Construction safety — the numbers that matter
Construction remains one of the most dangerous industries in the UK. HSE statistics consistently tell the same story: construction accounts for a disproportionate share of workplace fatalities and major injuries relative to its workforce size.
In a typical year, the UK construction industry records 30-50 fatal injuries — roughly a quarter of all workplace deaths despite employing around 6% of the workforce. Non-fatal major injuries (fractures, amputations, serious burns) number in the thousands. And these are the reported incidents — near-misses and minor injuries are significantly under-reported.
The most common causes of fatal injury in construction are:
- Falls from height — consistently the single largest cause, accounting for approximately 50% of construction fatalities
- Being struck by a moving vehicle or object — including crane operations, delivery vehicles, and falling materials
- Being trapped or crushed — collapses, overturning equipment, trenchworks
- Contact with electricity — overhead lines, underground cables, live installations
AI cannot prevent all of these. But it can provide systematic monitoring that catches hazardous conditions before they become incidents, supplements human safety management, and creates a documented record of site conditions.
Falls from height account for approximately 50% of construction fatalities. Which AI application has the most potential to reduce this specific risk?
CDM 2015 duties — where AI supports each role
The Construction (Design and Management) Regulations 2015 establish a framework of duties for managing health and safety on construction projects. AI does not create new duties — it provides tools that help duty holders discharge their existing responsibilities more effectively.
The Client must make suitable arrangements for managing a project, including ensuring sufficient time and resources for each phase. AI-generated progress reports and safety dashboards provide clients with evidence-based information about whether these arrangements are working.
The Principal Designer must plan, manage, and monitor pre-construction phase health and safety, and coordinate design to eliminate or reduce foreseeable risks. AI can assist by analysing design drawings for potential safety hazards — identifying where edge protection will be needed during construction, where temporary works are required, and where confined space work is unavoidable.
The Principal Contractor has the most direct operational safety management duties:
- Plan, manage, and monitor the construction phase
- Ensure the site is secured
- Ensure welfare facilities are provided
- Prepare a construction phase plan
- Organise cooperation between contractors
- Ensure suitable site inductions are provided
AI supports these duties through continuous site monitoring (PPE, exclusion zones, housekeeping), automated safety reporting, and evidence that the principal contractor is actively managing safety — not just producing paperwork.
Contractors must plan, manage, and monitor their own work. AI safety monitoring of a subcontractor's work area provides objective evidence of compliance (or non-compliance) with the site safety rules.
Workers must report any safety concern. AI-assisted near-miss reporting — where a worker takes a photograph and AI categorises the hazard — lowers the barrier to reporting and creates a structured record.
A principal contractor installs AI-powered cameras across the site and receives daily safety reports. A worker then suffers a fall from an unprotected edge that was not in view of any camera. The principal contractor argues they had 'comprehensive AI safety monitoring.' Does this discharge their CDM 2015 duties?
PPE and exclusion zone monitoring in practice
PPE monitoring systems:
The practical deployment of AI for PPE monitoring involves fixed cameras at site access points and work areas, with AI processing the video feed or captured images.
Access point monitoring. Cameras at the site entrance and key internal access points check that every person entering is wearing the required PPE — typically hard hat and hi-vis vest as a minimum. Non-compliant individuals are flagged, and the gateman or site manager is alerted. This catches the most common PPE failure: workers entering site without their helmet, particularly after breaks.
Work area monitoring. Cameras in active work areas monitor ongoing compliance. The AI distinguishes between workers (who must wear PPE) and the background environment. It flags any person detected without the required PPE for that zone. High-risk zones may require additional PPE — harnesses in areas of work at height, face shields near grinding operations.
Accuracy realities:
- Hard hat detection: 92-96% accuracy in good lighting with clear sight lines
- Hi-vis vest detection: 90-95% accuracy
- Harness detection: 75-85% accuracy (harnesses are partially concealed by clothing)
- Eye/ear/hand protection: 60-70% accuracy (too small to reliably detect from typical camera distances)
Exclusion zone monitoring:
Critical areas on construction sites — crane operational zones, deep excavations, areas below lifting operations, zones with temporary propping — should be clear of unauthorised personnel.
AI monitors defined exclusion zones by detecting any human presence within the zone boundary. When a person is detected, an alert is raised — typically a visual alert on a site monitoring screen and a notification to the site safety manager.
The system requires:
- Fixed camera positions with clear views of the exclusion zone
- A defined zone boundary overlaid on the camera view
- Processing fast enough to alert before the person reaches the danger point (ideally within 5-10 seconds)
For crane operations, this is particularly valuable: the AI monitors the exclusion zone beneath the crane's operational radius and alerts if anyone enters while a lift is in progress.
Near-miss detection and permit-to-work verification
Near-miss detection:
Near-misses — hazardous events that could have resulted in injury but did not — are the most valuable leading indicator of safety performance. An increase in near-misses predicts an increase in actual incidents. But near-misses are systematically under-reported on construction sites because reporting takes effort and workers fear being blamed.
AI can help in two ways:
-
Automated detection from camera feeds. AI can identify hazardous conditions (materials stacked unsafely, access routes blocked, missing barriers) and hazardous events (a person stumbling, a load swinging unexpectedly, a vehicle reversing near pedestrians) without a worker needing to fill in a form.
-
Simplified reporting. A worker takes a photograph of the hazard on their phone. The AI categorises the hazard type, severity, and location, and generates a near-miss report that the worker simply confirms and submits. Reducing the reporting effort from a 5-minute form to a 30-second photograph significantly increases reporting rates.
