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 sees and interprets visual and spatial information for inspection, monitoring, and analysis. Heavily clustered at leading-edge: object detection, OCR, and quality inspection have proven deployments but most organisations lack the labelled data or edge infrastructure for production scale. Only one practice reaches good-practice. Most trajectories are stalled, waiting on hardware costs and data pipeline maturity to unlock broader adoption.
The headline: The cameras and scanners now work almost everywhere — the technology is no longer the problem. What stops AI vision from going mainstream is regulation, legal liability, and the hard organizational work of fitting it into real operations.
AI that sees and interprets images — medical scans, faces, store shelves, satellite photos, traffic, factory floors — has largely solved the technical challenge. In trial after trial it matches or beats human experts. Yet across almost every use we track, adoption is stuck below the mainstream. A telling number from this cycle: 90% of major US hospitals have deployed AI to read medical images, but only 19% say it actually works for them. The leaders are pulling ahead in narrow, high-value lanes — wildfire detection, mineral exploration, government ID systems, airport security — where the payoff is clear and the rules are settled. Most organizations are stuck in pilots that never scale, not because the AI fails, but because their data, workflows, and people are not ready for it. This is the most heavily regulated corner of AI, because it touches faces, bodies, and public spaces — so for anything involving people, the law now sets the ceiling, not the engineering.
Turkey banned facial recognition for employee access control outright. Its data-protection regulator ruled the technology unlawful for clocking staff in and out, ordering companies to switch to cards, PINs, or key fobs. It joins a fast-growing list — China, the EU, 20-plus US states — treating face scanning as something you must justify, not just deploy. Any organization using facial recognition for staff or customers should treat "is this even legal here?" as the first question, not the last.
A US court forced police to disclose how their face-matching tools work. New Jersey's Supreme Court ruled prosecutors must reveal the AI tool used, its error rate, and the source photo in criminal cases — the strongest US safeguard yet — even as federal immigration agents pushed a face-recognition app to over 1,220 local police departments. The legal accountability gap around these tools is starting to close, and the liability is shifting toward whoever deploys them.
US regulators fast-tracked AI that writes radiology reports on its own. The FDA gave breakthrough status to two systems that draft medical imaging reports autonomously, a real capability milestone. But a Stanford-Harvard audit found that of 1,200-plus approved medical AI devices, fewer than 15% are actually used routinely — approval and adoption have come apart entirely. The lesson for any buyer: a vendor's regulatory clearance tells you almost nothing about whether the tool will deliver in practice.
Starbucks killed its AI shelf-counting system after nine months across 11,000 stores. The post-mortem blamed governance and change management for 97% of the failure — not the technology, which worked. It is the clearest warning of the cycle that buying the AI is the easy 3%; fixing your data and processes first is the other 97%.
The EU AI Act's workplace-monitoring deadline lands August 2026. From that date, AI that watches employees for safety — hard-hat detection, restricted-zone alerts — is classified as high-risk in Europe, demanding impact assessments, human oversight, and audit logging, with fines up to €35M or 7% of global revenue. If you operate in the EU or employ people there, compliance work needs to be underway now, not in August.
Demographic bias has become a courtroom fact, not a debate. Face-recognition systems still misidentify darker-skinned women up to 40 times more often than light-skinned men, and courts and regulators now treat that as established. Lawsuits are surviving dismissal on exactly this point. Before deploying any face- or body-based AI, demand demographic accuracy data in writing — it is now a liability question.
Sensor hardware is consolidating toward Chinese suppliers while the physics stays unsolved. Over 90% of automotive-grade LiDAR (the laser sensors behind 3D vision) comes from Chinese vendors, and the sensors still fail badly in fog, rain, and darkness. Anyone betting on camera-and-sensor AI for safety-critical use should watch both the supply-chain concentration and the weather-reliability gap before committing at scale.
The technology is the easy part; your organization is the hard part. Across radiology, retail, and satellite analysis, the same pattern holds — roughly 95% of AI-vision pilots never reach production, and the failures trace to data quality, workflow, and change management, not the AI. The capability you can buy; the readiness you have to build.
The law moves faster than the technology here. Because vision AI touches faces, health, and public spaces, regulators are imposing outright bans and prohibitions, not just guidelines. In several markets, no amount of accuracy improvement makes a deployment legal — the constraint is permission, not performance.
"It passed the test" doesn't mean "it works in the field." Approved, benchmark-topping systems repeatedly underperform once deployed — 90% of hospitals run imaging AI but only 19% find it effective; safety cameras flag false alarms 90%-plus of the time in production. Regulatory clearance and lab scores are necessary but nowhere near sufficient evidence of real-world value.
Go deeper: the full Computer Vision & Sensing briefing — the longer analytical write-up, plus every practice we track in this domain with its maturity rating, the tools to consider, and the evidence behind our assessment.