The State of Play

A living index of AI adoption across industries — where established practice meets the bleeding edge
UPDATED DAILY
← All domains

👥 People & Talent

AI for hiring, developing, engaging, and managing workforce. Skews mature: resume screening is established, and most practices — candidate sourcing, skills assessment, workforce planning — sit at good-practice. Learning and development is advancing. Bias and fairness concerns constrain adoption in hiring; most trajectories are stalled as organisations balance efficiency gains against regulatory scrutiny.

17 practices: 2 established, 9 good practice, 6 leading edge

The Headline

AI is now routine in hiring and pay, but the latest research shows its scores rewarding the wrong signals. The real advantage lies in governing those decisions, not in buying more tools.

17
practices tracked
6
at the frontier
1
moving this fortnight
3,600
evidence items

The Picture

Most employers already use AI to screen applicants, draft job ads and onboarding material, and answer policy questions. The big HR suites are buying up the specialist vendors: Workday reportedly paid $1.1 billion for Paradox, SAP agreed to acquire the skills-inference vendor TechWolf, and Pearson has moved to buy Workera. Nobody has yet shown that any of this produces better hires. Greenhouse's benchmark data put applications per job up 111% between 2022 and 2025, while SHRM's median cost per non-executive hire fell from about $1,600 in 2017 to $1,200 in 2025. A small group of employers is pulling ahead by treating AI outputs as evidence a person must examine rather than as decisions. Everyone else keeps the legal liability for every score their vendor produces, whether or not they know how it was reached.

This Fortnight

  • An unreviewed Harvard Kennedy School and MIT Sloan preprint found that all four AI models tested advised men to seek higher salaries than otherwise identical women. Across 86,000 negotiation-coaching queries, the gaps ran as high as $15,915. The models also steered parental leave towards women and equity towards men. A pay-equity check that looks only at base salary will not catch that second pattern, so widen yours to cover benefits and offer terms before AI-drafted offers spread.

  • A University of Washington study presented on October 25 found that general-purpose AI models scored Chinese- and Nigerian-accented English lowest, even when every speaker read the same script. The researchers did not test commercial hiring products. But a separate audit found that AI evaluators prefer résumés written by particular AI models, regardless of how good the candidate is. If you score spoken interviews or résumés with AI, ask your vendor for results broken down by subgroup, not averages.

  • Production data from the recruiting firm LatamCent showed its AI's top résumé grade barely predicted who passed the next stage. Résumés rated "Strong Yes" were shortlisted after interview 61% of the time, against 63% for a plain "Yes". On that evidence, the grade decides who gets a call, not who succeeds, so set cut-off scores with that in mind.

  • In a beqom survey of 202 US compensation decision-makers, 51% said approved pay decisions are often reopened or corrected. AI users reported this more often than non-users. The likely cause is AI exposing disconnected pay data rather than fixing it. Before expanding AI in pay, fix the data feeding it and keep a record of why each decision was made.

  • Merck gave about 5,000 employees Valence's AI coaching tool ahead of year-end performance reviews. This is review season's most visible deployment, and it arrives as a Nielsen survey for Monday.com finds more than half of senior leaders say AI-assisted work is harder to evaluate. Tell staff where AI shaped their review. Most employers currently don't.

Coming Up

  • LinkedIn begins rolling out Hiring Assistant 2 this month, and SAP's onboarding agents go live on November 15. Both are agents (software that acts on its own without being prompted), and both arrive mainly with vendor claims. LinkedIn says recruiters review 83% fewer profiles. Before switching either on, record your own baseline for time to hire, shortlist-to-offer rate and early attrition.

  • EU employers with 250 or more staff must file their first gender pay gap report under the Pay Transparency Directive by June 7, 2027. The directive also requires salary ranges to be disclosed to applicants and bans questions about salary history. So far, no named HR AI vendor has announced product changes in response. Get written commitments from your recruiting and compensation vendors now, not during the reporting crunch.

