UK AI JobsExpected · in use · happened
The gap · tracked every ONS print

Firms expect a bloodbath. The payroll data hasn't flinched.

Firms' AI-displacement EXPECTATIONS are rising exponentially while REALISED displacement is still nearly flat, and softened on the latest print. The gap between the two, tracked over time, is the falsifiable signal this instrument exists to measure.

Realised through 2026-07-21 print (PAYE RTI to June 2026)Exposure model 2026.03 (static)Next node · 2026-08-18 ONS Labour Market Overview
Expectations doubled · realised stayed low
Firms' AI-cut expectations vs realised, with scenario fork0%5%10%15%20%Q4 2025Q1 202621 Jul →Expected 13.7%≥5% cut, doubled in a quarterRealised 8%converge?or hold?gap ~1.7×
Expected (BoE DMP) Realisedsolid = observed · dashed = scenarios for the next print
The three ledgers · the front-door numbers, tracked from here print by print

Between the fear and the fallout sits a third number nobody was reading: what AI is measurably doing right now.

① Expected
13.7%
of UK firms expect AI to cut ≥5% of their staff within 3 years
BoE Decision Maker Panel · doubled from 6.3% in one quarter
② Measured now
47.5%
of real AI usage augments the worker rather than automating the task. A near-even split.
AEI · 12.3M automation-leaning vs 11.1M augmentation-leaning of 23.4M placed workers
③ Realised
8%
report AI has actually reduced their headcount
BoE DMP · realised, not expected

The fear is a forecast. What AI is actuallydoing splits almost evenly between doing the task and assisting the person, and that near-even measured middle is part of why the expected cuts haven't reached the payroll data yet. AEI real Claude.ai usage (Apr+May 2026, Claude users only, ~9% of codes unmapped, usage ≠ jobs lost); expected/realised = BoE DMP. Method →

The catch · 4AEI releases, Feb ’25 → Jun ’26

But that measured middle is sliding toward automation.

The augmentation share of Claude usage (how often AI assists rather than does the task) has fallen from 57.4% in early 2025 to 51.2% now. As models got more capable, real usage tilted toward doing the work. If that trend holds, the near-even split is temporary, and the gap starts closing from the usage side.

57.4%Feb '2556.9%Mar '2548.9%Sep '2551.2%Jun '26

Across AEI releases, Feb 2025 to Jun 2026. Early points are computed from the published interaction-type split; later points use Anthropic's published augmentation_pct / collaboration_bucket_augmentation_pct at GLOBAL. The 5-mode taxonomy is stable across releases, but sample, model mix and task-mapping evolved, read the TREND, not precise levels. Two intermediate releases (2026-01-15, 2026-03-24) ship raw counts only and are omitted. The three 2025 points are transcribed by hand from their published release files, source recorded per point. The newest point is derived by scripts/refresh_aei.py from the committed slice aei_global_overall.csv. Method →

The realised side · what actually landed

Re-aging holds. The acute youth signal softened, but stayed high.

PAYE RTI to 2026-06 · YoY change
Under 25
-0.6% · -25,140
Still negative, but the fall narrowed by a third between May (-38.1k) and June (-25.1k) on this print's vintage. What is left of it is entirely the under-18s.
Under 18
-9.0% · -40,099
The steepest fall of any band and accelerating (-6.6% in March, -9.0% now). Down 25% from its Sep 2022 peak. Confounded: this is the Saturday-job cohort, and April's step down lands on the minimum-wage uprating.
18-24
+0.4% · +15.0K
Positive for a second print and strengthening. The entry cohort a displacement wave would hit first is growing.
25-34
-1.9% · -130,344
The largest absolute fall in the table, and the band that has barely improved while every other one did.
35-49
+0.7% · +65.7K
50-64
-0.7% · -55,264
65+
+6.0% · +73.6K
The only fast-growing band. The re-aging signal, monotonic since ChatGPT.
All UK
-0.2% · -71,460
The exposure gradient · corrected 1 Jul
16–24 is the only band whose exposure×cohort gradient stays negative under every specification we ran (-0.9pp), and it is small. The −3.1pp shipped before 1 July was our own join bug; the correction is logged below.
The one clean tech signal
Information & Communication -2.5% YoY: unchanged for a second print while the UK total improved from -0.4% to -0.2%, so its gap to the national rate widened. The one clean AI-ish industry signature.
Since ChatGPT
Under-25 payroll -3.89%, 65+ +19.43%. Structural re-aging is monotonic since ChatGPT. What thinned inside the under-25s is the under-18 band, not the 18-24 entry cohort, which is now growing.

The one dial already flashing amber: 16–24 unemployment at 16.4%, an ~11-year high.

ONS LMS · to 2026-04

MGWY = 16-24 incl. full-time students (the 'decade-high' basis). AIXT = excl. students. MGSX = 16+ headline. Quote the basis with the figure. On the month the headline 16+ rate (4.9%) held flat while the 16–24 rate rose. If AI hits the entry rung first, this is where it shows.

16–24 incl. students16–24 excl. students16+ headline
The self-audit · scored vs measured vs reality

We benchmarked our own model against realised job change. It failed. The measured data didn't.

Same test for all three: rank correlation against realised 2024→2025 employment change, occupation by occupation, on the same 272 unit groups. Our Gemini-scored exposure and outlook read as noise. The measured automation share carries a weak but real signal. That is why measured data leads this site and the scored layer stays a labelled call.

