Anthropic has spent years publishing research about how AI changes other people’s jobs. On the weekend it turned the analysis on itself, releasing an internal employment study built on staffing records, calendar telemetry, hiring scorecards and interviews with 17 employees. The findings are the most detailed first-person account yet of what happens to a workforce when agents arrive at scale — and they land somewhere more complicated than either the panic or the reassurance.
🔍 THE BOTTOM LINE
The share of junior staff among new technical hires at Anthropic fell from 27 percent to 9 percent since early 2025. By March 2026, AI agent hours company-wide had surpassed human work hours for the first time; by August, the agents directed by individual employees were working six times the hours of the humans commanding them. Code commits per engineer are running at five times their mid-2026-earlier velocity, and hiring managers now assess candidates on autonomy and judgment over narrow domain expertise. But the same report shows heads still growing — sales, finance and recruiting expanded faster than engineering — and describes an early-warning system funneling displaced staff into new internal roles. Anthropic’s economists caution the study is one company at the extreme of adoption, its economists putting it roughly a year and a half ahead of the rest of the economy. It is an early indicator, not a verdict.
Junior hiring fell fastest where agents work hardest
The decline was not uniform. Within Anthropic, discussion of shrinking demand for junior technical staff surfaced in January 2026 in the teams leaning hardest on Claude for their own work — with one manager noting the performance-engineering uplift “basically overnight changed what we were looking for.” By April, engineering and research teams had de-emphasised hiring at lower technical bands; a third of machine-learning managers and a fifth of software-engineering managers reported hiring less because of AI. Among all technical staff, the junior share slid from 21 percent to 14 percent, with entry-level postings falling sharply since December 2025.
The report is careful about what this is: a case study of a single company that uses its own agents more intensively than almost anyone. It also cuts against the crudest version of the displacement story — this is a company that has tripled in size in a year while junior shares fell. The absolute count of technical hires kept rising. What changed is the mix.
That detail matters for readers of earlier coverage. The company’s CEO forecast half of entry-level white-collar jobs gone within five years (Anthropic’s own five-year entry-level forecast), and our earlier read of the evidence found no systematic unemployment rise in AI-exposed occupations. Anthropic’s own data now shows what the adjustment actually looks like inside one firm — not a layoff wave, but a hiring-pattern shift.
The bottleneck flipped from doing to deciding
The productivity numbers are striking. Engineers produce five times as many commits as six months earlier, and the report finds no quality penalty — comparing commits written by agents versus humans in the same directory, line-revert rates were no different. Lawyers process more contracts, accountants more invoices, salespeople close more deals. Sales staff who adopted agents reported Claude absorbing about 20 percent of their former workload, letting them see 30 percent more customers.
But the constraint moved. Interactive Claude sessions sit idle roughly 80 percent of the time — waiting for a human to answer, approve or steer. Employees outside engineering would need more than half a work week just to read everything their agents produce. So they built filters: written agent instructions grew from a few hundred to 8,500 in a year, and “meta-work” — building visibility tools, monitoring pipelines, reviewing other agents’ output — jumped from 10 percent of interactive sessions to 22 percent. Meeting time is up 20 percent. In an April survey, one in five employees described their role as shifted from doing work to managing Claudes.
For career planning, that inversion is the real signal: in the most AI-saturated workplace on record, human value concentrated in judgment, coordination and taste — not in the implementation tasks that got automated first.
The counterweights: redeployment and new metrics
Anthropic also did things most employers have not. It stood up an early-warning system tracking job families whose demand and productivity metrics are shifting, then prioritised moving affected staff internally. Cross-departmental moves among staff of a year or more ran 50 percent higher each month in 2026 than in 2025. Writers whose product copy was scaled out by a context graph rotated into writing model evals and prompts — “using language to drive Claude behavior.” The report argues internal reallocation beats external hiring during fast capability shifts, and that companies committing to it tend to steer AI toward new tasks rather than pure headcount substitution.
The limits get honest treatment too. Hiring-pattern changes do nothing for people never hired — protection for incumbents doesn’t reach people locked out of first roles. Anthropic’s economists call for public retraining and income-support policy to sit alongside private adjustment, a theme the broader evidence keeps reinforcing: Stanford-led work found significantly reduced hiring of young workers in AI-exposed occupations like software development and customer service, while a Danish study found no entry-level gap between AI-using and non-using firms — and Stanford’s Bharat Chandar concluded AI’s overall employment effect is “likely small” so far.
What it means for NZ job seekers
The junior-share collapse is the story with teeth for New Zealand’s 14.8 percent youth underutilisation — NZIER recently described young job seekers as locked out at “crisis level.” If an AI lab that tripled headcount hires juniors at a third of its earlier rate, the adjustment pattern is heading toward every sector that adopts agents seriously — including the grad roles and entry-level pipelines NZ employers are already thinning. The Mercor benchmark map analysis made the same point from the measurement side: the roles getting automated first are the ones with the dashboards watching them.
The optimistic read is Anthropic’s own staffing mix: a company can shrink its junior share while growing total employment, expanding sales and finance and recruiting — because the productivity gains get spent on more humans doing different things, not just margin. Workers’ practical playbook comes straight from the report. Hiring-manager scorecards there tripled mentions of creative AI tool use for technical roles and quadrupled them for non-technical ones — demonstrated judgment with AI tools now outweighs domain credentials. The meta-work layer is genuinely new labour: people who write specs, review agent output and keep pipelines honest are doing a job that did not exist two years ago. And the strongest protection named in the report is organisational, not technical: employers committed to redeployment kept people through the disruption. For NZ employers — mostly small firms without Anthropic’s margins — that is the one finding sitting closest to a policy lever.
FAQ
What did Anthropic’s employment report find? That between early 2025 and 2026, junior technical hires fell from 27 percent to 9 percent of new hires, agent hours surpassed human hours by March 2026 and reached a 6-to-1 ratio by August, and hiring managers now prioritise autonomy and judgment over domain expertise.
Is Anthropic hiring fewer people overall? No. Technical hiring continued growing in absolute terms while the junior share fell, and sales, finance and recruiting grew faster than engineering. The technical share of the workforce slipped from 61 percent to 56 percent.
Does this predict the same changes for other companies? Not directly. The report stresses it is a single-company case study of an unusually intensive AI user — its economists describe Anthropic as being roughly 18 months ahead of the broader economy in adoption.
What should entry-level job seekers do? The report’s hiring data suggests demonstrated, creative use of AI tools and evidence of judgment weigh more heavily than narrow domain execution — consistent with Stanford findings that AI-exposed entry roles are contracting first.
— CJ Murden, editor of Singularity.Kiwi. Former digital technologies teacher, author of AI-focused books. Writing with a New Zealand focus.