  • Mobley v. Workday goes to a class-certification hearing in March 2027, and the EU's high-risk obligations for recruitment AI have been pushed to December 2027. Employment lawyers quoted by SHRM argue that a match score which fast-tracks a candidate to interview already counts as a hiring decision, and the liability for it sits with the employer. Map every point where an AI score advances or rejects someone without a person reviewing it.

What's Hard About This

  • Faster screening has been absorbed by more applications. Applications per job more than doubled between 2022 and 2025, while SHRM's median cost per non-executive hire fell by about a quarter between 2017 and 2025. AI-written applications are part of the reason. Buying a faster tool buys a bigger funnel unless you also redesign the funnel itself.

  • Human sign-off is the norm, but it is often only on paper. Some AI interview products let a pass mark reject candidates with no recruiter involved, and one vendor found hiring teams commented on only 1.6% of finished AI interviews. A reviewer who never opens the underlying record does not reduce your liability.

  • The data under the tools is weaker than the tools. Visa fills about 35% of its openings internally, yet calls its skills-gap analysis "really manual". In a Workday survey of 3,200 employees who use AI, 89% said fewer than half of the roles at their organization had been updated to reflect it. A smarter agent working from stale role profiles produces better-presented versions of the same errors.

2012View this domain's timeline →Today

Practices in this Domain (17)

PRACTICETIERTREND
Absence & attrition pattern analysisGOOD PRACTICE— Steady
Candidate assessment — structured scoring supportLEADING EDGE— Steady
Candidate sourcing & outreach automationESTABLISHED— Steady
Compensation & benefits benchmarkingLEADING EDGE— Steady
DEI analytics & reportingLEADING EDGE↓ Declining
Employee engagement & sentiment analyticsGOOD PRACTICE— Steady
Employee onboarding automationGOOD PRACTICE— Steady
HR policy Q&A chatbotGOOD PRACTICE— Steady
L&D adaptive assessment & knowledge testingGOOD PRACTICE— Steady
L&D content & personalised learningGOOD PRACTICE— Steady
Organisational network analysisLEADING EDGE— Steady
Performance review generation & 360 synthesisLEADING EDGE— Steady
Recruitment content & compensation modellingGOOD PRACTICE— Steady
Resume screening & candidate matchingESTABLISHED— Steady
Skills mapping & career developmentGOOD PRACTICE— Steady
Workforce planning & demand forecastingLEADING EDGE— Steady
Workplace analytics & space utilisationGOOD PRACTICE— Steady
Read the full technical briefing (1,533 words) →

Where AI Stands in People & Talent

People & Talent has the widest AI deployment of any domain and the least proof that it helps. Hiring leads. Resume Genius found that 58% of US hiring managers use AI to screen résumés and 44% use it to match candidates to roles. Applicant tracking systems are close to universal in the Fortune 500. LinkedIn says more than 20,000 companies use its agent-based recruiting products. The economics have not followed. L.E.K. Consulting studied 31 mid-cap US technology firms and found applications per open role up 111% between 2022 and 2025. Median cost per non-executive hire has stayed at $1,300 since 2017. Speed gains are being absorbed by volume, and evidence that AI-screened hires perform better is thin: the most-cited field result was co-authored by the founder of the vendor it evaluated, and it did not measure job performance.

Momentum is strongest where AI drafts and weakest where it judges. Drafting covers onboarding journeys, course content, job adverts, review narratives and answers to policy questions. The large suites are absorbing the specialists. Workday bought Paradox for a reported $1.1bn, SAP agreed to acquire the skills-inference vendor TechWolf, and Pearson has moved to buy Workera. Agent features are now shipping from Workday, SAP, Oracle, ServiceNow and Docebo. Judging means anything that scores a named person: flight-risk models, interview ratings, network maps, pay recommendations. Progress there is slow, and the cause is data and trust rather than the models. Korn Ferry's survey of 1,600 senior leaders found that only 34% trust their organisation's talent insights. Visa fills about 35% of its openings internally, yet its talent-acquisition chief calls its skills-gap analysis "really manual". In a Workday survey of 3,200 employees who use AI, 89% said fewer than half the roles at their organisation had been updated to reflect it.

