a slide this week as "AI adoption".

They are not measuring the same thing, and the difference is not that one is wrong. It is who was asked, what exactly was asked, how many answered, and whether the method is published at all. This post sorts adoption and return-on-investment figures into four evidence classes, tables eight live numbers with their methods beside them, and ends with the three questions to ask before any of them goes in a deck. It analyses other people's evidence; it adds no survey of our own.

Key takeaways
  1. 01
    Four classes: counted telemetry, vendor-reported programme data, self-reported survey, analyst forecast.Each answers a different question. Telemetry counts what a platform's own users did. Programme data is a vendor describing its own results. A survey is a sample's answers to a question. A forecast is a model of the future. None of them is 'the adoption rate'.
  2. 02
    The 20% and the nine-in-ten differ by population, question and weighting, not by accuracy.Census asks a 1.2-million-business sample whether it used AI in any business function in the past two weeks. McKinsey asks 1,719 mostly large-organisation respondents whether their organisation uses AI regularly in at least one function, weighted by national GDP. Different denominators, different verbs.
  3. 03
    A figure with no published method is a claim, not a measurement.Of the eight figures tabled, the analyst forecast's method is gated, one vendor figure was spoken with no method, and one survey states its sample and dates but not its question wording. The rest publish enough to be checked.
  4. 04
    Quote the class, the population and the date, or do not quote the number.'19.8% of US businesses, Census survey, two weeks to May 3, 2026' is citable. 'Nearly 20% AI adoption' is not, because the next slide will have nine in ten from a different class and the audience will conclude one of you is lying.

01The taxonomyFour classes of evidence

Every adoption or ROI figure in circulation belongs to one of four classes. The one-line test for each is a question about who did the measuring and what they could see.

Class 1
Counted telemetry
Test: did a system count it?

A platform counting its own usage: conversations classified, sessions logged, tokens billed. Precise about its own users and silent about everyone else. Anthropic's Economic Index is the model of the class: a classifier over sampled Claude conversations, with the population stated as Claude users.

Strong on what, weak on who
Class 2
Vendor-reported programme data
Test: is the vendor describing itself?

A company reporting a figure about its own operations or customers: share of code AI-generated, tasks an agent leads. Each figure has its own definition, chosen by the reporter. Comparable only when the definitions are published, which they usually are not.

Each one its own unit
Class 3
Self-reported survey
Test: did a sample answer a question?

A sample of people or businesses answering a question. Value depends entirely on the sample, the question wording, the weighting and the response rate. Runs from official statistics with 1.2 million businesses to an investor-commissioned panel of 5,000 adults, and the class label alone does not tell you which.

Read the questionnaire
Class 4
Analyst forecast
Test: is it about the future?

A model of what will happen, published by a research firm and usually gated. A forecast is not an adoption figure at all; it is a projection with assumptions you cannot read. Useful for a sense of scale, unusable as evidence of what is happening now.

Not a measurement

02The evidenceEight live figures, sorted

Each row gives the figure as its author states it, the class it belongs to, who measured it and how, and whether the method is published. Where we could not find a sample size or a question wording, the row says so rather than filling it in.