AEI measured automation share
measured usage
ρ=-0.150p=0.014
signal, weak
Site scored exposure (Gemini, 0-10)
our scored call
ρ=+0.048p=0.433
no signal
Site scored outlook
our scored call
ρ=-0.051p=0.401
no signal

Honest about the effect size: even the measured signal explains ~2.2% of rank variance, and on the widest join it is ρ=-0.08 (p=0.15, N=329), short of significance. Realised displacement is still nearly flat. That is the gap story, and it cuts both ways. The full audit →

The honest part · what would break this

The claim, and exactly what would sink it.

A tracker is only worth trusting if it names its own falsifiers, keeps the strongest evidence against it in view, and logs its own corrections. Each watch below was written before the print that settles it, and the verdict stays on the card whichever way it lands. And here are the outside voices who say there's no AI footprint at all.

Does the 25-34 payroll fall (-1.9%, -130k) keep resisting while every other band improves?
If yes · The squeeze moved up a rung from the entry cohort to early-career, which is what a task-automation story predicts next.
If no · The 25-34 fall was the same cyclical softening as the rest of the market, arriving on a lag.
Does 18-24 payroll stay positive on 21 Jul?
Resolved · 2026-07-21
Yes, and by more. 18-24 payroll came in at +0.43% (+15.0k) against +0.19% (+6.4k) on the June print, so the acute 2025 youth-thinning did not resume. By our own pre-registered reading that weakens the acute-displacement story. It does not clear the youth picture: 16-24 unemployment hit 16.4%, an ~11-year high, and the under-18 band fell 9.0%, so the pressure sits at the hiring gate and one rung below, not in 18-24 payroll levels.
If yes · The 2025 acute youth-thinning was largely an artefact, the expectations-reality gap is real and the displacement story weakens.
If no · Acute displacement resumed, the gap is starting to close from the realised side.
Does the realised-cut share (~8%) rise toward the expected share (13.7%)?
If yes · Expectations were a leading indicator, the gap closes and the displacement thesis strengthens.
If no · Firms are over-forecasting their own AI cuts, expectations are sentiment, not realised.
Does Information & Communication (-2.5%) keep falling while other sectors hold?
If yes · The tech-junior-ladder mechanism is the real, narrow AI signature.
If no · The I&C fall was cyclical, not AI.
Corrections · what we got wrong, on the record
2026-07-01
Found · Our realised-change join treated SOC2010 codes from the ONS ad-hoc as if they were SOC2020 codes. 90 of 209 joined occupations, 8.19M workers (24.8% of employment), were matched to a different-meaning occupation: 9233 is Cleaners in SOC2010 but Exam invigilators in SOC2020; Nurses, Programmers and Secondary teachers were mis-keyed the same way.
Fixed · Rebuilt the join on the ONS dual-coded SOC2010 to SOC2020 relationship table (committed crosswalk + scripts). Scored coverage rose from 209 to 306 of 369 occupations.
Effect · The headline youth exposure gradient fell from -3.1pp to -0.9pp and is no longer unique to 16-24 in the simple read. The self-audit correlations were re-derived on the fixed join before publication. Full trail: adhoc3410_join_audit.json + aei_self_audit.json.
2026-07-01
Found · The 'most learning-led' card on /methodology and /cuts was hand-picked during the exploratory pass and skipped its own rank 4: telecoms network installers (29% learning share) passed every stated filter but was left out, so IT user-support technicians held the sixth row instead.
Fixed · The list is now emitted mechanically by the committed script (scripts/refresh_aei.py, emit_learning): top 6 by learning share across the 146 work-filtered occupations. Telecoms network installers enters at rank 4; IT user-support technicians, rank 7, drops off the card.
Effect · One row of six changed on each card. The mean (~9%), the population (146 occupations) and the maximum (~37%) are unchanged.
2026-07-24
Found · The four realised macro series on The cascade (unemployment rate, youth unemployment rate, vacancies, UK payroll change) were labelled ONS but came from a separate local project, at an older vintage and with no committed provenance. Against the published ONS series they were out by up to 0.3pp on the headline unemployment rate, up to 1.1pp on the youth rate and up to 98k on one vacancies quarter, and the payroll headline was still showing Nov 2025 (-171k) seven months on.
Fixed · Rebuilt all four straight from the committed ONS files by scripts/build_realised_series.py (LMS quarterly rows, VACS02 calendar quarters, PAYE RTI sheet 28). The rebuilt payroll series reproduces the ONS bulletin's own headline figures exactly: May 2026 -85,000 (-0.3%) and June 2026 provisional -71,000 (-0.2%).
Effect · Every point on those three charts moved slightly and the payroll headline moved from -171k (Nov 2025) to -71k (June 2026). The shapes are unchanged. Vacancies now run one quarter further, to 2026-Q2, and their plotted peak corrects from 2022-Q1 to 2022-Q2.
The skeptic wall · the strongest evidence against
→
NY Fed (14 May 2026)
No economy-wide AI employment footprint; exposed-occupation decline predates ChatGPT.
→
Yale Budget Lab (15 Jun 2026)
No measurable AI break in the aggregate occupational mix.
→
EPI (21 May 2026)
AI-exposed employment trends track pre-existing structural decline, not a new shock.
→
Confound, April NICs + min-wage
Youth-heavy retail/hospitality hit by labour-cost shock; part of any youth fall is not AI.

The honest, defensible read is the NARROW one: a tech/ICT junior-ladder signal plus firm expectations accelerating, NOT mass economy-wide displacement. This tracker is built to show exactly where that claim holds and where it would break.