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 analyses video feeds for people counting, crowd management, and anomaly or incident detection in public and commercial spaces. Includes crowd density estimation and behavioural anomaly alerting; distinct from workplace safety monitoring which targets occupational rather than general security contexts.
AI-powered video analytics for surveillance, crowd management, and incident detection remains a practice in transition: vendor maturity is accelerating, documented real-world ROI exists, but operational barriers persist and are now clearly articulated. The practice uses computer vision and deep learning to monitor feeds in real time—counting people, estimating crowd density, detecting behavioral anomalies, and alerting operators to incidents. Production deployments across 12+ industries (stadiums, hospitals, retail, memory care, data centers) have documented cost savings ($115K+ first-year in stadium deployments) and incident prevention (49-59% crime reduction, 67% crowd-incident reduction at major venues). Vendor ecosystems have matured: Hikvision DeepinView, Cisco Meraki Gen-3, and Axis Object Analytics now offer edge-embedded ML at scale with simultaneous multi-scenario detection (PPE, violence, gathering, intrusion). Market momentum is undeniable: global video surveillance market projected to reach €88B (2031) from €56.1B (2025); 93% of security leaders plan to deploy AI video analytics; 85% of organizations achieving full ROI payback within 12 months across retail, manufacturing, and logistics verticals.
However, two unresolved barriers now define the practice's ceiling. First, the false-alarm crisis: operator alert fatigue remains ubiquitous (83% of alerts in mature security operations are false positives; 74% of companies struggle to achieve value within 90 days of deployment). A critical evaluation gap was identified in 2026: frame-level VAD benchmarks show state-of-the-art AUC-ROC >52% but translate to event-level precision <10%, revealing fundamental misalignment between published metrics and operational surveillance. Academic research on vision language models documents consistent failure patterns: false-positive rates of 31-96% on safe scenes, with models unable to distinguish genuine emergencies from visually similar benign situations (CPR training vs. cardiac arrest). Second, post-deployment governance: while technical performance improves under controlled conditions, integration failures and operator disengagement compound in the 90-day window, driving organizational adoption failure even when algorithmic capability is present.
Regulatory pressure now represents an explicit maturity ceiling, not a distant constraint. The EU AI Act (in force February 2025) prohibits real-time biometric identification, emotion recognition, and biometric categorization systems; France's Council of State (January 2026) blocked the City of Nice's automated vehicle detection system for lacking explicit legislative basis—demonstrating that regulatory gatekeeping now constrains technical deployments. The US regulatory landscape fragments further: 21 states have enacted biometric privacy laws; 16 states mandate school threat detection (Alyssa's Law) but lack unified standards; healthcare facilities face mandatory workplace violence prevention requirements; while NDAA Section 889 restricts hardware from specified Chinese manufacturers. Deployment concentration remains high in occupancy, retail, and event safety; broader public surveillance adoption is stalled by convergence of technical (false-alarm misalignment), operational (90-day adoption cliff), and regulatory (prohibition + fragmented mandates) barriers. This is a practice with proven tactical ROI in specific, well-defined contexts—but without resolution of systemic false-alarm and governance barriers, it remains confined to experimental/early-adoption contexts rather than mainstream public surveillance deployment.
Vendor momentum and real-world ROI accelerate, supported by ecosystem maturity and institutional investment. Hikvision's DeepinView and Guanlan AI claim 90% false-alarm reduction; Cisco Meraki Gen-3 MV cameras integrate cloud dashboard with cross-camera tracking and query-driven investigation (April 2026); Axis Object Analytics supports 16+ simultaneous algorithms. Graymatics (NITI Aayog validated) operates multi-continent deployments across six verticals (smart cities, manufacturing, waste, healthcare, education) with documented cost reductions >25% and 90% accuracy in specialized domains. The Hikvision DeepinMind NVR exemplifies production-ready ecosystem maturity with 16 simultaneous models per engine. Market growth validates institutional confidence: video analytics expanding €56.1B (2025) to €88B (2031, 7.8% CAGR in EU); global market projected $6.41B→$24.18B (2035); 83% of security professionals rate market positively; Motorola acquisition of Blue Eye ($79M, February 2026) signals major institutional entry. Edge AI camera segment growing even faster ($1.8B→$9.2B, 19.2% CAGR) as privacy regulations and latency requirements drive on-device processing adoption.
