UK Job Adverts Fell 39.7%. AI Isn't the Axis the Data Points To

Postings for UK programmers and software developers fell 73.4% between 2022 and 2024. That looks like a profession being eaten. But the same dataset carries an inconvenient companion fact: the occupation this site measures at exactly 0.0% observed AI exposure also lost postings over that window — while the care job we rate as more AI-exposed grew faster than the classroom job we rate as less exposed.

Something large did happen to UK hiring. It just doesn't sort by automatability.

The ONS released updated online job advert volumes on 21 August 2026, covering January 2017 to July 2026, broken out by 4-digit occupation code. The figures come from Textkernel, which scrapes roughly 90,000 job boards and company sites, with ONS removing duplicates. What follows is a read of that file — and an argument with the conclusion most people will draw from it.

The number that gets quoted, and the one that gets skipped

UK online job adverts peaked in 2022 at 13,780,433 for the year. By 2024 they were 8,312,308. That is a 39.7% fall across the entire labour market. [Fact]

That 2024 figure is also 27.3% below 2019's 11,441,508 — so this was not a return to pre-pandemic normal. It went straight through it.

Hold that next to the headline. Programmers fell 73.4% in a market that fell 39.7%. The occupation-specific damage is the gap between those two — about 33.7 percentage points worse than the market. Real, and the steepest in the file. But the market did most of the work, and any story that explains only the tech numbers is explaining the smaller half.

Read in relative terms, the picture inverts. Of the 182 occupation unit groups with at least 20,000 adverts in 2022, 94 outperformed the market average. More than half of UK occupations did better than the overall collapse.

Where the losses actually cluster

The steepest declines from 2022 to 2024 sit in a tight band. Programmers and software development professionals fell 73.4% (and 77.1% against 2019). IT professionals not elsewhere classified fell 68.8%. IT business analysts, architects and systems designers fell 67.9%. Management consultants and business analysts fell 65.6%. Cyber security professionals fell 64.6%. Data analysts fell 64.0%.

Look at what unites that list. Graduate-entry. Project-funded. Paid out of discretionary budgets that move with interest rates. That is a coherent category — but it is not the category "work a language model can do."

The counter-evidence sits in our own data

Here is the part where this site's own figures cut against the easy interpretation.

Nursery education teaching professionals lost 7.8% of adverts from 2022 to 2024, and are down 26.3% against 2019. Our page for preschool teachers republishes an Anthropic Economic Index observed-exposure figure of 0.0% — the floor of the scale. An occupation with no measured AI exposure at all still shrank. [Fact]

Now the reverse case. Childcare workers carry an OpenAI rubric human-rated exposure of 33.0% and observed exposure of 1.2%. Teaching assistants are rated lower on the same rubric at 28.3%. If exposure drove hiring, teaching assistants should have held up better. They didn't: early education and childcare practitioners grew 29.0%, higher-level teaching assistants grew 15.1%.

Then there is a problem with the categories themselves. SOC code 2134 bundles software developers and computer programmers into one line. On our pages those two sit far apart on measured exposure — 28.8% observed for developers against 74.5% for programmers, a 2.6-fold gap. ONS reports them as a single number that moved as a single number. A figure spanning two populations our own data rates 2.6 times apart cannot be reading exposure; it is reading a budget. [Claim]

The honest objection is worth stating plainly: adverts fell most in office work that AI demonstrably assists, so isn't automation the simplest explanation? It would be, if the ranking held up. It doesn't. Cyber security is adversarial, defensive, judgment-saturated work that is among the harder things to hand to a model, and it fell 64.6% — in line with the rest of the tech band. Meanwhile the occupations at the bottom of every exposure scale split in both directions, some growing 29% and some shrinking 8%. A variable that predicts nothing within either group is not the variable doing the work.

The likelier account is duller: a 2022 technology hiring bubble unwinding into higher interest rates and frozen discretionary project spend. I cannot prove that mechanism from advert counts alone, and I am not going to pretend otherwise. What the data supports is the negative claim — not that AI is irrelevant, but that automatability does not sort this table.

2025 was a trough, not an ending

Skipping the recovery would be its own distortion. Programmer adverts went from 98,969 in 2024 to 113,816 in 2025, a 15.0% rise. IT business analysts rose 35.9%. [Fact]

That is a partial reopening, not a restoration. Programmer volumes in 2025 still sit roughly 69% below the 2022 peak. The entry gate narrowed violently and has since opened part-way.

What this dataset cannot tell you

These are official statistics in development — the designation formerly called experimental — and they carry real handling instructions. March 2026 is fully suppressed. October 2025 through February 2026, and June and July 2026, are partially imputed. ONS advises against reading month-to-month movements after November 2025 and recommends long-run trends instead.

So every claim above stops at 2024, the last fully clean calendar year, with 2025 used only for direction. The most recent month in the file is not a headline and should not be treated as one.

Three further limits. Online adverts are not jobs: word-of-mouth hiring, internal promotion and direct sourcing on LinkedIn all remove a posting without removing a role, and the mix of those channels has shifted. This is one country, and UK-specific shocks do not travel. And beating a collapsing market is not safety — of those 94 outperforming groups, plenty still shrank. [Estimate]

What to do with it

The structural read is that UK demand rotated toward work done in person, with people, under supervision. Early education and childcare adverts are up 99.6% against 2019. Higher-level teaching assistants are up 62.0%. Sports coaches and instructors are up 68.6%. Those are five-year gains in a market that lost more than a quarter of its volume overall.

The relative frame is the useful one for planning. Nursery teaching declined 7.8% — and still beat the market by roughly 32 percentage points. If you are deciding where to put two years of training, the question is not "does this job shrink" but "does it shrink less than everything else."

For anyone in or entering digital work: the honest summary is that the door narrowed and is reopening, not that the room emptied. The 2025 rebound is real. So is the fact that it starts from a floor 69% below the peak, which means the competition for junior roles stays unusually sharp for a while. The practical move is to be the candidate who can show delivered work rather than a credential — and to treat the management consulting and data analysis routes as cyclical rather than closed, because that is what the 2025 numbers say they are.

Sources

• Office for National Statistics, Labour demand volumes by Standard Occupation Classification (SOC 2020), UK: January 2017 to July 2026, released 21 August 2026. Dataset: labourdemandbyoccupation.xlsx, Table 3 (Total UK by 4-digit SOC). Underlying adverts collected by Textkernel; de-duplication by ONS. Official statistics in development. https://www.ons.gov.uk/employmentandlabourmarket/peopleinwork/employmentandemployeetypes/datasets/labourdemandvolumesbystandardoccupationclassificationsoc2020uk
• Contains public sector information licensed under the Open Government Licence v3.0.
• Exposure figures quoted from occupation pages on this site are republished from the Anthropic Economic Index (CC BY 4.0) and the OpenAI "GPTs are GPTs" rubric (MIT License), reproduced without modification. Year-on-year percentage changes and market-relative comparisons are our own arithmetic on the ONS dataset.

AI-assisted analysis. This article was drafted with AI assistance from the primary ONS dataset and reviewed before publication. Figures were read directly from the ONS release rather than from secondary coverage.