How to read an AI vendor's adoption number

A vendor publishes a milestone. On 31 August 2026, OpenAI announced that ChatGPT Ads had reached $1 billion in annualised revenue run rate in less than 200 days after launch, that the platform is used by tens of thousands of advertisers in more than 40 countries, and that the advertising-supported free tier helps keep ChatGPT available to more than 1 billion weekly active users [8]. Figures like these arrive in your feed with an implication already attached, which is that everyone else has moved and you have not. That feeling is the actual product being shipped. Adoption figures are written by marketing teams for exactly this effect, and they work, because for most people they are the only data that ever appears about how many other businesses are really using this stuff.

This guide is a procedure for taking one of those numbers apart and working out what it should change. The short answer is that it should almost never change your shortlist, it can sometimes change your calendar, and it should never be the only input to either. If you are an investor trying to value a vendor, this is the wrong guide, because your questions are about revenue quality and retention rather than about whether to buy two seats. This is written for someone running a business who keeps seeing these figures and quietly wonders whether they are behind.

A user count is a claim, not a measurement

There is a useful benchmark for what a serious operating metric looks like. When a public company puts one in its financial disclosures, the SEC expects three things alongside it: a clear definition of the metric and how it is calculated, a statement of the reasons why the metric provides useful information to investors, and a statement of how management uses it in managing or monitoring the performance of the business [5]. If the company later changes how the metric is calculated, it should consider disclosing the differences compared with prior periods, the reasons for the change, and the effects on amounts previously reported [5]. That guidance took effect in February 2020, and it exists because undefined metrics are how companies flatter themselves without technically lying.

A product announcement is under none of those obligations. It often carries no definition, no denominator, no time window and no restatement history. The number is not necessarily false. It is unfalsifiable, which is a different and worse property, because you cannot check it, you cannot compare it with the same company’s figure from last year, and you cannot compare it with a competitor’s figure that was built a different way.

Start by opening the vendor’s own page before you repeat anything you read in coverage of it. Numbers detach from their source quickly and travel further than the sentence they came from. OpenAI’s announcement of its small-business program, published on 21 July 2026, is a fair specimen to practise on. Read today, it names ChatGPT Work and describes virtual training webinars, in-person AI academies across the US, new guides and customer stories, short-form videos, and integrations with partner tools from companies including Dropbox, Shopify, Intuit, Slack, Atlassian and Wix [1]. What it does not contain is any user count or count of businesses at all. The only figures in its text are two percentages from a workshop and one customer’s estimate of the hours saved on a single document [1].

Four moves make a number bigger than the thing it counts

The first is combining. A figure that covers two products at once, a general work assistant and a coding tool, cannot tell you whether both are growing or one is carrying the other. Combination is a choice, and it is made when the split is less impressive than the total.

The second is the missing time window. “Reached” is not a rate. A count with no period attached cannot be turned into growth, and growth is the only part of an adoption number that would tell you anything about where the market is heading. When a vendor has a good rate, it publishes the rate.

The third is the definition of a user. Seats sold are not people working. A trial that converts to a workspace is not a habit. ChatGPT Business, for instance, requires a workspace to hold at least 2 Standard or Premium seats in total [2], so the smallest possible customer contributes two of whatever the vendor is counting, whether or not the second person ever signs in. Nothing about that is dishonest. It is simply a reason that seat-derived counts run ahead of human usage, and the gap is invisible from outside. Notice the contrast in wording, too. “More than 1 billion weekly active users” [8] at least names a window and an activity. A bare count of users names neither, and the vendor chose which one to print.

The fourth is the self-selected sample, which is where the two percentages on that small-business page live. OpenAI writes that at its Small Business AI Jams “last year”, “78% of participants built a functional AI workflow in a single day”, and that 42% saved more than five hours a week with the help of AI [1]. Look at the population before the percentages. These are people who gave up a working day to attend an AI event run by the vendor, which selects for enthusiasm at every stage. The outcomes are self-reported. No sample size is given. And 42% saving five hours also means 58% did not. US consumer-protection guidance is built around exactly this shape of claim: where an advertisement features an endorser reporting exceptional results, the advertiser needs proof that the endorser’s experience represents what people will generally achieve using the product as described, and otherwise the ad must make clear to the audience what the generally expected results are [6]. A first-party workshop statistic with no denominator does not clear that bar on its own, and you should not treat it as though it had.

Independent counts exist, and they are much smaller

The counterweight to a vendor’s number is a number the vendor did not produce. In the US, the closest thing to a national instrument is the Census Bureau’s Business Trends and Outlook Survey, whose sample consists of approximately 1.2 million businesses split into six panels of approximately 200,000 cases each, with businesses in each panel asked to report once every 12 weeks and data collection occurring every two weeks [4].

