How Vision AI Is Changing Health and Safety on UK Industrial Sites

From reactive reporting to real-time monitoring

Vision AI is moving health and safety away from incident forms filled in after the event and towards continuous, auditable monitoring of the site as it operates. Systems now detect PPE compliance, exclusion-zone breaches, vehicle-pedestrian conflicts and other site hazards in real time, adding a monitoring layer that reinforces existing controls rather than replacing them. For UK installers, specifiers and end-clients, the question is no longer whether the technology works, but how to deploy it in a way that satisfies the Health and Safety Executive (HSE), aligns with data protection law and stands up to audit.

The evidence behind the shift is becoming harder to ignore. Protex AI reports an 80% reduction in workplace incidents at Marks & Spencer. That is a vendor-reported figure from a single programme, so it warrants scrutiny, but the direction of travel across warehouse and logistics settings is consistent.

• 80%

  Reduction in workplace incidents at M&S, per the Protex AI case study

• 1974

  Year of the Health and Safety at Work etc. Act, the basis for HSE's AI approach

The UK regulatory position

The HSE applies the goal-setting principles of the Health and Safety at Work etc. Act 1974 to AI that affects workplace safety. In practice that means a vision AI system is treated like any other hazard control: you carry out a risk assessment and put in place proportionate controls, including cyber risks, 'so far as is reasonably practicable'. The HSE is developing its regulatory approach to AI and contributing to wider standards work rather than introducing a single, separate AI regime.

Sector-specific rules apply directly and are reinforced, not replaced, by vision AI. CDM 2015 governs PPE and zone management on construction sites; COMAH and DSEAR cover process safety in chemicals and explosive atmospheres; ATEX applies in hazardous areas; and GMP controls apply in pharmaceutical environments. In each case, a system that produces timestamped, tamper-evident event logs gives both HSE inspectors and internal compliance teams audit-ready evidence without retaining continuous video.

Where the EU AI Act fits for UK readers

UK organisations that supply into the EU, or operate sites there, need to track the EU AI Act. Under Article 6, vision AI used as a safety component of a product covered by Annex I legislation, or listed in Annex III, is classified as high-risk. That triggers obligations around risk management, data governance, human oversight, conformity assessment (often third-party) and CE marking. The European Commission's draft guidelines on high-risk AI systems address how safety components are defined and classified. Even where the Act does not bite directly, UK vendors increasingly benchmark their designs against its high-risk requirements to smooth cross-border deployment.

The standards that matter

A cluster of standards now frames responsible deployment. They are worth knowing by name when you specify or tender for a system.

• •ISO/IEC 42001: an AI management system framework covering risk, governance and continual improvement.
• •Functional safety standards for safety-related control systems and general principles for machinery.
• •Machine-vision camera and sensor performance standards for the underlying imaging hardware.
• •Occupational health and safety management standards, where continuous monitoring and reporting support accreditation.
• •UKCA and CE marking for the underlying hardware, plus UK GDPR and Data Protection Act alignment.

Edge versus cloud: the practical trade-offs

For safety-critical detection, latency and connectivity matter. Edge-first systems process video on dedicated, calibrated hardware at the site, triggering barriers or alerts locally and continuing to function during a connectivity outage. The Security Industry Association recommends hybrid edge approaches for latency-sensitive safety applications and notes that installation complexity favours professional integrators for mission-critical setups. The trade-off is higher upfront hardware cost and more customisation effort, offset by reduced cloud dependency, lower bandwidth needs and stronger privacy posture.

Dynamic Safety's SAiFI Edge platform reflects this approach. It provides real-time vision AI for PPE monitoring, vehicle-pedestrian segregation and hazard detection, integrates with existing gates, barriers, access control and PLCs without a rip-and-replace project, and logs events for HSE audits without retaining continuous video. Site-specific model training improves accuracy for local hazards, PPE rules and zoning, though it adds data collection and calibration steps for the installer up front.

What this means for specifiers and installers

• ✓Treat the AI as a control to be risk-assessed under HSWA 1974, not a bolt-on, and document it 'so far as is reasonably practicable'.
• ✓Confirm whether EU AI Act high-risk obligations apply to your supply chain or sites.
• ✓Specify log-based, tamper-evident evidence rather than continuous video to satisfy audits and UK GDPR.
• ✓Check UKCA/CE marking on the hardware and ask how the vendor benchmarks against ISO/IEC 42001.
• ✓Choose edge or hybrid for latency-sensitive safety triggers; reserve cloud for non-critical analytics.
• ✓For mission-critical environments, use professional integrators and plan for site-specific model training.

The picture is one of a maturing practice focused on measurable safety gains and data-driven compliance rather than enforcement actions. In researching this article we found no verified HSE prosecutions, product recalls or major incidents involving vision AI safety systems reported in the sources reviewed. That makes now the right moment to get the foundations right.