Permit-to-work verification:
High-risk activities on construction sites require permits to work — hot works, confined space entry, electrical work on or near live systems, excavation near existing services. The permit system ensures that all safety precautions have been verified before the work begins.
AI can assist by:
- Reading the active permit register and checking that a permit is in place for the area and activity being conducted
- Processing site photographs to detect hot works (grinding, welding) and checking whether a permit is logged for that area
- Flagging areas where controlled work appears to be in progress but no corresponding permit is registered
- Verifying that permit conditions are met: fire extinguisher visible near hot works, atmosphere monitoring equipment present in confined space work
This does not replace the permit system. It provides a second check — an independent verification that the system is being followed.
A worker photographs a poorly stacked pallet of bricks on their phone and submits it to the AI near-miss system. The AI categorises it as 'materials storage — unstable stack — medium severity.' What happens next in a well-designed system?
COSHH assessments and AI-assisted material hazard review
Construction sites use hundreds of substances — adhesives, sealants, paints, solvents, cement products, insulation materials, timber treatments — many of which are hazardous. The Control of Substances Hazardous to Health Regulations 2002 (COSHH) require employers to assess the risks from hazardous substances and implement controls.
AI can assist COSHH compliance by:
- Cross-referencing materials against safety data sheets (SDS). When a specification calls for a particular product, AI can retrieve the SDS and summarise the key hazards, required PPE, first aid measures, and storage requirements.
- Generating draft COSHH assessments. Based on the product SDS and the proposed method of use, AI can draft a COSHH assessment covering the hazards, exposure routes, control measures, and emergency procedures.
- Flagging substitution opportunities. If a specified product has significant hazards, AI can identify lower-hazard alternatives that meet the same performance specification — supporting the COSHH hierarchy of control (elimination, substitution, engineering controls, administrative controls, PPE).
The QS's draft assessment is exactly that — a draft. It must be reviewed by someone competent to assess the risks in the context of the specific site and work activity. AI does not have the contextual understanding of site conditions, ventilation, proximity to other workers, or exposure duration that a competent person applies.
The ethical dimension — surveillance vs safety
AI-powered monitoring on construction sites raises genuine ethical concerns that must be addressed honestly, not dismissed.
The surveillance concern. Workers may perceive AI cameras as surveillance rather than safety tools. If workers feel they are being watched to catch mistakes rather than to keep them safe, the technology will generate resentment and resistance. Implementation must be transparent about what is monitored, why, and what happens with the data.
What to monitor — and what not to:
- Monitor PPE compliance — this directly relates to worker safety and is a legal requirement under CDM 2015 and the Personal Protective Equipment at Work Regulations
- Monitor exclusion zones — preventing entry to dangerous areas is a clear safety function
- Monitor general site conditions — housekeeping, material storage, access routes
- Do NOT use facial recognition — identifying individual workers by face crosses the line from safety monitoring to personal surveillance. PPE compliance can be monitored without knowing who the non-compliant person is. If individual follow-up is needed, the supervisor addresses it through normal management, not algorithmic identification.
- Do NOT use monitoring data for performance management — timing how long workers take on tasks, monitoring break durations, or tracking individual productivity using safety cameras is a misuse of the technology that will destroy trust.
- Do NOT retain footage beyond the reasonable period needed for safety analysis (typically 30-90 days unless an incident occurred)
Data protection. AI safety monitoring involves processing personal data (images of identifiable individuals). This engages GDPR requirements:
- A lawful basis for processing (legitimate interest in worker safety)
- Transparency (workers must be informed that monitoring is in place)
- Data minimisation (collect only what is needed for the safety purpose)
- Retention limits (do not store footage indefinitely)
- Rights of access (workers can request access to footage of themselves)
Getting buy-in. The most successful implementations involve workers in the design of the system. Consult the safety committee or worker representatives. Explain what the system monitors and does not monitor. Share the safety data it produces. If workers can see that the system is catching hazards that protect them — a missing guard rail, an unstable scaffold, a vehicle in a pedestrian zone — they are more likely to accept it as a safety tool rather than resent it as surveillance.
Your company wants to implement AI safety monitoring. A union representative asks: 'Will this be used to identify and discipline individual workers?' What is the right answer?
Key takeaways
- Construction is disproportionately dangerous — 25% of workplace fatalities despite 6% of the workforce. Falls from height account for approximately half of construction deaths.
- AI supports CDM 2015 duties but does not discharge them. Cameras do not replace guard rails. Monitoring does not replace management.
- PPE monitoring is reliable for large items (hard hats at 92-96%, hi-vis at 90-95%) but less so for small items (gloves, eye protection).
- Near-miss reporting benefits from AI — photograph-based submission with AI categorisation lowers the reporting barrier and increases reporting rates.
- Do not use facial recognition on construction sites. Monitor conditions, not individuals. The ethical boundary is clear and crossing it destroys trust.
- GDPR applies to AI monitoring of construction sites — lawful basis, transparency, data minimisation, retention limits, and access rights must all be addressed.
Module 9 — Final Assessment
Under CDM 2015, what does installing AI safety cameras on a construction site achieve in terms of the principal contractor's duties?
Why should facial recognition NOT be used in construction site AI safety monitoring?
What is the most effective way to increase near-miss reporting on construction sites?
An AI system detects hot works (grinding) in progress on site. No hot works permit is registered in the permit system for that area. What should the system do?
