Blog
Evidence3 min read

The AI statistics that survive checking

We tried to source every number on this site. Several of the most-quoted figures in the industry did not survive the attempt. Here is what happened to each of them.

A long row of bound volumes on a library shelf, receding into the distance.
EvidencePhotograph via Unsplash

This site has one rule that governs everything on it: a figure appears only with the organisation that published it, the period it covers, and a link to the publication itself rather than to coverage of it. It sounds like a small piece of housekeeping. In practice it removed about a third of the statistics we started with, including several that are quoted in almost every presentation given on this subject.

What follows is what happened when we actually checked. It is worth reading before you quote anything in this field, including us.

The one that got away

You have seen this figure. It is on slide three of most AI strategy decks written in the last year: 95% of corporate generative AI pilots produce no measurable return, from a study by MIT's NANDA initiative.

It is a genuinely useful finding and we wanted it. The PDF it comes from now redirects to a group overview page. Anyone following the citation lands somewhere that does not contain the number. So it is not on this site, and every deck still quoting it is quoting something the reader can no longer verify.

A statistic you cannot follow back to its source is not evidence. It is a rumour with a decimal point.

The one that is marketing

The single most quoted figure in search marketing is that 53% of all website traffic comes from organic search. It comes from BrightEdge, who sell SEO software.

That is not an accusation of dishonesty. It is a structural point: a firm measuring the value of the category it sells into has an interest in the answer, and no external party can audit the sample. We do not publish it. The same test removed almost everything available on AI's effect on marketing, because almost all of it is published by firms selling marketing services.

The one that could not be reconciled

KPMG runs a quarterly pulse survey on AI agent deployment. One quarter reports 42% of organisations having deployed agents. The next reports 26%. Both are presented as growth.

The definitions almost certainly changed between the two, and the published material does not say how. A figure we cannot explain the movement in is a figure we cannot put in front of a client, so neither number appears here.

The one that was rounded before it reached us

A US Census figure we wanted was reported as “around 37%”. The rounding was done by the publisher, not by us, which does not help: a reader has no way to know whether that is 36.5 or 37.4, and a chart drawn from it would be precise about something imprecise. Dropped.

What survived, and why it is worth more

What is left is thirteen publications: national statistics offices, one research institute, two consultancies who publish their methodology, and one peer-reviewed paper. The figures are less dramatic than the ones we lost. They are also checkable, which is the only property that matters after the first time somebody in the room says “where did that come from?”.

20.0%
of EU enterprises with 10 or more employees used AI technologies, up from 13.5% a year earlier. Denmark reached 42.0% and Romania 5.2% in the same year.2025 · Eurostat

That last pair is the reason averages should be distrusted generally. One region, one definition, one year, and an eight-fold spread between the top and the bottom. A business benchmarking itself against the EU average of 20.0% is comparing itself to a number that accurately describes almost nowhere.

How to read a statistic in this field

Four questions, in order. They take about ninety seconds and they will discard most of what you are shown.

  • Who published it, and do they sell the thing being measured? If yes, it is a marketing claim. It may still be true. It is not evidence.
  • What exactly was counted? “Businesses using AI” means firms with ten or more employees in the UK, establishments in Saudi Arabia, and firms of any size in the United States. These are not the same measure and they are routinely charted together.
  • What is the ceiling? Half of firms reporting a cost saving sounds transformative until you read that most of those savings are under ten per cent. The second half of that sentence is the one that gets left off the slide.
  • Can you reach the source in one click? If not, treat it as unverified, however respectable the name attached to it.

Apply those four to this page as well. Every figure above links to the organisation that published it, and if one of those links ever stops resolving, the figure should come off this site the same way the others did.

Sources

04

New writing, when there is some

A short note when something is published: what it is about and a link. No more than a couple a month, and nothing else.

Start with
the diagnostic.

A scoping conversation costs nothing and ends with a straight answer about whether there is enough here to be worth doing. If there is not, we will tell you.

Book a scoping callinfo@skaiyon.com

Questions. Asked before every engagement.

Your data is yours. You can export it at any time, and it is exported to you in a documented format before any engagement closes. The software itself is licensed: we build it around your business, host it, and run it, and you pay monthly for that. If you'd rather own it outright, that's possible. It's a different kind of engagement and it's priced accordingly. Licensing keeps maintenance our problem rather than yours, which is why it is the default.

Two weeks' notice, either side. Your data is yours. You can export it at any time, and it is exported to you in a documented format before any engagement closes. The system stops running. If you'd rather keep it running, the ownership option is available at that point as well. Ending the retainer doesn't force you to lose what was built.

Typically four to eight weeks from signing to a working system, depending on how many tools it connects to and what state the data is in. A scoping conversation and a written plan come first, so the timeline is agreed before anything is committed to.

Access to the tools and data the system will work with, one person who can make decisions, and a few hours in the first two weeks while we map how work actually moves through your business. After that, very little. The point of the engagement is that it runs without your attention.

Usually. Most business software exposes an interface we can build against, and where one doesn't there's normally a way around it. Which connections are viable is settled in the scoping conversation, so you find out before committing rather than after.

Hosting region for a system we build for you is decided at the start of the engagement, not inherited from a default. If data residency is a hard requirement, raise it in the first conversation and we will tell you plainly whether we can meet it. This website is separate and its processing is already fixed: enquiries, quote requests, bookings and chat are handled by Supabase in Tokyo, Resend in Tokyo, Cloudflare, cal.com and Moonshot AI, which the privacy policy names individually along with where each one processes. Nothing from a client system is ever sent to the chat assistant.

That's the normal starting point, and mapping them is part of the work. Automating a process nobody has examined just makes the confusion faster, so we don't start there. The first phase establishes how things actually happen, as opposed to how they're supposed to.