Operational safety
Human oversight in safety AI
Computer vision can make risk conditions visible earlier. It should support safety teams, not replace their judgement — and it should watch the environment, not judge people.
Most workplace incidents are preceded by warnings nobody recorded: a forklift that passed too close, a pallet left in front of an exit, someone walking under a load. Near misses are the most valuable safety data a site has — and the least collected.
Computer vision can help collect them. How it is designed decides whether it helps or harms.
Watch conditions, not people
The purpose is to detect risk conditions in the environment: vehicles and pedestrians on shared routes, people inside exclusion zones, blocked emergency exits, missing area delimitation, smoke. It is not to measure productivity, rate individuals or build a disciplinary record.
That distinction should be built into the deployment: what is detected, what is stored, who can see it, how long it is kept.
Inform, then let people decide
A good safety system does three things:
- Detects a condition reliably enough to be worth attention.
- Informs the right person at the right time — the area supervisor, the HSE team.
- Records the event so patterns become visible.
Then people act: stop the lift, clear the exit, change the route marking, update the procedure, adjust training. The system raises awareness; the decision stays human.
Deploy with the people it protects
Responsible deployment involves the safety officer, the data protection officer and employee representatives from the start, and is assessed against applicable data protection, workplace and AI regulations at each site.
Prevention works when people trust the system. That trust is earned by design.