What sets the domain apart is that almost every output is a decision about an identifiable person, and the employer keeps the liability whatever the vendor's part in it. The litigation covers both discrimination law and consumer-protection law: Mobley v. Workday heads to a class-certification hearing in March 2027, and Kistler v. Eightfold is a credit-reporting class action. State rules in California, Colorado and Illinois now cover pay as well as hiring. The EU's high-risk obligations for recruitment have been pushed to December 2027 but have not gone away. In practice, then, a human signs off on most decisions. In Singapore, 85% of employers using AI in HR still require one, according to the national employers' federation. The domain's main unsolved problem is governance, not capability: making that sign-off mean something.

What's New, 2026-10-20 to 2026-11-03

The fortnight's evidence went mostly against AI scoring. A University of Washington study, presented at EMNLP on 25 October, found that general-purpose models from Google, OpenAI and Alibaba ranked Chinese- and Nigerian-accented English lowest, even when every speaker read the same script. The researchers did not test commercial hiring products. An unreviewed preprint from Harvard Kennedy School and MIT Sloan examined 86,000 negotiation-coaching queries. All four models tested gave men higher salary targets than otherwise identical women, with gaps of $1,594 to $15,915, and also steered parental leave towards women and equity towards men. An audit of 11 résumé-writing and 9 evaluator models found that evaluators favour particular writer models regardless of candidate quality. Production data from LatamCent showed AI "Strong Yes" résumés shortlisted after interview 61% of the time against 63% for plain "Yes". On the pay side, a beqom survey of 202 US compensation decision-makers found that 79% use AI and 51% often see approved pay decisions reopened. AI users report this more often.

Disclosure was the second theme. A Highwire survey fielded in August found that 78% of managers had used AI for feedback or reviews, yet only 16% of individual contributors were told. A Nielsen survey for Monday.com found that more than half of senior leaders find AI-assisted work harder to evaluate. Greenhouse found that at least four in five job seekers had been rejected in the past year without speaking to anyone at the employer. Gallup's third-quarter reading put US engagement at 29%, the lowest since 2011. Vendors kept shipping all the same:

  • LinkedIn begins rolling out Hiring Assistant 2 this month.
  • Visier put a driver diagnostic for resignation spikes into early access, with general availability not due until winter 2027.
  • Merck gave about 5,000 staff Valence's coaching tool ahead of year-end reviews.
  • Oracle: a partner reported a telecoms deployment that halved onboarding effort per hire, from four hours to two.
  • Pigment raised $250m at a $2.3bn valuation, and L'Oréal has selected it (not yet gone live) for salary planning covering more than 95,000 staff.

The sobering counterpoints came from operators, not vendors. HGS's CHRO said predictive attrition models missed engaged staff who left for career moves. Only 4% of Irish HR leaders surveyed by Ibec expect AI to cut headcount significantly.

Key Tensions

  • Deployment has outrun proof of better hiring. Recruiters work faster but are no more confident: in TestGorilla's survey of 1,428 talent professionals, 72% find candidates faster, yet a third of them report no gain in confidence and 18% report a loss. LatamCent's data shows résumé grades decide who gets a call, not who passes it. Without evidence on the quality of hires, faster sourcing and screening is buying a larger funnel rather than better staff.

  • Scoring models reward the wrong signals. This cycle's research found models penalising accent, favouring résumés by their writing model, and pricing work by gender and age: a peer-reviewed study of eight models found age premiums of up to 45.97% that survived corrective prompts. In each case the bias sits in features a rubric does not mention, so audits that pool results or test only base salary miss it. Pay-equity checks that look only at base salary will not catch benefits steered by gender.

