👥 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.
The Headline
HR owns more AI than any other function and trusts it least. The tools work; courts, regulators and your own workforce now decide whether you can use them.
The Picture
Almost every large employer has already bought AI for hiring, pay and performance: resume screening runs at 99 percent of the Fortune 500 and 91 percent of managers lean on it for assessments. Yet 88 percent of HR leaders report no significant business value and only 3 percent of US employers use HR AI extensively, so being in the pack means owning tools you neither measure nor govern. A small group is pulling ahead by pairing mature platforms with real governance and a willingness to act: Kmart's structured chat interviews more than doubled First Nations hiring, a 20,000-person health system cut turnover 29 percent, and Workday's own three-year learning rollout lifted completion a quarter. The pack is now exposed rather than merely behind. A screening vendor has been held liable alongside the employer across 1.1 billion rejected applications, Germany's courts have made companies own every word their HR chatbot says, and standard commercial insurance stopped covering AI losses in January. Buying more capability will not move you; deciding what you will let the software do, and proving a person actually checked, will.
This Fortnight
The strongest trial yet shows AI-conducted structured interviews improved hiring outcomes across 70,884 applicants at a Teleperformance subsidiary. Job starts rose 18 percent and early retention by a similar margin with no productivity loss, but a companion study found AI scorers reward candidate embellishment unless the evaluation criteria are disclosed in advance. The method is validated; whether you tell candidates what is scored, and whether a person reads the transcript rather than approving the number, is what separates a defensible deployment from a lawsuit.
Two vendors shipped real-time pay benchmarking at scale while new survey data showed 49 percent of employers run no pay equity review at all. Payscale's Ascent platform reached general availability (out of beta) on 10 million employee records and Compa raised $35 million, yet of employers that do review pay, only one in six has ever adjusted anyone's salary after finding a gap, and Apollo Global's analysis suggests benchmarking data is being used to manage pay pressure rather than match the market. Better data is not your constraint; deciding in advance what you will do when it shows a gap is.
Gartner now forecasts that at least one in three roles eliminated for AI will be refilled by 2029 at higher cost. A large study of US firms found heavy AI adopters actually grew headcount around 10 percent, and Ford, IBM and Klarna have all rehired within months of AI-justified cuts. If your workforce plan models AI as a headcount reduction, it is modeling the scenario the evidence says fails; redeployment and capability planning are the harder, better-supported case.
Gallup's four-wave panel shows frequent AI use raises displacement fear and lowers engagement regardless of how much support the organization provides. Manager quality buffers the effect by 11 points, and Meta suspended its employee-monitoring program this month after a staff petition against it. Sentiment dashboards will not fix what they measure; investing in managers, and not surveilling people to train models, will.
A pre-registered audit of 40,726 AI screening requests found that audit design, not model bias, drives most of the demographic disparity measured. Ranking versus rating, and whether the model knew it was being tested, moved results more than the applicants' demographics did. The bias audits New York, Colorado and Illinois now require may certify very little; treat a passed audit as a compliance receipt, not evidence of fairness.
Coming Up
Colorado's automated-decision rules take effect in January 2027 and require human review independent of both the AI and the original decision-maker. Every framework now in force, from the EU AI Act to the federal Office of Personnel Management's hiring guidance, rests on the same person-checks-the-machine assumption, and research shows recruiters mirror the AI 90 percent of the time. Inventory every HR tool that scores, ranks or predicts people, and define what a reviewer must actually examine before a regulator asks you to prove it.
Standard commercial liability policies stopped covering AI-related losses in January, and vendor contracts are not filling the gap. A German court has held a company strictly liable for its HR chatbot's hallucinations (when an AI tool confidently makes things up), Mobley v. Workday is heading toward a nationwide trial, and 88 percent of AI vendor contracts cap the vendor's liability. Ask your broker for affirmative AI coverage this quarter and renegotiate indemnities at renewal; until then the exposure is yours, not the vendor's.
The US is dismantling the demographic data infrastructure that pay-equity and diversity analytics have relied on since 1966, while Europe is expanding it. The EEOC has voted to rescind EEO-1 reporting and the DOJ's $21.5 million Deloitte settlement makes goal-linked demographic management a False Claims Act exposure for federal contractors, while the EU Pay Transparency Directive brings reporting deadlines in June 2027. Decide now, with counsel, whether you keep collecting the data privately for European compliance and litigation defense, because a gap in the series cannot be backfilled later.
What's Hard About This
The human oversight the law demands does not work the way most companies practice it. Recruiters given an AI score follow it 90 percent of the time, and UPS hired its seasonal workforce in seven minutes per person with no human examination at all. When review is a rubber stamp, courts are treating the vendor as the real decision-maker, and the liability travels with the decision to whoever made it.
