AI Computer Vision in Health and Safety: The Shift From Watching to Acting

Active safety, not another camera on the wall

Most cameras on an industrial site are passive. They capture footage that someone reviews after a near miss, a claim, or an investigation. The record is useful, but it arrives too late to change the outcome. AI computer vision is now being used the other way round: watching the live scene, recognising a hazard as it forms, and doing something about it before the incident lands. That distinction, between evidence after the fact and intervention in the moment, is the whole point of the category, and it is the part that changes how a SHEQ director should think about the business case.

The technology itself is not novel to the regulator. In its published approach to artificial intelligence, the Health and Safety Executive treats AI in the workplace as a risk to be managed under existing goal-setting law rather than a special case requiring new powers. That is a helpful starting point, because it means the questions a board will ask (is the risk assessed, are the controls proportionate, is there human oversight) are questions the existing framework already answers.

What the regulators and standards actually require

In the UK, systems that detect hazards in real time sit under the Health and Safety at Work etc. Act 1974. That means a documented risk assessment and controls applied so far as is reasonably practicable, with the HSE acting as market surveillance authority for workplace machinery and equipment. The HSE has set out that it will manage AI risks through the existing framework rather than through new powers. In July 2025 the Construction Plant-hire Association's plant safety group published good-practice guidance for machine-mounted human form recognition systems, which is the clearest UK signpost yet for what a defensible deployment looks like.

For any operation with an EU footprint, two instruments matter. The AI Act (Regulation (EU) 2024/1689) prohibits emotion recognition in the workplace except for medical or safety purposes, and classes AI safety components in critical infrastructure and products as high-risk, with high-risk obligations phasing in on a staged timetable. The Machinery Regulation (EU) 2023/1230, which replaces the old Machinery Directive, requires state-of-the-art safety where AI is integrated into a machine's safety function. The technical standards underneath are worth naming in a board paper because they show the ground is settling, not shifting.

• •ISO/IEC TR 5469:2024 sets out properties, risks and processes for AI used in safety-related functions.
• •ISO/TR 22100-5:2021 addresses how artificial intelligence and machine learning affect machinery risk assessment.

What the deployment evidence shows

The published results are consistent enough to build a case on, provided you read them as directional rather than guaranteed. Reported reductions in unsafe behaviour cluster in a wide band because sites, cameras and tuning differ, but the direction is the same across independent deployments.

• 42%

  Reduction in PPE deviations reported by Alliad using viAct AI-powered PPE detection

• £433,000

  Fine after a worker suffered a skull fracture in a recent UK prosecution

• No missed actions

  Result reported by Ultimo from a trial of its AI safety agent

Two things are notable in the record. First, deployment often follows a conventional failure: in New Zealand, a company introduced AI-enabled CCTV to flag hazards after a chemical burn incident. The lesson is cheaper to learn before the enforcement notice than after it. Second, the enforcement picture in the UK still centres on traditional causes. Recent UK prosecutions, including a £433,000 penalty after a worker's skull fracture and a machine-entanglement case in Shetland, turned on conventional physical hazards, not vision technology. That is the gap active safety is built to close: the moment a person and a machine occupy the same space and nothing intervenes.

Computer vision or smart PPE? Read the trade-offs

Specifiers usually weigh site-wide computer vision against wearable smart PPE. They are not rivals so much as different tools, and many mature sites run both: vision for broad compliance, wearables for a handful of high-risk roles.

Industry guidance from the CPA's plant safety group recommends a risk assessment per machine type, human-form detection validated for accuracy with false-positive controls, defined inner and outer danger zones, operator training, and clear data ownership. Crucially, it positions the technology as a safety aid, not a replacement for physical segregation. That framing matches how SAIFI is designed to work: as the last line that catches what a banksman, a yellow line or a mirror inevitably misses on the hundredth hour of a shift.

Where SAIFI Edge fits

SAIFI Edge runs the detection on-premises using existing ONVIF-compatible IP cameras plus an edge appliance, so video is not streamed to the cloud and personal data stays on site, which eases the privacy-by-design and GDPR questions an auditor will raise. It recognises the events that dominate incident logs (a forklift entering a pedestrian aisle, a missing hard hat, an exclusion-zone breach, a process deviation) and, unlike a review-later camera, it triggers a physical response in real time: a barrier drops, a gate closes, a projected warning appears on the floor, an alarm sounds, or an access action fires. It sits alongside your existing controls and supervision rather than replacing anyone.

• ✓Run a per-machine or per-zone risk assessment before fitment, in line with CPA good practice.
• ✓Validate human-form detection accuracy against your own conditions, not a vendor's demo footage.
• ✓Define inner and outer danger zones with escalating, zoned alerts to avoid alert fatigue.
• ✓Start in mixed mode with human oversight, and keep a record of false negatives and positives.
• ✓Confirm edge processing, minimal retention and a clear data-ownership position for GDPR.
• ✓Check the vendor's cybersecurity posture and data-handling controls.