Almost every firm has adopted AI. Almost none of them can find it in their numbers.

Four links sit between a capable model and a changed job. Only three of them get measured.

1. 1Capability
   What can the model do?
   Measured by
   Benchmarks, METR time horizons
   State: Racing
2. 2Adoption
   Is it in the building?
   Measured by
   Anthropic Economic Index, DMP
   State: Near-universal
3. 3Absorption
   Did the organisation change shape?
   Measured by
   Nothing measures this
   State: The gap
4. 4Outcome
   Did employment move?
   Measured by
   ONS PAYE RTI, LFS
   State: Almost flat

The Anthropic Economic Index measures link two, because it watches what people actually do with the model. The ONS measures link four, because it counts jobs. Nothing measures link three. When the two ledgers either side of an unmeasured step disagree, the disagreement is not noise. It is the size of the step.

If you cannot redesign the organisation, there is exactly one way to bank any of it.

The one lever that banks part of the capability without changing anything is declining to replace the junior who left. No committee, no redundancy cost, no workflow redesign. It is the path of least organisational resistance, which is why it is the first effect to show up and so far the only one.

Read the middle number again. Firms sitting on a technology that compresses the tasks they bring to it by roughly sevenfold expect it to move their productivity by 1.4% over three years. That is not a forecast about the technology. It is a forecast about themselves.

It is not landing on the jobs young people do. Those are no more automatable than anyone else's.

This is the explanation everyone reaches for, and the data does not support it. Join the measured usage of every occupation to the ages of the people holding it and the line is flat. 2.2 points separates the most exposed cohort from the least.

Which forces the mechanism somewhere more uncomfortable. If it is not which occupations young people are in, it has to be their position within an occupation. Not the paralegal job, the first two years of the paralegal job. And no official dataset has a seniority column. Occupation by age exists. Occupation by rung does not exist anywhere.

Split the young by whether they are students and the two halves move in opposite directions.

PAYE payroll counts a student's Saturday shift and a graduate's first career job as one employee each. So when student work rises and career entry falls, payroll shows nothing and can even look healthy. It is why the 18-24 payroll line and the 16-24 unemployment rate currently point opposite ways, and the payroll line is the one that gets reported.

For the group this is a career rather than a job, the unemployment rate has gone from 10.7% to 14.3% since 2023 NOV. Controlling for a youth population that grew over the same period, the 16-24 employment rate fell from 51.3% to 50.5%.

They are not being pushed out of work. They are being pushed down.

Young people gained least in the rungs the labour market grew most, and most in elementary occupations. A displacement wave takes the headcount. This takes the rung and leaves the headcount, which is why it clears every check that looks at employment levels and none that look at what the employment is.

There is a reason this compounds rather than settling. The first rung was never only a pay grade. It was the mechanism by which you acquired the judgment that makes the higher rungs worth paying for. AI substitutes for codified knowledge, which is what a junior sells, and complements the tacit kind, which is what a senior sells. Removing the bottom rung does not shorten the ladder. It disconnects it.

A third of the payroll collapse everyone reported was never there.

PAYE RTI is built from employer submissions that arrive late, so the newest one or two months always undercount and are always revised up. Every payroll headline is read off exactly those months. The bias runs one way and only at the newest edge, so every month delivers a number that reads worse than it will turn out to be, and the correction arrives after the story has been written.

Applied to the current print, the headline -71,460 for June 2026 is more likely to settle near -39,082. Against a workforce of 30.3 million that is a rounding error. The payroll contraction peaked at -127,194 in November 2025 and has been closing since.

Every net job the UK has added since ChatGPT sits in two age bands.

PAYE payroll change since ChatGPT shipped. The UK added jobs overall, and every net job sits in two bands. The most durable pattern in the data is the workforce getting older, and nobody frames it as an AI story. It also runs in exactly the direction the mechanism above predicts. What AI substitutes for is the documented, checkable, teachable part of a job. What it does not touch is judgment, relationships and knowing which thing usually goes wrong. The first is what you sell at 22. The second is what you sell at 55.

The honest limits of all of this.

• The youth split is Labour Force Survey data, which the ONS classes as official statistics in development with acknowledged volatility. The payroll and revision findings are administrative and much firmer.
• None of this is causal. A hiring freeze caused by weak demand, higher employer National Insurance and a minimum-wage uprating would leave a similar mark on entry-level employment. The confounds are real and they are large.
• The clearest counter-signal is in the sector with the strongest AI story. Information and communication payroll is still falling year on year, and its job vacancies are up 17% over the same period. A sector being automated away does not advertise more jobs.
• The absorption argument predicts something testable that we have not tested: firms that redesigned should show large effects while everyone else shows none, so the spread across firms should be wide and two-humped rather than smooth.