You are paying for outcomes you cannot yet count. Workday alone has passed $600 million in AI annual recurring revenue, much of it from HR buyers, while most HR organizations do not formally measure whether any of it works. Until you define the outcome metric before purchase, the vendor's revenue is the only number that will be real.
Measurement has outrun the will to act on it, and no dashboard fixes that. Attrition models forecast departures accurately, 57 percent of employers run engagement surveys, and pay-equity tools find gaps, yet the feedback rarely changes anything and most employers that find a pay gap never adjust pay. The missing investment is manager conversations, incentives and a pre-agreed remediation budget, none of which a vendor sells.
Practices in this Domain (17)
| PRACTICE | TIER | TREND |
|---|---|---|
| Absence & attrition pattern analysis | GOOD PRACTICE | — Steady |
| Candidate assessment — structured scoring support | LEADING EDGE | — Steady |
| Candidate sourcing & outreach automation | ESTABLISHED | — Steady |
| Compensation & benefits benchmarking | LEADING EDGE | — Steady |
| DEI analytics & reporting | LEADING EDGE | ↓ Declining |
| Employee engagement & sentiment analytics | GOOD PRACTICE | — Steady |
| Employee onboarding automation | GOOD PRACTICE | — Steady |
| HR policy Q&A chatbot | GOOD PRACTICE | — Steady |
| L&D adaptive assessment & knowledge testing | GOOD PRACTICE | — Steady |
| L&D content & personalised learning | GOOD PRACTICE | — Steady |
| Organisational network analysis | LEADING EDGE | — Steady |
| Performance review generation & 360 synthesis | LEADING EDGE | — Steady |
| Recruitment content & compensation modelling | GOOD PRACTICE | — Steady |
| Resume screening & candidate matching | ESTABLISHED | — Steady |
| Skills mapping & career development | GOOD PRACTICE | — Steady |
| Workforce planning & demand forecasting | LEADING EDGE | — Steady |
| Workplace analytics & space utilisation | GOOD PRACTICE | — Steady |
Read the full technical briefing (1,905 words) →
Where AI Stands in People & Talent
People and talent is the domain where AI adoption is most complete and least trusted. Resume screening runs at 98.8% of the Fortune 500 and 83% of employers; 87% of organisations use AI somewhere in recruiting; 91% of surveyed managers lean on it for performance assessment; Workday's Sana platform has passed $600M in AI annual recurring revenue and ServiceNow's AI contract value has crossed $1B, much of it HR and employee service. By the measure of tools bought and switched on, this is a mature market. By the measure of value realised, it is not. Gartner and SHRM put the share of HR leaders reporting no significant business value from AI at 88%; ManpowerGroup and Everest Group found that of the 90% of large employers who have deployed hiring AI, fewer than 5% describe the outcome as transformational; Gallagher's survey of 3,700 US employers found only 3% using AI extensively in HR at all. The pattern repeats in every corner of the domain: purchase is universal, integration is shallow, and outcomes are unmeasured.
What distinguishes this domain from its neighbours is why it has stalled. In most fields the binding constraint is capability or data plumbing. Here the tools work well enough and the barrier is that the people being measured, hired, paid and reviewed have rights, and the courts and regulators have started to enforce them. Mobley v. Workday, now a nationwide collective action covering 1.1 billion rejected applications, has established that a screening vendor can be liable as the employer's agent. Kistler v. Eightfold has extended that liability into consumer-protection law by treating undisclosed candidate scores as credit reports. Germany's OLG Hamm has ruled that a company owns every statement its HR chatbot makes, hallucinated or not. The EU AI Act's emotion-recognition ban has moved from statute to enforcement, and its Annex III high-risk regime covers recruitment, assessment, attrition prediction and performance evaluation. Most damaging of all, the field's favourite safeguard has failed its own audit: peer-reviewed work at FAccT found recruiters mirror biased AI recommendations 90% of the time, which means "human in the loop" as usually practised is a rubber stamp rather than a control.
