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 monitoring of workplaces for safety compliance including PPE wearing, exclusion zone violations, and unsafe behaviour. Includes hard hat detection and restricted area monitoring; distinct from construction site monitoring which tracks progress as well as safety.
Computer vision for workplace safety monitoring works. Forward-leaning manufacturers, logistics operators, and construction firms are running it in production and documenting 30-90% reductions in hazards and incidents. The technology -- real-time detection of PPE violations, exclusion zone breaches, and unsafe behavior via existing camera infrastructure -- has matured past proof-of-concept into a vendor ecosystem with demonstrated ROI and edge-to-cloud deployment options. Yet the practice remains leading-edge, not mainstream. Regulatory fragmentation across privacy, biometric, and labor law creates a compliance burden that deters all but the most motivated adopters. Employee resistance runs deep: surveys consistently find a majority of workers view continuous monitoring as unethical or psychologically harmful. The defining tension is stark -- the safety economics are compelling, but the legal and organizational friction required to deploy responsibly keeps adoption confined to early-mover sectors. Until privacy frameworks stabilize and worker trust models mature, this practice will continue to deliver strong results for those who can navigate the constraints while remaining out of reach for most.
Production deployments continue accelerating across sectors and geographies, with evidence of both sustained ROI and unresolved adoption barriers. Intenseye remains the scale leader, protecting over 100,000 workers across 25+ countries with Sentinel edge hardware achieving sub-second machinery intervention at industrial sites like Oldcastle APG. Swire Coca-Cola's multi-site deployment across the U.S. and Asia cut its Lost Day Rate by 27%. Fortune 500 adoption demonstrates compelling unit economics: Americold's 500K+ sq ft facility achieved 77% injury reduction and eliminated 100% lost-time days ($1.1M EBITDA savings) within 12 months; NSG Group expanded from a single PPE monitoring pilot to 20+ global facilities after seeing 62% violation reduction; ServiceCenter Metals achieved airbag safety compliance improvement from 25% to over 90% within two months. These are sustained operations, not pilots. The vendor ecosystem has stratified by architecture: Intenseye leads on scale and Sentinel hardware; viAct offers 50+ modular detection modules with 62% forklift near-miss reduction documented in production (Saudi Arabia); DeepX deploys on-premises processing with model drift detection for construction/mining; AWS Rekognition Workplace Safety now includes custom label detection and named enterprise deployments; Chooch runs hazard detection on existing cameras without sensors. Ferrovial and DroneDeploy rolled out agentic AI frameworks in Oct 2025, deploying 30+ Safety AI agents across construction sites using existing camera feeds for PPE detection.
Government adoption is accelerating policy-driven deployment. Singapore's Ministry of Manpower published official WSH 2028 Strategy guidance in March 2026 endorsing video analytics as core safety infrastructure across all industries, with mandated deployment on construction sites valued ≥$5M. South Korea expanded allowable AI safety equipment budget allocation from 10% to 20% of occupational safety spending in construction works, signaling policy-driven adoption in aging-workforce sectors. Economic drivers remain compelling: estimated 36.42 trillion won in annual accident losses and SAPA penalties up to 1 billion won creating strong ROI justification for deployment.
The regulatory and technical barriers have simultaneously hardened, creating a compliance ceiling that blocks mainstream expansion. The EU AI Act's August 2, 2026 enforcement deadline classifies workplace monitoring as high-risk, requiring mandatory Fundamental Rights Impact Assessment, bias testing, human oversight, automated logging (6+ months), and transparency disclosures before deployment — penalties of €7.5M to €35M or 1.5% to 7% global turnover for non-compliance, with extraterritorial scope affecting all global employees. Multi-jurisdictional regulatory frameworks are tightening: NSW Australia formally extended WHS employer duties to AI and algorithmic management systems (Feb 2026); Turkey's data protection authority permits occupational safety as a legitimate camera-use purpose but requires strict proportionality and data minimization. A critical technical adoption barrier persists: Vision Language Model detection systems exhibit "overreaction" problems where models detect individual danger cues (smoke, flames, person lying down) without contextual reasoning, producing high false-positive rates. Traditional monolithic detection pipelines run false-positive rates as high as 98%; modular design reduces false positives by order of magnitude but remains incomplete at scale. Detection maturity varies sharply by hazard category: PPE accuracy ranges 92-96% under normal lighting but drops significantly for complex equipment (harness/fall detection) and degraded conditions (night/low-light). Infrastructure constraints compound the problem: 90% of high-hazard construction zones lack full-time dedicated power, and dangerous worker behaviors often occur in 1-5 seconds, exceeding practical response times even with sub-second detection. This technical-operational gap generates alert fatigue that costs the North American security industry over $4.5B annually and causes 74% of enterprises to struggle post-deployment, with 80-95% alert volumes being false positives and 83% analyst misclassification rates. Worker sentiment remains a drag: 71% of employees view monitoring as unethical. Architectural evolution is underway: vendors are shifting from rule-based object detection toward LLM/VLM semantic understanding (e.g., "worker near unguarded edge without harness = imminent fall risk") enabling context-aware reasoning and site-adaptation without retraining. Successful deployments increasingly depend on trust-building measures; a Viso case study documented 54% near-miss reduction only after phased rollout with union endorsement. The practice sits at a regulatory and organizational inflection point: the technology and ROI are proven, government mandates are emerging, detection accuracy for core hazards (PPE, restricted zones, proximity) is validated at 92-96% under normal conditions, and 76% of EHS professionals believe AI reduces administrative burden with 34% reporting measurable ROI; however, regulatory complexity is accelerating, technical limitations in generalized hazard detection persist, infrastructure constraints bind, and psychological resistance constrains adoption velocity. The PPE detection analytics market reached $1.2B in 2024 with projected 19.7% CAGR to $5.8B by 2033, but converting that market signal into mainstream labor force adoption requires solving the regulatory, operational, and organizational problems that technology improvements alone cannot address.
— Technical guide documenting peer-reviewed detection performance (92.11% precision, 0.95 recall, 90% accuracy) from 2024 industrial study; phased deployment strategy and integration with ERP/EHS platforms.
— NSW legislation (Feb 2026) extends employer WHS duty to AI, algorithmic management, and automated monitoring; establishes formal legal framework for workplace safety video analytics deployments.
— EU AI Act compliance framework for workplace monitoring: prohibits emotion inference and real-time facial ID; classifies hazard detection as high-risk requiring FRIA, human oversight, 6+ month logging; €35M penalties.
— VelocityEHS 2026 survey (1,008 EHS professionals, 10M user base): 76% believe AI reduces administrative burden, 70% trust AI insights, 34% report measurable ROI from AI tools; indicates practical adoption momentum.
— Technical evolution from rule-based object detection to LLM/VLM semantic understanding for safety (e.g., 'worker near unguarded edge without harness = imminent fall risk'); documents architectural maturity shift in video analytics.
— Curated thought leadership featuring 10 safety experts including 2026 ASSP Safety Professional of the Year on industry shift from lagging to leading indicators via real-time risk visibility and video analytics adoption.
— Technical research on critical VLM limitation: false-positive overreaction when models detect danger cues (smoke, flames, person lying down) without contextual reasoning; introduces VERI benchmark for contrastive testing.
— Turkish regulatory framework establishing OHS as legitimate camera-use purpose while prohibiting productivity-only surveillance; sets proportionality and data-minimization principles for occupational safety deployments.