Scovai Scovai
Talent Intelligence 2026-11-09 1 min read

Domain Expertise Lost Ground on Anthropic's Interview Scorecards. Its Junior Technical Hire Share Went From 27% to 9%.

DSL

Dr. Sarah Liu

Domain Expertise Lost Ground on Anthropic's Interview Scorecards. Its Junior Technical Hire Share Went From 27% to 9%.

The junior share of technical hires at Anthropic fell from 27% to 9%. That number has been covered to exhaustion. The one underneath it has not: an automated coder read the hiring-manager interview scorecards for thousands of employees hired since 2024, and found that among technical candidates "the single biggest increases were in comments on applicants' autonomy and judgment," while "discussion of domain expertise shrank as a share of overall assessment" (Anthropic, 2026).

Most of what we know about AI and entry-level hiring comes from job ads โ€” what employers say they want before they meet anyone. Anthropic's 16 October employment report does something rarer: it reads what interviewers wrote after the conversation.

Two numbers, two different objects. The headcount number is contested, as I will show. The scorecard number is the one your own interview rubric is not ready for.

What the Interview Scorecards Actually Measured

The report is an internal-data study of a single company: AI usage logs, output metrics, employee surveys, Slack and calendar data, hiring decision records, plus 17 employee interviews. Lead author Nathan Wilmers is at MIT Sloan; the work is published by Anthropic itself.

On composition, two distinct figures matter and should not be merged. The share of junior-level employees inside technical departments "slipped from 21% to 14% since early 2025." The share of junior technical hires "fell even faster, from 27% to 9%." The second is a mix shift in who gets hired, not a headcount cut โ€” the report states plainly that technical hiring "has continued growing in absolute terms" while the company "tripled in size in the last year." Anyone reading 27% to 9% as a 65% reduction in junior hiring has misread it.

What managers said tracks the data. Surveyed in April 2026, "a majority of managers described hiring for a different set of skills due to AI," while "around a third of ML and a fifth of SWE managers reported hiring less due to AI." The majority changed criteria. A minority cut volume. That ordering is the whole story.

The scorecard layer

The distinctive evidence is textual. In 2024, hiring managers noted a candidate's track record with AI tooling in "around 5% of assessments for technical roles, and 7% for non-technical roles." By 2026 those rates had tripled for technical roles and quadrupled for non-technical roles. The report gives the 2024 bases and the multipliers in words only, so the endpoints are not published โ€” and inventing them would manufacture precision the source does not have.

Underneath that, the deeper rebalancing. For technical candidates: autonomy and judgment up by the largest margin, domain expertise down as a share of assessment. For non-technical candidates: "customer and mission orientation increased, while discussion of execution skills fell." One manager's explanation is the cleanest summary of the logic in the entire document: "I expect models to continue to be jagged, so for my team I want them to be fungible across domains."

Note what has no number attached. The autonomy-and-judgment increase is reported directionally, as the largest change, with no percentage. That is a limitation of the evidence, not a reason to discount it โ€” but it means this finding supports a direction of travel, not a magnitude.

One Firm Is Not a Benchmark

The report's authors are blunter about this than most commentary has been: "There are many parts of Anthropic's employment practices that cannot be replicated across the economy." The company generates between two and four times as much revenue per employee as Meta and Google did as they crested. It is n = 1, self-published, unaudited, and sitting at the far end of AI intensity. Nothing in it establishes an economy-wide trend.

The wider literature is actively contested on exactly the point most people take from it. Lambert and Schindler's analysis of 243 million new hires and 407 million job postings across the US, UK, Canada and Australia found that a two-standard-deviation increase in either generative-AI exposure or remote-work exposure was each associated with roughly a five-percentage-point fall in junior share by 2025 when tested alone โ€” but when both were entered together, the AI coefficient "attenuates sharply and often becomes statistically indistinguishable from zero" while the remote-work effect stayed robust (Lambert & Schindler, 2026). Their reading: the broken junior ladder may be a story about where work happens, not about what automates it.

Firm-level spending data cuts the other way again. Linking observed AI spend across 21,559 US firms to workforce records, Ramp and Revelio Labs found that high-intensity adopters grew entry-level headcount by about 12% over the two years after adoption, while low-intensity adopters showed no statistically significant change at all (Ramp, 2026). Heavy AI investment, in that dataset, was associated with more junior hiring, not less.