Sources as named in each row, read on September 22, 2026. The Census and Anthropic rows are linked in the text; consultancy, vendor and analyst sources are cited by name and date.
Figure, as statedClassWho measured it, population, question, n, dateMethod published?
19.8% of US businesses used AI in the past two weeks (May 3, 2026)Self-reported survey, official statisticsUS Census Bureau Business Trends and Outlook Survey; all US employer businesses excluding farms. Sample of about 1.2 million businesses in six panels, each asked every 12 weeks; collected every two weeks. Question, as revised in November 2025: whether the business used AI 'in any business function' in the past two weeks. Story dated May 26, 2026.Yes, in full
37% of firms with 250 or more employees reported using AISelf-reported survey, official statisticsSame survey, over the December 2025 to May 2026 window the story covers rather than a single collection date. The national rate for firms with four or fewer employees is under 20%. The gap by size is the point.Yes, in full
Nearly nine in ten respondents report regular use of AI in at least one business functionSelf-reported survey, consultancyMcKinsey Global Survey, The state of AI in 2026, published August 25, 2026. Online, in the field May 4 to June 8, 2026; 1,719 participants in 97 nations; weighted by each nation's share of global GDP. 36% of respondents work at organisations over $1 billion in revenue.Sample, dates and weighting stated; the definition is footnoted, the questionnaire item is not published
40% of respondents from organisations over $1 billion report scaling AI agentsSelf-reported survey, consultancySame survey. Up from 27% a year earlier; the figure for smaller organisations is 22% and flat. About two in ten overall report scaling software coding agents.As above
64% of US adults use AI; 24% of AI users use an agent regularlySelf-reported survey, investor-commissionedMenlo Ventures, 2026 State of Consumer AI, published September 16, 2026, from a Morning Consult survey of 5,067 US adults fielded in July 2026. Our September 16 post tables more than 40 of its figures with the base for each.Sample, field date and bases stated
93% of chat and Cowork conversations produce an artifact; 54% of Claude Code sessions are served by Opus against 10% of chat and CoworkCounted telemetryAnthropic Economic Index report, June 2026. An output classifier over chat and Cowork conversations, which the report calls 'Claude conversations'; the Opus comparison adds Claude Code sessions. The population is Anthropic's own users, not the market. Anthropic publishes no n for either figure.Yes, methodology and footnotes
75% of Google's new code is AI-generated and engineer-approved; Anthropic says Claude leads 26% of its AI R&D work, weighted by person-timeVendor-reported programme dataGoogle's CEO at Cloud Next, April 22, 2026, with no denominator published; Anthropic's post of September 17, 2026 with a published method, figure as of August 2026. Each counts something different; our September 18 census keeps every definition.Anthropic yes; Google a spoken figure with no method
AI agents and assistants spending of $29.2 billion in 2026, $65.5 billion in 2027Analyst forecastGartner press release, September 16, 2026, from its worldwide AI spending forecast; 2025 is given as $16.5 billion. No sample, question or method is published; the forecast document is available to Gartner clients only.No; the underlying document is gated

The Census figures come from the Bureau's May 26, 2026 story on its Business Trends and Outlook Survey, which also reports the national rate hovering between 17% and 20% from December 2025 to May 2026 and sector rates of 39.7% for Information and 33.9% for Finance and Insurance; the sample of about 1.2 million businesses in six panels, asked every 12 weeks and collected every two weeks, is from the Bureau's survey methodology page. The telemetry row is from Anthropic's Economic Index report. The vendor row draws on our census of disclosed figures, which keeps the exact definition beside each of 28 numbers from 13 companies, and the consumer survey row on our September 16 statistics post.

A methodology note, as published

McKinsey's own note for its 2026 survey reads: "The online survey was in the field from May 4 to June 8, 2026, and garnered responses from 1,719 participants in 97 nations representing the full range of regions, industries, company sizes, functional specialties, and tenures." That sentence is what makes the figure quotable. A figure without one is not.

03The worked caseWhy 20% and nine in ten are both true

Put the two headline figures side by side and read the methodology rows. The Census asks a business whether it used AI in any business function in the past two weeks, samples about 1.2 million businesses of every size, and reports that firms with four or fewer employees are under 20% while firms with 250 or more are at 37%. McKinsey asks a respondent whether their organisation uses AI regularly in at least one business function, reaches 1,719 people of whom 36% work at organisations over $1 billion in revenue, and weights the answers by each nation's share of global GDP.

4.5×
the ratio between the two headline figures, nine in ten over 19.8%
The same word, two denominators

Most of the gap is who is in the sample, and the rest is the verb

One sample is dominated by the smallest businesses in the economy, because that is what the economy is made of; the other is dominated by large enterprises, because that is who answers a consultancy's survey. Both now ask about any business function; one asks whether it happened in the past two weeks, the other whether use is regular. Restrict the Census to its largest firms and the two figures move toward each other without either changing.

The same reading applies within a class. Two surveys, one official and one consultancy, differ by population and weighting. Two telemetry figures from two platforms differ by whose users were counted. Two vendor figures differ by definition, which is why the disclosed-share census refuses to rank them. A forecast differs from all of them by being about a year that has not happened.

04The filterThree questions before you quote one

Each question maps to a column of the table, and each has an answer that ends the conversation. Our citation checklist for original research covers the same ground from the publisher's side.

