By Craig Collier, Director, StructAssure Ltd.
Artificial intelligence has arrived on the construction site faster than most safety management systems have been updated to account for it. Cameras that flag a missing hard hat, software that predicts which activities are most likely to produce an incident, tools that draft a method statement in seconds: all of this is now sold to contractors as a way to make sites safer.
Some of it genuinely does. Some of it quietly moves risk around rather than removing it, or creates a new risk of its own. The job of anyone responsible for safety is to tell the difference, and to remember that whatever the technology does, the legal duty to manage the risk still sits with a person.
That is worth saying plainly at the start, because the sales pitch rarely says it. AI is a powerful tool. It is not a safety manager, a competent person or a duty holder, and no procurement decision changes who the law holds responsible when something goes wrong.
What AI is actually doing on site
Strip away the marketing and the current uses of AI fall into a few groups.
- Computer vision.
Cameras linked to trained models watch the site in close to real time and flag conditions as they arise: a worker without a hard hat or harness, someone inside an exclusion zone, a person too close to a moving plant. Instead of an inspection at a single point in time, the site is watched continuously wherever there is a camera. - Wearables and sensors.
Smart helmets, vests and wristbands can track location, detect a fall, and flag signs of fatigue or heat stress, alerting a supervisor before a lapse becomes an injury. - Predictive analytics.
Models trained on incident, inspection and near miss data try to identify which activities, locations or times carry the most risk, so that supervision and resources are aimed where they are most needed. - Document drafting.
Large language models will produce a risk assessment, a method statement or a permit from a short prompt, and summarise long standards or reports in seconds.
Where AI genuinely helps
Used well, these tools do things people cannot. A camera does not get tired, distracted or talked out of raising a concern, and it watches every shift, not only the ones an inspector happens to attend.
Analytics can see a pattern across thousands of records that no individual would spot, such as a particular combination of trade, location and time that keeps producing the same near miss.
Drafting tools can take the mechanical load off a competent person so that time goes on judgement rather than formatting. The common thread is that AI is at its best when it handles volume, repetition and speed, and leaves the decisions to people.
Take a simple, real example. On a busy site, keeping people clear of moving plants is a constant battle, and a reversing excavator is one of the most dangerous moments of the day.
A camera and proximity system can watch that interface continuously, sound a warning the instant someone steps into the danger zone, and log every incursion. That record then tells the site manager where the segregation is failing and where to change the traffic management plan.
The technology has not removed the hazard, and it has not written the plan, but it has given a competent person better information to act on, sooner. That is the right division of labour: the machine watches and reports, the person decides and acts.
Where AI must not replace judgement
The limits matter more than the features, because this is safety.
Accountability does not move.
The Health and Safety at Work etc. Act 1974 and the Construction (Design and Management) Regulations 2015 place duties on people and organisations, not on software. A model can inform a decision; it cannot be the duty holder, and “the system did not flag it” is not a defence.
A model is not a competent person.
AI is trained on past data and performs well on the situations it has seen. The site that does not look like its training data, the unusual sequence, the one off temporary works arrangement, is exactly where it is least reliable and where human competence matters most.
Automation complacency and alert fatigue.
If people come to trust the camera to catch everything, they stop looking. If the system raises too many low value alerts, the real ones get dismissed with the rest. Either way the technology can make a site less safe by dulling the attention of the people on it.
Data quality and bias.
A model is only as good as the data behind it. Gaps, errors or an unrepresentative history produce confident but wrong outputs, and a drafting tool will state a withdrawn standard or an invented clause with exactly the same confidence as a correct one. Everything it produces has to be checked by someone who knows the subject.
Monitoring people.
Cameras and wearables that watch workers carry real obligations under data protection law. There is a difference between monitoring a hazard and monitoring a person, and treating worker data casually damages the trust a safety culture depends on.
None of this is an argument against the technology. It is an argument for keeping it proportionate. HSE’s long standing position on managing risk is that effort should be proportionate to the level of risk, spent where it reduces real harm rather than to be seen to be doing something. That principle does not change because the tool is new.
The drafting trap
The use spreading fastest is also the one that needs the most care: using a large language model to write risk assessments, method statements and permits.
It is quick, and the output reads well. That is exactly the danger. A model produces fluent, confident text whether or not it is right, and it does not know your site. Left unchecked, it will give you a generic document that lists hazards you do not have and misses the one that matters, cites a standard that was withdrawn years ago, or invents a clause number that sounds plausible.
A risk assessment is a record that a competent person has thought about this job; a document generated in ten seconds and signed without that thought is worse than useless, because it looks like control while providing none.
Use these tools to draft and to prompt your thinking, then have someone who knows the work check every line against the real task and current standards before it is issued.
How to responsibly adopt AI in safety
Treat AI as a tool inside the safety management system, not a bolt on beside it. In practice that means:
- Before you buy, decide what human decision the tool supports and who makes it. The output is an input to a competent person, never the final word.
- Validate it on your own sites, and keep checking. Trial it against what your team already knows, and keep a person in the loop on anything that triggers an action.
- Consult the workforce. The people being monitored should understand what is collected, why, and what happens to it. Involvement builds the trust that makes any safety measure work.
- Govern the data. Manage it under data protection law and, increasingly, under the emerging management standard for AI itself, ISO/IEC 42001, alongside your ISO 45001 occupational health and safety arrangements.
- Keep the records. If AI informs a safety decision, be able to show what it recommended, who reviewed it and what was decided, the same as any other part of the system.
The duties that do not change
AI can sharpen a good safety culture. It can watch more, spot more and take routine work off skilled people so they can think. What it cannot do is care about the outcome, take responsibility for it, or exercise the judgement of a competent professional who knows the site. The contractors who get the most from it will be the ones who treat it as a powerful assistant to a well run safety system, and never as a replacement for one.
About the author: Craig Collier
Craig Collier is Director of StructAssure. He’s a chartered construction and quality professional (MCIOB, CQP MCQI, MAPM) with more than twenty years on major UK programmes including Crossrail, the Woodsmith Mine and Transport for Wales. He writes about the practical side of health, safety and quality assurance in construction, and how the records that prove a job was done safely can be made to do real work rather than sit in a file.
References
- HSE. Health and Safety at Work etc. Act 1974. Available from: https://www.hse.gov.uk/legislation/hswa.htm
- HSE. Construction (Design and Management) Regulations 2015. Available from: https://www.hse.gov.uk/construction/cdm/2015/index.htm
- HSE. Managing risks and risk assessment at work. Available from: https://www.hse.gov.uk/simple-health-safety/risk/index.htm
- ISO 45001:2018, Occupational health and safety management systems.
- ISO/IEC 42001:2023, Information technology, Artificial intelligence, Management system.
- Information Commissioner’s Office. Monitoring workers. Available from: https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/employment/monitoring-workers/data-protection-and-monitoring-workers/?search=imbalance&ref=cleanlog.ghost.io
