People & Talent — Biweekly Brief
The headline: AI is reliable at organizing people data and unreliable at interpreting it. Six in ten managers are using it for raise and layoff decisions anyway.
The Picture
HR has absorbed more AI than any other business function and has less to show for it. Nearly every large employer now screens resumes automatically, yet 88 percent of HR leaders report no significant business value and only 6.6 percent have wired AI end to end through hiring. Asked what blocks them, the most common answer is not bias or privacy — it is plain systems integration. A small group with clean employee data and managers who actually use the tools is getting real returns: Chipotle cut time-to-hire by three quarters, Unilever saved 50,000 recruiter hours a year, Workday's own team recovered 24,000 hours in nine months. Most companies own comparable software and have not moved. If that is you, you are in the pack — but the pack is now accumulating legal exposure while still waiting for the payoff.
This Fortnight
- Three separate studies found AI is good at sorting people data and bad at interpreting it. A peer-reviewed benchmark of seven leading AI models across 84 employee-feedback tasks scored them 64–82 percent on categorizing but as low as 33 percent on synthesis — reading feedback and forming a view. A parallel survey found 60 percent of managers already use AI in raise and layoff decisions, under 20 percent of them trained. Draw the line explicitly: AI can collect, route, and summarize, but whoever decides the outcome should read the source material.
- One in five employers has frozen entry-level hiring because of AI — on almost no evidence. A Gartner survey of chief HR officers found 22 percent froze junior hiring citing AI automation, while only 5 percent report any measurable employment impact from AI. Separate research finds 55 percent of leaders who cut jobs for AI now regret it. Before you cut a graduate intake, ask what measured evidence you have rather than what the market is doing.
- Compensation benchmarking has a governance problem on two fronts. An investigation opened into whether one major pay-benchmarking platform's 9,000-company salary pool lets competitors coordinate wages — an antitrust question, not a fairness one. Separately, 84 percent of compensation professionals run high-stakes pay analysis through general-purpose ChatGPT rather than a compensation-specific tool, with no audit trail. Find out this quarter which of those describes your team.
- Enforcement landed twice, and a clean audit did not save anyone. The US Department of Justice settled with OpenAI and its subsidiary Statsig for $3.2 million over discriminatory hiring practices, and a federal jury awarded $15 million in punitive damages against Nike in a gender pay equity case despite the company's analytical work. Budget for legal review of your pay audit, not just the audit.
- Employee engagement hit a twenty-year low. Gallup's annual benchmark puts global engagement at 20 percent, with manager engagement down nine points since 2022. The same data shows engagement nearly doubles where managers actively support AI use — making manager enablement the highest-return line in your AI budget.
Coming Up
- December 2, 2027 is the EU deadline that requires work now. Disclosure and emotion-detection rules took effect on August 2; the heavy documentation obligations for hiring, pay, and performance AI land in December 2027, with penalties up to 7 percent of global revenue. The prerequisite is slow: inventory every tool touching hiring, pay, or performance, then vendor due diligence, per-decision logging, and bias testing on your own data. Get the inventory into your Q4 budget round.
- Your US measurement baseline is disappearing while liability grows. The Equal Employment Opportunity Commission has proposed scrapping the workforce demographic reports in place since 1966, and public diversity disclosure among large US companies fell 65 percent year on year — even as landmark cases covering 1.1 billion rejected applications expand employer and vendor liability. Decide deliberately whether to keep collecting the data voluntarily rather than losing your baseline by accident.
- Vendors are shipping agentic AI — software that acts on its own without being prompted — into the judgment tasks the evidence says it handles worst. Two dozen distinct agent products launched across the major hiring platforms in the last month, while compensation vendors publicly advise against fully autonomous agents in pay. At your next renewal, require audit rights, per-decision logs, and a written rule that no termination, promotion, or pay change executes without a named human.
What's Hard About This
- "A human reviews it" is not the safeguard regulators think it is. A controlled study of 528 reviewers across 1,526 screening scenarios found they went along with biased AI recommendations up to 90 percent of the time, including when they rated the output as poor. Running the same screening tool twice on identical candidates produced only 14 percent overlap in the shortlist — so there is no stable answer for a human to check against.
- Cheap sourcing has not produced better hiring. The cost of finding candidates has fallen roughly tenfold, but senior hire failure rates remain around 30 percent, unchanged since 2018, because the bottleneck was never finding people. Outreach is saturating too: reply rates on the main professional network fell from 32 percent to 22 percent in a year.
- Managers, not models, are the variance. Only 34 percent of organizations see strong manager participation in the skills and mobility tools they bought, and 46 percent of managers resist internal moves even though internal hires cost 50–60 percent less and stay 60 percent longer.
Go deeper: the full People & Talent briefing — the longer analytical write-up, plus every practice we track in this domain with its maturity rating, the tools to consider, and the evidence behind our assessment.