The AI landscape doesn't move in one direction — it lurches. Some techniques leap from experiment to table stakes in a single quarter; others stall against regulatory walls, technical ceilings, or organisational inertia that no amount of hype can dislodge. Knowing which is which is the hard part. The State of Play cuts through the noise with a rigorously maintained index of AI techniques across every major business domain — classified by maturity, evidenced by real-world adoption, and updated daily so you always know where you stand relative to the field. Stop guessing. Start knowing.
A daily newsletter distilling the past two weeks of movement in a domain or two — delivered to your inbox while the index updates in the background.
Each dot marks the weighted maturity of practices within a domain — hover for a brief summary, click for more detail
AI that helps interviewers evaluate candidates consistently by structuring scoring rubrics and flagging evaluation biases. Includes calibration support and rubric enforcement; distinct from resume screening which evaluates documents rather than interview performance.
AI-assisted structured scoring has achieved operational maturity in high-volume hiring but remains trapped behind cascading validity, fairness, and regulatory barriers that are now crystallizing into systemic deployment risk. Multinational enterprises and large-scale recruiters—HireVue (800+ clients), Metaview (3,000+ customers), Curatal, Interviewer.AI—sustain deployments with documented efficiency gains: 27–71% time-to-hire reductions, £3,000/month CV screening savings, and 20-point improvements in final interview pass rates. The underlying methodology is theoretically robust: decades of research validate structured interviews as 2.2× more predictive of job performance than unstructured alternatives (Sackett et al. 2022), and 85% of systems designed with explicit fairness guardrails meet bias thresholds. Yet adoption has stalled and new structural validity threats have emerged in June 2026. Five reinforcing barriers now constrain expansion: (1) LLM self-preference bias—peer-reviewed research shows AI screeners prefer their own stylistic outputs 67–82% of the time regardless of actual quality, a structural property unfixable by prompt engineering; (2) audit methodology failures—vendor bias audits aggregate across jobs, masking job-level disparities that emerge under EEOC scrutiny (Stanford's 3.4M-application study found vendor audits gave "clean" results while job-by-job analysis revealed 26% of Black and 15% of Asian applicants faced adverse impact); (3) GenAI cheating and response gaming—39% of applicants use AI to optimize answers; LLM-scored assessments show 18–23% scoring bias toward AI-generated text; (4) candidate trust collapse at 26% fairness confidence and offer acceptance falling from 74% (2023) to 51% (2026), despite positive user experience in live interactions; (5) regulatory fragmentation—federal vacuum after EEOC guidance removal (Jan 2025) created patchwork of state standards (CA, IL, CO, TX) and EU AI Act high-risk classification (August 2, 2026 enforcement, up to 7% global revenue penalties), with Mobley v. Workday class certification (~1.1B applications) establishing vendor liability precedent. The result is deepening bifurcation: enterprises with compliance infrastructure and high-volume hiring needs sustain deployment despite validity and fairness risks; mid-market and risk-averse organisations remain blocked by unresolved structural threats and implementation costs. The practice is production-grade for compliant enterprise use but not yet enterprise-safe at mass market scale.
The vendor ecosystem continues scaling despite mounting validity and fairness concerns. HireVue serves 800+ enterprise clients—Emirates, Unilever, Philips, Nestlé—reporting $500k to £1M annual savings per deployment. Metaview's 3,000+ customers cite 30-minute-per-interview time savings and 30% reduction in interviews-per-hire. Interviewer.AI reports 66% of hires closing within one week. UK SME adoption jumped to 54% (from 35% in 2025) with documented 71% cost-per-hire reduction and £3,000/month CV screening savings. Production case studies show measurable outcomes: LNER cut hiring from 7 weeks to 3 weeks (71% reduction); William Hill compressed time-to-interview from 15 days to 1.8 days (88% reduction); TTS Talent client reduced first-year turnover 48% and achieved 5% performance uplift through structured assessment. These deployment outcomes remain credible across sectors and geographies, validating operational ROI in controlled settings.