Which class is it, and does the class answer the question I am using it for?
A forecast cannot show current adoption. Telemetry cannot show market share. A vendor's programme figure cannot show anything outside that vendor. If the class does not match the claim on the slide, no sample size will rescue it.
Class
Who was in the population, and what was the exact question?
Find the questionnaire or the methodology note. If the wording is unpublished, the figure is quotable only with the author's own phrasing attached; if the population is unpublished, it is not quotable as a rate at all.
Population and wording
What is n, what is the date, and is the method public?
A sample under a few hundred, a field date older than a year, or a gated method each moves the figure from evidence to anecdote. Say which of the three applies rather than dropping the number silently.
n, date, method

05In the deckHow to cite this in a deck

The honest slide is not harder to make than the misleading one; it has one more line. Give the figure, then the class, the population, the question in a few words, and the date, with the author named. Two examples from the table.

Two citable lines

"19.8% of US businesses used AI in the past two weeks (Census Bureau survey of about 1.2 million businesses, as of May 3, 2026)." And: "Nearly nine in ten respondents said their organisation uses AI regularly in at least one function (McKinsey Global Survey, 1,719 respondents, mostly large organisations, May to June 2026)." Put both on the same slide and the audience learns something true about the difference between small and large organisations. Put one on the slide as "AI adoption" and they learn nothing.

Our collection of agentic AI statistics carries the source for each of its figures for the same reason. If your team needs a defensible evidence base for an AI investment case, sorted this way and dated, that is part of what we do under AI transformation.

06ConclusionThe numbers do not disagree; the questions do

Which to use

Use the class that matches your claim, name the population and the date, and put the method beside the number

For the state of the whole economy, the official survey with the published questionnaire. For what a platform's own users do, that platform's telemetry, labelled as such. For a vendor's claim about itself, the vendor's figure with its definition attached. For a sense of where spending is heading, a forecast, labelled as a forecast. Any of them can go on a slide. None of them can go on a slide alone.

Digital Applied

An AI investment case built on evidence you can name.

We assemble the adoption and ROI evidence for an AI business case with the class, population, method and date attached to every figure, so it survives the first hard question.

Evidence sortingSource verificationBoard-ready figures
Your next project

A deck that survives the methodology question

  • Every figure sorted by evidence class
  • Population, n and date on the slide
  • Vendor claims labelled as vendor claims
Questions and answers

Applying this post

There is not one. The Census Bureau's survey gives 19.8% of US businesses using AI in the two weeks to May 3, 2026, with 37% for firms of 250 or more employees. McKinsey's survey of 1,719 mostly large-organisation respondents gives nearly nine in ten respondents reporting regular use in at least one business function. Each is the answer to its own question, and the honest citation names the question.
Digital Applied newsletter

Deep dives on AI, marketing and development.

Practical guides and fresh insights by email. No recycled takes.

Related dispatches

Continue reading

AI Development

AI Research Claims: What Has Actually Been Verified?

Assess AI research claims with a practical evidence matrix. Separate formal proofs, measured results and demos, and record what each check establishes.

September 9, 2026 · 6 minRead
AI Development

Your AI Agent Didn't Get Worse. Your Metrics Broke.

A real capability regression and a broken measurement pipeline look identical on one dashboard. Two tests separate them, five common breaks, a first-hour plan.

September 21, 2026 · 7 minRead
AI Development

Give an Agent the Facts Before It Rewrites Your Archive

Before an AI agent rewrites hundreds of pages, index every claim, figure and source already in them. The schema, three rules, and what to do when a claim fails.

September 21, 2026 · 7 minRead
AI Development

Why Parallel AI Agents Quietly Skip Part of the Work

Fan a batch out across subagents and items vanish with zero errors. The fault is the hand-written work list, not the agents. Three corruption points, one fix.

September 21, 2026 · 7 minRead
AI Development

AI Agent or Fixed Workflow? A Business Decision Table

Choose rules, an AI-assisted workflow or an autonomous agent by checking task uncertainty, verification and consequences with a business decision table.

September 13, 2026 · 5 minRead
AI Development

Deleting AI Agent Memory: Where Stored Copies Survive

Deleting AI agent memory takes more than clearing a chat. Map stored copies, retrieval indexes and backups, then verify what your system can still recover.

September 4, 2026 · 6 minRead