Real-world deployments now demonstrate both capability and economic validation at enterprise scale. Fortune 500 campus deployment (800 cameras) achieved 90% threat detection accuracy, 85% false-alarm reduction, and incident response time cut from 22 minutes to 4 minutes with ROI within first month. Multi-site transit authority deployment realized 69% crime reduction and 67% assault reduction in first quarter. Transit systems in Singapore and South Korea achieve 76% false-alarm reduction and 94% detection accuracy (Mordor Intelligence market analysis); 47-store grocery deployment validates retail loss-prevention capability (23% shrinkage reduction, ~$880K recovered, 9-month ROI at $30K-$60K per store). IntelliSee documents 13 deployments across 12 industries with $115K+ first-year cost savings. Industry benchmarking (Wavestore 2026) shows 85% of organizations achieve full payback within 12 months; manufacturing ROI 90-95%, retail 30-100%, logistics 100-300% over 12-18 months. Crowd density prediction at MetLife Stadium (67% incident reduction), Singapore Changi, and Coachella achieves 92% accuracy.
Yet the false-alarm barrier crystallizes as the definitive operational ceiling. Skeletal-analysis violence detection achieves 0.98 precision in controlled environments but 0.72 in real surveillance (26-point accuracy loss from oblique angles and high-altitude deployments), validating technical feasibility while exposing production tuning barriers. Vision Language Models exhibit consistent failure: false-positive rates of 31-96% on safe scenes, unable to distinguish genuine emergencies from visually similar but benign situations (CPR training, fireplace video, fire drills). Consumer deployments reveal systemic failures: Wyze, Ring, Blink systems misidentify fire, animals, and interactions, documenting widespread failure across 75M deployed home cameras despite vendor claims. A peer-reviewed evaluation study (April 2026) exposed frame-level vs. event-level misalignment: state-of-the-art models achieve AUC-ROC >52% but event-level precision <10%. Post-deployment governance failures compound technical barriers: 74% of companies struggle to achieve value within 90 days; alert fatigue (83% false positives in mature SOCs) drives operator disengagement.
Regulatory barriers now explicitly constrain deployment scope and have become binding maturity factors. EU AI Act (effective February 2025) prohibits real-time biometric identification, emotion recognition, and biometric categorization; narrow law-enforcement exceptions for targeted victim search and imminent terrorism prevention require prior judicial authorization and are time/location-limited. France's Council of State (January 2026) blocked the City of Nice's automated vehicle detection system, ruling that continuous algorithmic analysis requires explicit legislative basis—a precedent limiting EU public-sector surveillance deployments. US regulatory landscape fragments into fragmented incentives and constraints: 16 states mandate school threat detection (Alyssa's Law); 20+ states require healthcare workplace violence prevention; 21 states have enacted biometric data privacy laws; NDAA Section 889 restricts surveillance hardware from specified Chinese manufacturers. Deployment concentration remains high in occupancy, retail, event safety, and industrial monitoring; mainstream public surveillance adoption stalled by unresolved false-alarm crisis (systemic precision failures, 90%+ baseline false-alarm rates), post-deployment governance gaps (90-day adoption cliff, operator disengagement), and regulatory gatekeeping (EU prohibition + judicial blocking + US fragmentation).
— Enterprise deployment across 800-camera 50-acre campus: 90% threat detection accuracy, 85% false alarm reduction, incident response time cut from 22 minutes to 4 minutes, ROI within first month; validates operational effectiveness in Fortune 500 environments.
— Comprehensive 2026 ROI analysis: 85% of organizations achieve full payback within 12 months; manufacturing 90-95%, retail 30-100%, oil & gas 200-400% ROI; false alarm reduction 90%, investigation time cut from hours to seconds.
— Staqu JARVIS media aggregator documents production deployments: RCB cricket stadium crowd monitoring, UP Police 71-prison state-wide deployment (900km perimeter, 700 cameras), Nagpur Police; demonstrates breadth of adoption across stadium and law enforcement verticals.
— Analyst perspective shows 93% of security leaders mainstream adoption intent; McKinsey reports physical AI applications remain early-stage with full rollout requiring ~10 years; market projects $6.41B (2025) to $24.18B (2035).
— Compliance tracker documenting 137+ mandates across 40 US states and EU driving adoption (Alyssa's Law in 16 states, healthcare WPV in 20+ states) and constraining deployment (21 state biometric privacy laws); shows fragmented regulatory landscape for video analytics.
— European security market analysis projects video surveillance market growth from €56.1B (2025) to €88B (2031, 7.8% CAGR) driven by smart cities, critical infrastructure, and operational decision-support deployments; ecosystem shift from hardware to software/platform economics.
— Insurance industry analysis: 57% of facilities cite legacy systems as top operational barrier; edge AI enables modernization without replacement; global non-guarding security services market projected $117B; edge enables behavioral anomaly detection without full infrastructure rebuild.
— Authoritative reference on 8 prohibited AI practices (real-time biometric identification, emotion recognition, biometric categorization) with legal force since Feb 2, 2025, directly constraining video analytics deployment scope in EU.