So the junior-collapse causal claim is unsettled, and a Head of Operations has no business planning headcount off it. The criteria claim is a different matter.

What Converges Is the Criteria, Not the Count

Here is why the scorecard finding deserves attention despite n = 1. It agrees with evidence drawn from a completely different data type, at an entirely different scale.

PwC's 2026 Global AI Jobs Barometer analysed over a billion job ads across six continents, with a deep cut on 2.4 million US entry-level roles. AI-exposed junior roles were found to be seven times more likely than the least-exposed junior roles to demand traditionally senior skills โ€” judgment, strategic thinking, motivational leadership, emotional intelligence. Postings fitting that "seniorized" profile grew 35% since 2019, even while overall early-career postings flatlined in highly AI-exposed sectors (PwC, 2026).

Ads record what employers advertise. Scorecards record what interviewers rewarded after meeting the person. Those are independent measurement surfaces, collected by different parties with different incentives, and they moved in the same direction: toward judgment and autonomy, away from narrow execution.

Convergence across independent methods is weaker than a single causal estimate but stronger than two anecdotes. It is enough to act on, because the action is cheap and locally testable.

And the mismatch it exposes is live in most mid-market hiring stacks. The entry-level requisition now asks for senior judgment while the comp band, the 90-day ramp and the mentoring model all still assume a beginner. Meanwhile UK graduate recruiters reported 141 to 172 applications per vacancy in the most oversubscribed sectors (ISE, 2026) โ€” which means the screening layer is under more volume pressure than ever, at precisely the moment the thing worth screening for has changed.

What a Head of Operations Should Do This Quarter

The audit implied here is small enough to run yourself before the next hiring cycle closes.

  1. Pull your last twenty interview scorecards. Not the rubric โ€” the free-text comments your interviewers actually wrote.
  2. Code each line into one of four buckets: narrow domain depth, execution and implementation, autonomy and judgment, and evidence of working effectively with AI tools.
  3. Count. If the first two buckets dominate โ€” and in most 50-to-500-FTE companies they will, because that is what the rubric was built to elicit โ€” you are selecting hardest on the two categories that lost ground at the firm furthest down this road.
  4. Then check the harder thing: whether your process can even observe the other two. Autonomy and judgment do not show up in a CV or a timed technical exercise. They show up in how someone scopes an ambiguous problem, decides what not to do, and explains a tradeoff they got wrong.

That last point is the real constraint, and it is a measurement problem before it is a criteria problem. Degrees and job titles are proxies for domain expertise โ€” the attribute that is losing share. Judgment and autonomy need demonstrated evidence: structured assessment, worked problems, a record of decisions rather than a record of positions. The candidate-side instruments being built for this distinction, Scovai's Talent Passport among them, certify demonstrated capability with a per-skill percentile rank rather than history alone โ€” which is the shape the evidence would have to take to be screenable at volume.

Two cautions before you rewrite anything. Do not use this to justify cutting junior hiring; the causal evidence does not support it, and the firm-level spending data points the opposite way. And do not drop domain expertise from your criteria โ€” "shrank as a share of overall assessment" is a rebalancing, not a removal. In a 200-person operations team with no model-training function, domain knowledge is still the thing your customers are paying for.

The Decision in Front of You

The honest summary of the evidence is this: nobody can yet tell you whether AI will shrink your junior layer. Several credible datasets disagree. But two independent sources โ€” a billion job ads and thousands of interview scorecards โ€” agree on what the surviving roles are being assessed against, and most interview rubrics have not moved in response.

Your interview scorecards are a written record of what your organisation believes predicts performance. Anthropic's are two years of that belief changing in measurable text. Yours are sitting in your ATS right now, unread as a dataset.

Read twenty of them this quarter. If they are still rewarding the candidate who knows the most over the candidate who decides best, you have found this quarter's highest-leverage fix โ€” and it costs an afternoon, not a budget line.

Ready to go beyond the CV?

Scovai's AI-powered Talent Passport reveals what resumes can't: personality, potential, and true job fit.