However, June 2026 evidence reveals critical structural validity threats beneath the deployment metrics. Stanford HAI's analysis of 3.4M applications across 156 employers found that while individual vendors had published clean bias audits, job-level disaggregation (required by EEOC) exposed algorithmic adverse impact: 26% of Black and 15% of Asian applicants faced disproportionate rejection—a classic audit methodology failure showing vendor audits aggregate across jobs to mask disparities. LLM-based assessment systems exhibit self-preference bias (67–82% preferring own-model outputs) regardless of actual quality, and this is a structural property unfixable by prompt engineering. Assessments scored by LLMs show 18–23% bias toward AI-generated text, disadvantaging candidates using their own voice. On calibration side, while 69.6% of organizations use structured interviews and 47% conduct calibration, the vast majority (78.7%) retain human final authority—indicating tool adoption without methodology adoption. Candidate transparency gaps compound fairness perceptions: 70% of candidates are never informed upfront that AI is involved, and 38% abandon hiring processes due to AI assessment. Only 26% of candidates trust AI fairness despite positive user experience, and offer acceptance rates remain depressed at 51% (down from 74% in 2023). GenAI cheating remains prevalent: 39% of applicants use AI in responses; 49.6% optimize answers against known LLM assessments; RPO research documents candidates successfully using GenAI to pass video interviews with inflated ratings. Fairness metrics remain vendor-dependent and inconsistent (40% variance between implementations), yet audit frameworks that enforce structured rubrics (vs. black-box embeddings) demonstrate measurable bias reduction.
The regulatory environment has hardened into a compliance crisis. The EEOC removed AI hiring guidance in January 2025, creating a federal vacuum filled by conflicting state standards: California (FEHA), Illinois (HB 3773 prohibiting intent-independent discrimination), Colorado (SB 24-205 with annual impact assessments), Texas (TRAIGA), Ontario (AI disclosure mandate), Germany and UK (EU AI Act and Data Act reformed decision-making). The EU AI Act classifies recruitment as high-risk with mandatory agent inventories, automated logging, human oversight, and transparency—enforcement August 2, 2026 with penalties up to 7% of global annual revenue. Vendor liability is now established: Mobley v. Workday class certification (nationwide, ~1.1B applications) treats vendors as direct agents liable for disparate impact; Kistler v. Eightfold exposes FCRA liability for opaque scoring and missing dispute rights; HireVue faces ACLU/EEOC complaints on accessibility failures. This fragmented compliance burden favours caution and concentrates adoption among enterprises with compliance infrastructure. The practice has achieved operational maturity but faces a compliance and validity infrastructure gap that blocks mass-market expansion.
— U.S. District Judge Rita Lin denies Workday's motion to dismiss, establishing vendor liability as "agent" performing FEHA-regulated activities; first ruling expanding direct vendor liability for algorithmic outcomes.
— Morgan Lewis confirms candidate assessment and selection are Annex III high-risk systems under EU AI Act (enforcement Aug 2, 2026); mandates risk management, data governance, human oversight, technical documentation, and regulatory registration.
— GESI analysis documents proxy bias (names, career gaps, language patterns) and failure of human-in-the-loop safeguard (only 41% detect deliberate bias); provides negative signal on structural limitations and oversight gaps.
— Enterprise AI interview platform with 10M+ interviews conducted, 9.1/10 candidate satisfaction, 1.6k hours saved monthly; demonstrates production-scale maturity with explainable scoring, fairness metrics, and independent audit availability.
— UK ICO audited 30+ employers, issued compliance letters to 16; enforcement establishes test for meaningful human involvement and mandates bias/fairness testing; signals shift from best-practice to regulated-mainstream governance.
— Class action alleges Eightfold scored 1B+ worker profiles 0-5 without disclosure/consent, violating FCRA; establishes FCRA liability theory distinct from discrimination claims, covering opaque scoring without candidate access/dispute rights.
— Named global SaaS company raised offer acceptance by 12% after six months of disciplined calibration; demonstrates specific, measurable business outcome from structured scoring and panel alignment.
— Large-scale empirical study (3.4M applications, 156 employers) found 26% of Black and 15% of Asian applicants faced adverse algorithmic impact; position-level analysis reveals job-by-job disparities invisible in aggregate vendor audits.