Somebody has to count what the sellers will not.
Two problems sit underneath every number published about AI adoption. One is definitional. The other is commercial. Neither is being fixed by the people currently producing the evidence.
The measurement problem
Adoption covers an employee pasting text into a chatbot and an organisation that has rebuilt its operations around machine decisions. Most published research counts them the same way.
A measure that cannot separate a tool somebody uses from a process that would stop without it is not a measure. It is a headline. We ask one question of every system instead: if this were switched off tomorrow morning, what happens to the work?
The evidence problem
Most of the evidence is produced by the organisations selling the thing being evidenced. That research is competent and it is commissioned to support a sale, which means it stops where the sale stops.
It goes deep on large enterprise transformation, because that is where the fees are. Nobody is funded to count how many of last year's programmes were quietly abandoned.
What that leaves uncounted
Small business
Undisclosed vendor surveys, with samples nobody publishes and frames nobody describes.
Mid-market
Barely covered, despite being where most of the economy sits and where the unit economics are least understood.
Most of the world
Country depth exists for a handful of wealthy markets. India, the Gulf and Southeast Asia are described from outside or not at all.
Failure
Discussed constantly, measured almost never. The most quoted number in the field rests on one unreplicated study.
The services layer
In most categories the spend on making a system work exceeds the spend on the system. No published map covers it.
Abandonment
Programmes that stopped. Nobody publishes them, because nobody is paid to.
What we do instead
Six standing series, published on a fixed calendar so the record accumulates rather than arriving in bursts. Each one answers the same questions in the same order, which is what turns fifty-two items into a dataset rather than fifty-two articles.
If a sentence would read identically in a vendor's deck, it does not ship. The Forum is worth reading only where it says something a seller cannot afford to say.
Where this goes
The weekly record is raw material. In 2027 it becomes structured: two hundred and twenty AI use cases across eleven industries and eleven business functions, each answering the same questions in the same order.
What the work actually is. Where the machine sits in it. What good looks like, and which metric is the wrong one. What breaks it. What binds it. Who solves it, in products and in services. And who has done it, with an evidence grade attached.
That is the Use Case Genome. Every failure case, category map, deployment and obligation resolves into a cell in it.