UK & European Health and Safety Round-up: HSE Fatality Figures, Enforcement, and the AI Act Countdown

The week in UK and European health and safety

Two threads run through this week's headlines for anyone accountable for health and safety in heavy industry. The first is the HSE's annual statistics, which show long-term risk reduction has stalled and the human cost remains stubbornly high. The second is regulatory: the EU AI Act's high-risk provisions are now weeks from applying, and they name worker monitoring and safety-component AI directly. If you sit in a SHEQ or operations role in construction, manufacturing or logistics, both threads land on your desk. Here is the detail, and where active safety technology fits without displacing the controls you already run.

HSE statistics: the numbers behind the toolbox talk

The HSE's latest annual figures on work-related fatalities and its 2024/25 health and safety statistics tell the same story: a plateau rather than steady improvement. The recurring mechanisms behind the worst outcomes have not changed. Falls from height, vehicle-pedestrian conflicts and machinery incidents still dominate. For a warehouse or manufacturing site, the vehicle-pedestrian pattern is the one that reads closest to daily reality, and it is worth taking to your safety committee against the HSE's published data rather than relying on generic messaging.

A forklift crossing a walkway on the hundredth hour of a shift, a reversing manoeuvre near a blind corner, a person stepping outside a marked exclusion zone: these are the moments that pass without incident thousands of times and then, once, do not. Yellow lines, banksmen, mirrors and training all reduce that risk. None of them acts in the instant a person and a machine converge.

Enforcement this period: notices and repeat breaches

The HSE's prosecutions coverage and its register of convictions and notices continue to show where duty holders are falling short. The pattern is repetition rather than novelty. The register is where regulators publish the gap between a written control and its reality on the ground, and it is a useful benchmark for any site director building a board paper.

• •The HSE register of convictions and notices records prosecutions and enforcement notices as they are issued.
• •Recurring themes across the register include falls from height, vehicle and plant incidents, and repeat breaches of enforcement notices.
• •The duty extends to the quality of services a duty holder commissions, not just its own direct operations.
• •Enforcement outcomes are the practical measure of what a regulator considers an unacceptable control gap.

The common thread is repetition. These are not novel hazards; they are known risks that controls failed to contain on the day. For a site director building a board paper, the enforcement register is a useful benchmark: it shows what regulators consider an unacceptable gap between a written control and its reality on the ground.

The EU AI Act: safety AI is now named as high-risk

The regulatory story of the period is the EU AI Act. Its high-risk provisions become applicable on 2 August 2026, and they classify certain AI used for worker monitoring, task allocation and safety components as high-risk. That classification carries obligations: documented risk assessments, human oversight, transparency, and incident reporting for deployers. Prohibited practices and AI literacy duties came into force earlier in the Act's phased timeline, ahead of the high-risk provisions.

There has been meaningful timeline relief. Following provisional EU agreement on targeted amendments, high-risk rules for employment decisions, including performance monitoring and task allocation, are being delayed to December 2027. The same package of amendments defers certain sector obligations and simplifies parts of the framework. UK operators are not exempt in practice: sites supplying into, or operating across, the EU need to plan for these rules, and must in any case satisfy UK GDPR, UKCA marking and ALARP principles at home.

Where vision-based active safety fits

Vision-based active safety is not a new category of control sitting outside the hierarchy; it is an enhancement to it. It is most useful where a persistent risk (vehicle-pedestrian conflict, PPE non-compliance, exclusion-zone breach) survives every conventional control a site can apply. That framing keeps the technology inside the established risk-management approach rather than presenting it as a replacement for it.

This is the design point behind SAIFI Edge. It combines real-time computer vision with physical interventions: closing a barrier, dropping a gate, projecting a warning or triggering an alarm. The latency threshold is the line between active and advisory. A system that flags a hazard for a supervisor to notice is helpful, but a system that acts inside the reaction window is what turns a near miss into a non-event. SAIFI's offline-first edge architecture keeps that response running independent of site connectivity, with encrypted flows and audit trails that support the human-oversight and traceability requirements the AI Act sets out.

The practical route in is modular. A phased rollout can start with one or two cameras on the highest-risk crossing and expand across zones without replacing hardware, which keeps upfront cost proportionate to proven benefit. Configuration by zone, rule and schedule tailors detection to how a site actually runs, rather than applying generic detection that generates noise. The point is not to tear up your existing controls; it is to add the layer that catches what banksmen, mirrors and training cannot catch every single time.