Every safety manager who has run a behaviour-based safety programme knows the ritual of the observation card. A trained observer walks the floor, clipboard in hand, watching for safe and at-risk acts. He notes a worker reaching across a running conveyor, has a quiet word, ticks a box, and moves on. It is a good system, built on a genuinely sound idea: that most injuries are preceded not by faulty machines but by human behaviour, and that catching the behaviour catches the injury before it happens.
The idea is right. The trouble has always been the arithmetic.
The maths that quietly defeats BBS
Behaviour-based safety rests on a simple premise from Herbert Heinrich and later Frank Bird: beneath every serious injury lies a broad base of minor incidents, and beneath those, a vast foundation of unsafe acts that harmed no one — this time. Reduce the unsafe acts at the bottom of the pyramid and, in time, the injuries at the top fall away too.
To do that, you have to see the base of the pyramid. And the base is enormous.
A single worker might make dozens of small safety-relevant decisions in a shift. Multiply that across every person, every bay, every shift, every day. Now count your observers. Even a committed programme captures a few hundred observations a month, gathered in short bursts by people who also have other jobs. Against a population of hundreds of thousands of actual behaviours, that is not a sample. It is a rumour.
And the sample is biased in the worst way. Observations happen when observers are free, in daylight, in the areas that are easy to reach, on the shifts that are well staffed. The unsafe acts that matter most tend to happen at three in the morning, in the awkward corner, at the end of a double shift, when no observer is anywhere near. The programme measures where the light is good, not where the risk is high.
The observer effect, in a hi-vis vest
There is a subtler problem still. The moment a worker sees the observer, the behaviour changes. Gloves go on. The shortcut is not taken. The card fills up with safe acts, and everyone feels reassured. This is the Hawthorne effect wearing a clipboard, and it means the very act of observing corrupts the thing being observed.
None of this is an argument against behaviour-based safety. The methodology is one of the most effective ideas in the history of the discipline. It is an argument about capacity. BBS works precisely as well as your ability to observe — and human observation simply does not scale to the size of the problem.
What changes when the camera is the observer
This is the gap that computer vision fills, and it fills it without asking you to abandon a single principle of good BBS. The methodology stays exactly as it is. What changes is who does the watching.
An AI safety layer added to the CCTV you already have watches every camera, every shift, without fatigue, breaks or blind spots. It does not tick a box every few minutes; it evaluates behaviour continuously. And crucially, it does not change what it sees. The worker does not modify his behaviour for a camera that has been on the wall for years. For the first time, the observation is of the real workplace, not the performance staged for the observer.
The system categorises the same at-risk behaviours a trained observer would flag: a person entering a marked exclusion zone, a body too close to a moving vehicle, PPE not worn where it is required, someone reaching into a line that should be isolated, a fall or a person on the ground. Each is logged, timestamped and located, not as a single anecdote but as data — a running census of how work is actually done.
This capability was not built in a classroom. The detection was forged offshore, on drill floors, where heavy equipment moves in tight spaces around people and a moment's unsafe behaviour has no margin for error. Proven since in a national oil major's operations, a major international port and an international airport, it now runs at a sub-0.05% error rate and has been measured cutting unsafe behaviours by around 90%. It runs on-premise, so the footage never leaves your site — which matters a great deal when the thing you are recording is your own people.
From counting acts to changing them
The purpose was never to accumulate observations. It was to change behaviour. Automated observation makes that loop tighter and fairer in three ways.
It makes coaching timely
An unsafe act flagged the moment it happens can prompt a conversation the same shift, while the moment is fresh, rather than a statistic reviewed weeks later when no one remembers the day.
It shows you the pattern, not the person
Because the system counts everything, it reveals where at-risk behaviour clusters — the junction where people always cut the corner, the machine that invites the reach-over, the hour of the shift when compliance falls. That points you at the conditions that produce unsafe acts, which is where fixes actually work. Good BBS was always meant to be about the system, not blame. Complete data makes that honest.
It keeps the programme alive
Every safety leader knows the way BBS decays — enthusiasm at launch, cards drying up by the second year. An automated observer does not lose interest. The programme keeps running whether or not anyone is feeling energetic about it this quarter.
The observer that never blinks
Behaviour-based safety asked its observers to do something almost impossible: to watch everything, everywhere, all the time, without ever changing what they saw. People cannot do that. A machine can. Not to replace the supervisor's judgement or the coaching conversation — those remain irreducibly human — but to give them, at last, a complete and honest picture of the behaviour they are trying to change.
Applying behavioural safety AI in practice
The observation gap that AI fills
Traditional behaviour-based safety programmes depend on trained observers conducting structured safety observations and providing feedback to workers on unsafe behaviours. The limitation of this approach is coverage: even the most active observation programme cannot observe every worker in every area on every shift. A single trained observer may complete a dozen structured observations per week on a large site where hundreds of workers are performing hundreds of tasks. The behaviours that are observed are not necessarily the ones that represent the highest risk.
AI-based behavioural monitoring does not replace the structured observation and feedback conversation — that human interaction has value that technology cannot replicate. What it does is extend the observation coverage to every camera, every shift, generating a dataset of observed behaviours that no team of human observers could match in volume. The safety manager who has data on 10,000 behavioural events per month, rather than 50, is working from a fundamentally different evidence base when deciding where to focus the human observation programme.
From event detection to behavioural intervention
The pathway from AI behavioural detection to behavioural change runs through the human management conversation. An alert that a specific worker is displaying fatigue at a machine station is an opportunity for a supervisor to initiate a conversation — not about the alert, but about whether the worker is OK, whether there are factors affecting their performance, and whether the working arrangement needs to be adjusted. The AI provides the signal; the supervisor provides the response.
The highest-value use of AI behavioural data is not individual-level intervention but pattern analysis. Which areas, which shifts and which tasks generate the most behavioural safety events? The answer points to engineering controls, schedule changes and workload adjustments that address the root cause of the behaviour rather than the behaviour itself.
Implementation checklist for behaviour-based safety AI
- Define the monitored behaviours explicitly: before deployment, agree with the safety team which specific behaviours will be monitored and how they will be used — this is both a practical configuration requirement and an important communication point for the workforce
- Separate safety monitoring from performance monitoring: confirm in writing, and communicate to workers, that behavioural monitoring data is used for safety purposes only and is not used in performance management or disciplinary processes
- Establish a behaviour trend review cadence: agree a weekly or fortnightly review of behavioural event patterns with the safety team and line management — this is where the data becomes actionable
- Train supervisors in using AI data for safety conversations: the conversation between a supervisor and a worker following a behavioural alert requires specific skills — the alert should be the trigger for a supportive conversation, not an accusatory one
See what your cameras have been missing — book a demo.
