Apollo's AI paper finds a 6.7% decline in real wage growth in 11 high-exposure occupations. Its robustness check gives 1.89% and 1.74% at cutoffs of 0.4 and 0.6, and says 6.7% sits between them.
The 55% in this week's AI layoff coverage comes from an Orgvue survey fielded February to March 2025. It measures 55% of the 39% who made AI-driven cuts, and it carries no date in the story.
A think tank rebuilt techUK's 40,200-job projection from 20 planning applications and got 10,400. Both sides use the same metric. The government calls it misleading.
McKinsey's 2025 survey had 32% expecting AI head count cuts. Its 2026 survey has 14% reporting them, and 39% expecting them next year. Same survey, same cohort.
BLS revised July 2026 payrolls up 44,000, from -23,000 to +21,000. The same August release has its own -23,000, in information, and never uses the word AI.
Revelio Labs reports AI-exposed occupations down 6%, and 19% for workers aged 22 to 25, alongside AI-exposed firms cutting fewer staff. Challenger puts AI fourth for August and first for 2026.
Oracle's FY2026 10-K names AI as a cause of workforce reduction, twice. The four Oracle layoff figures in circulation measure four different things, and one is unsourced.
Salesforce's Q2 FY27 filing reports 7.0 billion Agentic Work Units. Its own definition makes that a count of discrete AI-executed tasks, not a headcount number.
HP's own SEC filings tie a 4,000-6,000 headcount cut to AI adoption. The only mechanism they name is a voluntary retirement offer for US staff aged 55 and over.
Stanford now puts employment for 22-to-25-year-olds in AI-exposed jobs 19% below a kept-pace path. Coverage pairs that with 13% from last year. On the same measure, last year was 15%.
The most-quoted AI layoff figure counts employers who named AI in an announcement. The same report keeps a second, hedged category for a contested case. It holds 12 people.
Andon Labs published the case on 14 August 2026. The AI logged 6 of 17 late arrivals, needed a human prompt to decide, and people reviewed and delivered the termination.
In June, Veritone told the SEC its workforce reduction would be substantially complete by late July 2026. The 10-Q it filed on 13 August 2026 says the first half of 2027.
A Richmond Fed brief published 12 August 2026 finds the job-finding decline is concentrated among strongly attached workers, particularly in AI-exposed occupations.
Rapid7 and N-able each filed an 8-K on 10 August 2026 disclosing workforce cuts. Both CEOs invoked AI. Neither legally binding disclosure mentions it. Why that gap matters.
BLS reported payrolls fell 23,000 in July 2026. Its own confidence interval on that monthly change is plus or minus 122,000. What the report can and cannot support.
The New York Fed says 1% of service firms laid off workers because of AI. Challenger says AI was cited in 33% of July cuts. Four reasons the numbers cannot be compared.
The ONS surveyed 38,637 UK businesses and the Fed asked nearly 750 executives. Both find adoption rising steeply and aggregate employment holding. The gap is the finding.
ADP Research and Stanford measured pay attached to individual tasks in IT jobs. Five declined after 2022, advisory work held. The sample is ~7,000 workers, and that matters.
The Digital Omnibus took force on 27 July and pushed high-risk AI obligations to December 2027. Article 50 transparency was left untouched and applies from 2 August 2026.
Amazon cut staff inside its own AGI unit in July 2026. If proximity to AI is not what protects you, what does? The payroll evidence points at task structure, not job titles.
The AI literacy duty has applied since 2 February 2025 and becomes enforceable on 2 August 2026. What the Digital Omnibus actually deferred to 2027 — and what it did not.
What 1,822 documented evidence entries across 506 companies say about where AI displacement is actually happening first — and where it has not started yet.
Most AI training is bought on faith and never measured. A practical framework: baseline first, then reclaimed hours, capability artifacts, and risk reduction.
Most AI training treats everyone the same, which is why most of it fails. The four maturity levels, how to self-diagnose yours, and what moves you up one.
Inside the evidence base behind the Protocol: what a documented evidence entry is, how exposure scores are computed, and what the data can and cannot tell you.
You do not need to write code to be AI-literate. A practical four-week starting point for managers, analysts, and other knowledge workers — no engineering required.
Want personalized insights instead of articles?
Take the free assessmentFree, 7 questions, scored in your browser.
