Every few months, someone declares the death or the resurrection of the labour market. Both camps are missing the structural thing.
In August 2025, Stanford's Digital Economy Lab (Brynjolfsson, Chandar, Chen) published the cleanest dataset we have on what generative AI has done to hiring: payroll records for 25 million US workers, tracked from ChatGPT's launch through mid-2025. Their finding:
Entry-level employment in AI-exposed occupations is down 13%.
Senior employment in the same occupations is up 6–9%.
Entry-level software developer roles fell roughly 20% (Stanford, Aug 2025).
AI didn't replace experts. It replaced the tasks we used to train juniors into experts. The bottom rung of the ladder is being sawn off.
The bifurcation
Zoom out to all US postings and the pattern repeats. Indeed Hiring Lab's January 2026 update:
- Total postings: +6% above Feb 2020 baseline.
- Postings mentioning AI: +134% above baseline. 4.2% of all US postings now.
- Software developer postings: 36% below baseline. 49% below for backend.
The market is bifurcated. Not shrinking, not booming. Splitting, right?
Salary data tells the same story from the compensation side. Lightcast analysed 1.3 billion job postings and found AI skills carry a 28% salary premium (~$18,000/year). Two AI skills triggers a 43% premium. 51% of AI-skill postings are now outside IT and CS (Lightcast, July 2025). PwC's 2025 AI Jobs Barometer puts the premium higher still, at 56%, with wage growth in AI-exposed sectors running twice as fast as non-exposed ones (PwC, June 2025).
If you're early-career, "will AI take my job?" is the wrong question. The real one: has AI taken the tasks that used to be your on-ramp?
The receipts
Here's what actually happened at named companies in 2024–2026. With caveats.
Klarna. February 2024: the AI customer-service bot, built on OpenAI, was handling 2.3 million conversations a month, the equivalent of 700 full-time agents. Headcount dropped from 5,527 at end-2022 to 3,422 at end-2024. A 40% cut. By May 2025, CEO Sebastian Siemiatkowski admitted the company had gone too far and started rehiring humans for complex cases (CNBC; Tech.co). The canonical "AI reversal" story.
Salesforce. Marc Benioff on the Logan Bartlett Show, September 2025: "I've reduced it from 9,000 heads to about 5,000, because I need less heads." Support costs down 17%. Agentforce now handles about 50% of support interactions (Fortune). Salesforce's official line is "rebalanced". Many were redeployed to sales and customer success. 4,000 roles going away the old way is still structural.
Amazon. Andy Jassy's June 2025 memo said the corporate workforce would shrink as generative AI rolls out. In October 2025, Amazon cut 14,000 corporate workers, its largest corporate reduction ever, on top of the 27,000+ cut since 2022 (CNBC). Challenger, Gray & Christmas attributed about 55,000 AI-linked US layoffs across 2025.
IBM. The AskHR chatbot now handles about 94% of routine HR queries. A couple hundred HR roles were replaced, but IBM used the savings to hire more programmers and salespeople. Net headcount grew (Entrepreneur). The useful shape of the pattern: AI shifts roles as much as it removes them.
Meta. Announced April 2026: 8,000 cuts targeting software engineers, recruiters, and middle managers, starting May 20. Internal AI tools (Metamate, DevMate) are claimed to handle up to 70% of routine coding and admin. Zuckerberg has telegraphed a 50:1 employee-to-manager ratio against the 7–15:1 industry norm (Tech Startups).
The pattern isn't uniform. Some companies are cutting, some are shifting, a few like Klarna have round-tripped. What connects them, honestly, is that AI is now a plausible board-deck justification for any structural change a company wanted to make anyway.
The productivity mirage
Which makes the next finding inconvenient for most narratives.
In July 2025, METR ran a randomised controlled trial on experienced open-source developers working on their own codebases, using Cursor with Claude 3.5 and 3.7 Sonnet. The finding: developers allowed to use AI were 19% slower. They believed they had been 20% faster (METR).
This is the single sharpest data point against the "one engineer does the work of three" story. The lift is real for juniors and unfamiliar codebases. On senior work in a codebase you know cold, it's negative. Self-reports overstate it by 39 percentage points.
Daron Acemoglu's 2024 NBER paper (published Economic Policy, January 2025) is the macro version. AI's decade-long total-factor-productivity gain is less than 0.66%, possibly less than 0.53% (NBER WP 32487). Nontrivial. Also not the economic revolution of earnings-call transcripts.
The honest picture of AI productivity in 2026 is jagged. Excellent on some tasks. Net-negative on others. Widely overstated in the aggregate by the people being helped. Worth holding onto when you read the next CEO quote about AI "10X-ing their team."
The AI-resilient sectors nobody talks about
While the white-collar laptop class absorbs the disruption that was supposed to hit truck drivers, the opposite story is playing out elsewhere.
BLS 2024–34 projections:
- Electricians: +9% employment growth. ~81,000 openings a year.
- Claims adjusters: −5.1%.
- Retail trade: single largest job-losing sector.
- Paralegals: roughly flat.
Randstad and CSIS both estimate the US needs 140,000+ more electricians, HVAC technicians and welders by 2030 just to build the AI data-centre infrastructure (Fortune, March 2026; CSIS). Median electrician pay sits around $62,000, top quartile over $100,000, and demand is baked into the hyperscaler capex cycle through at least 2030.
The people literally building the AI boom are in short supply. The people writing React components for that boom are oversupplied.. that's the 2026 market in one sentence.
Geography
AI work is still concentrated. San Francisco holds 32% of global AI-engineering listings, more than the next nine cities combined. New York and London each cross 4,000 open AI postings. Bengaluru's AI-professional population hit 2.35 million, up 55% YoY (Insight Global).
Remote AI work exists. But "AI job" and "remote" don't cluster the way "traditional software engineer" and "remote" did in 2022. Superstar cities still dominate this particular cycle. If you're hunting AI roles and not in one of them, your filter should start with remote-first companies (Ramp, Vercel, Linear, Anthropic's Applied AI team, a long tail of small AI labs) rather than the default LinkedIn sort.
Four things worth doing
A few practical moves fall out of the data.
Put real AI signal on your CV, even in a non-technical role. The 28% Lightcast premium and 56% PwC premium show up across industries, not just engineering. Two AI skills clears the 43% threshold. Stack-level specifics (LangChain, RAG, model names, eval frameworks) beat generic "AI literacy" claims on keyword-matched screeners.
If you're junior, compensate for the automated on-ramp. The tasks your seniors used to give you to cut your teeth are now API calls. Counter-moves: ship visible output (GitHub, case studies, a small newsletter), take contract AI work on Upwork (AI-integration freelance work grew 178% in 2025; AI-video-generation work +329%), bias toward employers that still invest in apprenticeship.
Don't play the volume game. Overall hiring is down 8.7% YoY and the marginal return on application #401 is near zero. Time is better spent on 20 well-targeted applications plus warm intros.
Mid-career and mobile? Seriously look at AI-resilient sectors. Not the ones you hear about. The ones that build data centres, keep hospitals running, fix power grids, tutor kids, run small businesses. Most are structurally undersupplied through 2030.
If you're trying to read the current market against your actual profile (not a LinkedIn-sorted front page), Flint is built for this. It scores every posting across 100+ company boards against your CV on six dimensions, so a Rust engineer and a CFO looking at AI-exposed roles get genuinely different front pages.
Try Flint free. 60 seconds, a CV, your top 10 scored matches.