  • Employees are rarely told what is being measured. Only 16% of staff knew AI shaped their review, and in Gallup's second-quarter data only 19% at AI-adopting firms had been told what workplace data is collected. A Censuswide poll found that 75% of UK workers did not know they could challenge an automated rejection. California's new curbs on emotion-reading workplace AI followed a Kaiser nurse finding sentiment scores on her recorded calls, so the gap is now attracting legislation.

  • Human sign-off is the norm but often nominal. Vendors and advisers say AI should be "the analyst, not the judge", and most employers require approval. Yet some interview products let a pass mark reject candidates with no recruiter involved, and one vendor found hiring teams commented on only 1.6% of finished AI interviews. Employment lawyers quoted by SHRM argue that a match score which fast-tracks interviews already counts as a hiring decision, with liability resting on the employer.

  • Suites absorb the tools, but the data underneath stays weak. Consolidation is putting skills inference, sourcing and planning agents inside HCM suites. The foundations they depend on remain weak: in a July survey Gartner found only 19% of companies plan their future workforce effectively, and Adobe found 63% of new starters re-entering the same personal data. A smarter agent working from unmaintained role profiles or patchy goal records produces better-presented versions of the same errors.

Top 10 Evidence Items

  1. AI steers women toward caregiving benefits, lower salaries: study (news-coverage) — Direct evidence for the tension that scoring models reward wrong signals, here steering pay and benefits by gender. https://www.hrreporter.com/focus-areas/automation-ai/ai-steers-women-toward-caregiving-benefits-lower-salaries-study/395073
  2. Accent bias: AI models penalize certain speakers in interviews and tests, UW study finds (news-coverage) — Independent accent-bias finding anchors the fortnight's run of evidence against AI judging candidates. https://www.geekwire.com/2026/accent-bias-ai-models-penalize-certain-speakers-in-interviews-and-tests-uw-study-finds/
  3. AI Resume Scores Barely Predict Interview Success (case-study) — Production data showing AI résumé scores barely predict interview success undercuts the case that screening gains translate into better hiring. https://latamcent.com/research-insights/ai-candidate-screening-authority/
  4. AI evaluators favor job applications written by certain AI models, independent of candidate quality (research-paper) — Shows bias sits in features no rubric names, exactly the audit blind-spot the briefing flags. https://arxiv.org/html/2610.41866
  5. When AI sets wages: biases and labor discrimination in generative pricing (research-paper) — Peer-reviewed age-premium result that survives corrective prompts is the strongest single data point for biased pricing. https://link.springer.com/article/10.1007/s43681-026-01472-x
  6. Ghostwritten Feedback: Managers' Hidden Use of AI in Performance Conversations, and How Leaders Can Protect Trust and Development (opinion) — The 78%-use/16%-told gap is the clearest evidence for the disclosure theme and nominal human sign-off. https://www.innovativehumancapital.com/article/ghostwritten-feedback-managers-hidden-use-of-ai-in-performance-conversations-and-how-leaders-can
  7. AI in hiring leaves most job seekers rejected without a conversation (news-coverage) — Greenhouse's four-in-five-rejected-without-contact figure makes the governance gap concrete for job seekers. https://www.hcamag.com/nz/news/general/ai-in-hiring-leaves-most-job-seekers-rejected-without-a-conversation/594138
  8. Why AI's bigger return will come from redesigning jobs (news-coverage) — Visa's 'really manual' admission shows capability has outrun the data foundations suites now sit on. https://sg.finance.yahoo.com/news/why-ais-bigger-return-come-210500768.html
  9. Talent leaders flag a blind spot in HR Analytics (news-coverage) — A named CHRO's own model failing on engaged leavers is the operator-side counterpoint the briefing contrasts with vendor claims. https://hr.economictimes.indiatimes.com/news/trends/hr-analytics/talent-leaders-flag-a-blind-spot-in-hr-analytics/135268936
  10. Can AI recruiting make money if it replaces recruiters? (opinion) — The Paradox and LinkedIn figures show consolidation economics absorbing specialists without proof of better outcomes. https://newmarketpitch.com/blogs/news/ai-recruiting-business-model