The exceptions prove the rule. Where organisations have paired mature platforms with real governance and a willingness to act, results are strong: Sapia's Kmart deployment lifted First Nations hiring from 3.2% to 8.25%, a 20,000-employee healthcare system cut turnover 29% in a matched-control study, Workday's own three-year Sana Learn rollout across 35 countries lifted completion 25% and cut content maintenance 70%. But these remain a handful of names against a backdrop of quiet retreat. DEI analytics is the one area where the momentum is actively reversing rather than merely stalled: the EEOC has voted to rescind the EEO-1 reporting framework that anchored the practice since 1966, Fortune 500 public disclosure has fallen 65% year on year, and the DOJ's $21.5M settlement with Deloitte has made demographic goal-linked management a False Claims Act exposure for federal contractors. Employee sentiment, meanwhile, is deteriorating in the background of every deployment: Gallup has global engagement at a 20-year low of 20%, and Glassdoor's review corpus shows positive mentions of AI at work falling from 81% in 2019 to 43% by mid-2026. The domain's tools have never been more capable, and the workforce has never been less convinced they are on its side.
What's New, 2026-09-08 to 2026-09-22
A fortnight is the normal cycle, so nothing here was researched over a longer period. The most consequential evidence lands on structured interviewing, where a randomised trial across 70,884 applicants at a Teleperformance RPO subsidiary found AI-conducted structured interviews raised offer rates 12%, job starts 18% and early retention 18–19% with no productivity loss, the strongest causal evidence yet that the methodology, not just the efficiency, holds up at scale. It arrives alongside a University of Georgia study showing AI scorers reward candidate embellishment that human raters penalise, unless the evaluation criteria are disclosed in advance, and the Office of Personnel Management's 27 August guidance requiring federal hiring AI to be subject to independent human review that examines the underlying record rather than approving a score. Taken together they sharpen the same point: the tool is validated, and governance discipline is what separates a fair deployment from a lawsuit. On the candidate side the trust numbers hardened further, with 70% of job seekers never told AI would assess them and up to four in ten withdrawing when they find out.
Compensation had a busy window. Payscale shipped Ascent to general availability, an AI-first benchmarking platform drawing on 10M+ HRIS-reported incumbents across 4,500 organisations, and Compa's $35M Series B underwrote real-time benchmarking across 42 countries. Against that vendor momentum, two findings complicate the ROI story: Apollo Global Management's analysis that real-wage growth for AI-exposed workers has run 6.7 percentage points slower since 2023, raising the uncomfortable possibility that benchmarking data manages pay pressure rather than matching the market, and a Payscale survey finding 49% of employers run no pay equity review at all and only 16% of those that do have ever adjusted anyone's pay. Workforce planning saw Gartner put a number on layoff remorse, forecasting that at least one in three positions eliminated for AI will be refilled by 2029 at higher cost, while a 21,559-firm analysis showed heavy AI adopters grew headcount around 10%; Workday's Decision Intelligence platform reached GA in the same window, with TX Group cutting a rolling forecast from three days to four hours. Elsewhere, Gallup's four-wave panel quantified the trust mechanism behind falling engagement, showing frequent AI use raises displacement fear and lowers engagement by 0.20 standard deviations regardless of organisational support, with manager quality buffering the effect by 11 points; Meta suspended its employee-monitoring programme after 1,800 staff petitioned against it; 15Five shipped AI-assisted reviews with guardrails that bar the model from setting ratings, while Lattice's survey found 74% of managers already draft with AI and 60% harbour ethical concerns about doing so. A pre-registered 40,726-request audit study found that audit design, not model bias, drives most measured demographic disparity in LLM screening, which undermines the validity of the bias audits several state laws now mandate.
Key Tensions
Human oversight is a legal requirement that does not work as practised. Every regulatory framework now in force, from Colorado's ADMT Act to the EU AI Act to OPM's federal guidance, rests on a human reviewing the machine's output, yet FAccT research shows recruiters mirror biased AI recommendations 90% of the time and cut deliberation by 55% when a score is in front of them. UPS hired 125,000 seasonal workers in seven minutes each with no human examination at all. The Stanford Law Review's functional-control framework and the Mobley ruling both point the same way: when review is a rubber stamp, the vendor becomes the decision-maker and inherits the liability.
Vendors are monetising what customers cannot yet measure. Workday reports $600M in AI ARR and 5,500 customers on organic agents growing 35% a quarter, ServiceNow's AI contract value has passed $1B and its agentic deployments rose ninefold in nine months. On the buyer side, 88% of HR leaders report no significant business value, 56% do not formally measure AI success at all, and Gallagher finds only 3% of employers using HR AI extensively. The revenue is real; the returns are still being estimated rather than counted.
Liability has inverted from employer to vendor and out of the insurance market. Mobley v. Workday treats the screening vendor as the employer's agent across 1.1 billion applications, Kistler v. Eightfold recasts undisclosed candidate scores as consumer reports under FCRA, and OLG Hamm makes companies strictly liable for chatbot hallucinations regardless of training data. ISO's January 2026 exclusionary endorsements have stripped AI-related losses from standard commercial liability cover, so an HR chatbot's six-figure policy error is now uninsured unless the buyer has bought affirmative AI coverage. Compliance cost is rising faster than fairness is improving.