The picture it produces is far more sober than any launch post. Across the six months of data collected from 14 December 2025 to 3 May 2026, overall AI use among US businesses hovered between 17% and 20%, while between 20% and 23% expected to be using it in the next six months [3]. The national rate as of 3 May 2026 was 19.8% [3]. Roughly one business in five, with another one in five saying it intends to get there.

The breakdowns matter more than the headline. Use is concentrated in large firms: as of 3 May 2026, 37% of firms with at least 250 employees reported using AI, and 32% of firms with 100 to 249 employees did [3]. Fewer than 20% of firms with four or fewer employees reported using it, and between December 2025 and May 2026 AI use increased among firms with at least 20 employees but did not change significantly among firms with fewer than 20 [3]. By sector on that same date, Information stood at 39.7% and Finance and Insurance at 33.9%, both above the national rate, while Retail Trade sat near 14% [3]. If you are a two-person business in retail, the honest reading is that most firms like yours are not doing this yet, whatever the milestone in your feed implied.

One more detail from that survey is worth more than all of its percentages. The Bureau revised the wording of its core question in November 2025, moving from asking about AI use in producing goods or services to asking whether the business used AI “in any business function”, and from 17 November 2025 it ran a second AI supplement measuring use across 15 business functions [3]. A national statistical agency moved its own headline number by rewriting one phrase, and it told everyone it had done so. That is the standard. Hold an undated, undefined vendor figure next to it and the comparison makes itself.

An adoption number can move your calendar, not your shortlist

There is a narrow, real use for these figures, and it is timing. A national rate of 19.8% [3] is neither a niche curiosity nor a settled default. It places you in the middle of an adoption curve rather than at either end, which means the cost of waiting another quarter is moderate and rising rather than either trivial or catastrophic. That is a genuinely different position from the one a vendor milestone implies, and it is the position the evidence supports.

So the legitimate move is to let a market signal change when you look, not what you buy. If you have not evaluated AI tooling against your actual work in the last six months, a widely reported milestone is a decent prompt to schedule that evaluation this month. If you have, the milestone changes nothing, because your own trial produced better evidence than any user count could. What a vendor’s adoption figure must never do is pick the tool, since it is not a measure of quality, it is not comparable across vendors who count differently, and the vendor with the biggest number is simply the vendor with the best-funded marketing department.

Turn that into a personal rule with one sentence: for any adoption claim that makes you feel behind, find one number about the same question that the vendor did not produce. If no such number exists, treat the claim as advertising, because that is what it is.

The only number that should move you is what finding out costs

The figure that deserves your attention is not on the vendor’s blog. It is the cost of answering the question yourself, and it is usually smaller than the anxiety the milestone produced.

Current published prices make this concrete. ChatGPT Business is $25 per user per month on the monthly plan, or $20 per user per month billed annually, with premium seats at $125 and $100, and a workspace minimum of 2 seats [2]. Claude Team is priced almost identically, at $25 per seat monthly and $20 billed annually, with premium seats at $125 and $100, while an individual Claude Pro seat is $20 billed monthly or $17 per month with the annual discount, billed at $200 up front [7]. Two seats for two months on monthly billing is $100. Choose monthly billing for a pilot even though it costs more, because the thing you are buying is the option to stop. Note also what the price parity tells you: on the two largest work assistants, seat price is not the variable you are choosing between.

The money is not the real cost. The real cost is the hours somebody spends putting genuine work through the tool, which is why the calculation below includes them and why you should be honest about the hourly figure you put in.

Run the pilot on work you were going to do anyway, not on the vendor’s demo, and finish it with something you actually shipped. That produces the one piece of evidence no adoption number can give you, which is whether the tool handles your material, your file formats, your clients and your standards. A milestone tells you about a market. A finished piece of real work tells you about you.

What still goes wrong

Independent surveys have their own defects and you should not swing from credulous to reverent. The Census figures are self-reported by whoever fills in the form at each business, AI use “in any business function” counts a marketing assistant who used a chatbot twice in a fortnight alongside a firm that rebuilt its support desk, the survey is published as an experimental data product [4], and the November 2025 question change means the current series cannot be compared cleanly with older ones [3]. A number being independent makes it harder to game, not automatically right.

The sceptical reading also overcorrects if you let it. Vendor adoption figures are weak evidence, not zero evidence, and a very large number that keeps growing is at least consistent with a technology that many organisations found useful enough to pay for twice. Being the last firm in your sector to try something is a real cost, and it is paid quietly, in work that takes three hours when a competitor spends twenty minutes. The failure mode this guide is written against is switching tools because of a headline. The opposite failure mode, deciding that all such numbers are noise and therefore doing nothing for two years, costs more.

Finally, everything priced or counted here is a snapshot taken on the date in the header, and prices, plan names, product names and survey question wording all move. That is the point of citing each figure to a page you can open yourself. Check the source before you spend money, and apply the same suspicion to this guide that it recommends applying to the vendors, since it is also a document written by someone with a view.