Almost every firm has adopted AI. Almost none can find it in their numbers.
Those two facts come from the same survey, about the same firms, in the same year. Together they rule out both easy answers: not too early, because adoption already happened, and not incapable, because the firms buying it say the capability is real. Something sits in between. This site measures it, occupation by occupation, and keeps the case against itself in plain sight.
Four links sit between a capable model and a changed job.
Two of them are well measured and they disagree with each other. The step between them is measured by nothing at all, which is exactly why they disagree. That step is the subject of this site.
- 1CapabilityWhat can the model do?Measured byBenchmarks, METR time horizonsState: Racing
- 2AdoptionIs it in the building?Measured byAnthropic Economic Index, DMPState: Near-universal
- 3AbsorptionDid the organisation change shape?Measured byNothing measures thisState: The gap
- 4OutcomeDid employment move?Measured byONS PAYE RTI, LFSState: Almost flat
Not on a job. On the first two years of one.
If you cannot redesign the organisation, the only lever that banks any of the capability is declining to replace the junior who left. Split the under-25s by whether they are still students and you can see it: the ones for whom work is a career are losing ground on both sides at once, while the headline barely moves.
Official statistics cannot see AI. So measure the usage directly.
AI is a technique applied across every job, not a category any official ledger has a row for, so it dissolves into whatever it passed through. Anthropic publishes what millions of real conversations actually do, job by job. Crossed with ONS employment, of 23.4M mapped workers, 12.3M lean toward the AI doing the task and 11.1M toward it helping the person. What it mostly produces is documents, not code.
What AI actually does →This could just be early. Here is exactly what would break it.
Each watch below is written before the print that settles it, and the verdict stays on the record whichever way it lands. One has already gone against us.
We benchmarked our own scored model against realised job change and published the failure: ρ = +0.05, no better than noise. The measured usage layer passed (ρ = −0.150, p = 0.014), so measured data leads the site. Every figure traces to a primary ONS, Bank of England or Anthropic source through a committed script, and the strongest evidence against this read sits next to the falsifiers that would sink it. The whole dataset is open.