Measurement has outrun the will to act on it. Attrition models forecast departures at AUC 0.85 and vendors themselves warn that prediction without action is an expensive alarm; 49% of employers run no pay equity review and only 16% of those that do ever adjust pay; 57% of employers survey engagement but feedback rarely changes anything; only 19% of facilities leaders base space decisions mostly on data despite 94% integrating automated sources. Korn Ferry's phrase for skills mapping, "Skills Everywhere, Decisions Nowhere", describes the whole domain. The dashboards are installed; the interventions, incentives and manager conversations that would make them worth having mostly are not.
The headcount-reduction thesis behind workforce planning is collapsing. Gartner now forecasts at least one in three roles eliminated for AI will be restored by 2029 at higher cost, 55% of leaders who cut for AI say they regret it, and Ford, IBM, Commonwealth Bank and Klarna have rehired within months. A 21,559-firm analysis finds heavy AI adopters actually grew headcount around 10%, while Apollo's data shows AI-exposed workers' real wages lagging by 6.7 points, suggesting the savings are coming from pay compression rather than fewer people. Planning tools built to model reduction scenarios are being asked to model augmentation, redeployment and capability debt instead, and most organisations have no strategy for that.
Top 10 Evidence Items
- Voice AI in Firms: A Natural Field Experiment on Automated Job Interviews (70,884 applicants, RCT) (research-paper) — This is the strongest causal evidence cited in the fortnight's roundup that structured AI interviewing itself works at scale, not just efficiently. https://arxiv.org/html/2607.28222v2
- Layoff remorse: Gartner says at least one in three positions eliminated by AI will be restored by 2029–at a higher cost (news-coverage) — Directly underpins the collapsing headcount-reduction thesis tension, quantifying the cost of premature AI-driven cuts. https://www.computerworld.com/article/4220413/layoff-remorse-gartner-says-at-least-one-in-three-positions-eliminated-by-ai-will-be-restored-by-2029-at-a-higher-cost.html
- AI Adoption and Employment Growth: Evidence from Enterprise Spending Data (research-paper) — Contradicts the displacement narrative with firm-level data, reinforcing that heavy AI adopters are growing headcount rather than shrinking it. https://www.innovativehumancapital.com/article/ai-adoption-and-employment-growth-evidence-from-enterprise-spending-data
- Using AI More Does Not Reassure Workers, Managers Do (adoption-metric) — Supplies the causal mechanism behind the domain's falling trust and engagement numbers cited throughout the briefing. https://www.gallup.com/workplace/713231/ai-not-reassure-workers-managers-do.aspx
- Embellishment in AI-Scored Interviews: When Deception Goes Undetected (University of Georgia, peer-reviewed) (research-paper) — Shows the RCT's positive result is conditional on governance discipline, sharpening the piece's central point that process, not capability, is the binding constraint. https://news.uga.edu/interviewing-with-ai/
- Apollo Global Management: AI-Exposed Workers See 6.7pp Wage Growth Decline (news-coverage) — Complicates the vendor ROI story by suggesting benchmarking tools may be suppressing pay rather than matching the market. https://www.cnbc.com/2026/09/13/ai-jobs-pay-inflation.html
- Audit design, not model bias, drives measured LLM demographic disparities: pre-registered 40,726-request study across five models shows format and ordering effects exceed demographic effects (research-paper) — Undermines the validity of the bias-audit regime that several state laws now mandate, a direct hit on the domain's favourite safeguard. https://commonplace.workforcefutures.net/paper/arxiv:2609.09048
- We Need Standards for AI in Hiring: Divergent Vendor Approaches and Governance Gaps (opinion) — Illustrates the human-in-the-loop failure mode with a concrete example, UPS hiring 125,000 workers with zero human review. https://www.ere.net/articles/we-need-standards-for-ai-in-hiring
- Meta suspends employee monitoring programme after 1,800 workers sign petition (news-coverage) — A high-profile instance of workforce backlash against surveillance-style AI, evidence for the deteriorating employee sentiment thread. https://www.hrkatha.com/news/meta-suspends-employee-data-programme-after-1800-workers-oppose-computer-monitoring/
- Candidate Backlash Against AI Interview Scoring (CNBC independent journalism) (news-coverage) — Grounds the candidate-side trust statistics in independent reporting rather than vendor-sourced surveys. https://www.cnbc.com/2026/09/15/job-seekers-refusing-ai-interviews-blacklisting-employers.html