Draft: under review before publication.
A mid-size construction site runs somewhere between eight and forty cameras. Multiply by 24 hours and you get a simple, uncomfortable number: a single site can record more footage in a day than its safety team could watch in a month of doing nothing else.
So nobody watches. The cameras earn their keep in other ways: theft deterrence, dispute evidence, progress documentation. All of it is backwards-looking. The footage becomes useful only after something has already gone wrong, when someone scrubs through the recording to answer "how did this happen?"
The near-miss is on tape
Here's what makes that genuinely painful: the evidence of the next incident is usually on camera before it happens. Safety research has pointed at the same pattern for decades: serious injuries are preceded by near-misses and routine violations of exactly the kind a camera sees. The missing vest at the gate. The worker cutting through the crane's swing radius because it saves ninety seconds. The pallet that's been leaning against the fire exit since Tuesday.
None of these are exotic. They are visible, repeated, and boring, which is precisely why no human reviews eight hours of footage to find them.
Why "watch more" doesn't work
The traditional answers all fail the same way:
- Hire someone to watch the wall of monitors. Attention research is brutal on this: sustained vigilance over multiple video feeds collapses within tens of minutes. It's also unaffordable for most sites.
- Review footage after the fact. Useful for investigations, useless for prevention. The injury has already happened.
- Walk the site more. Walkarounds matter and nothing replaces them. But a safety manager walking zone A is, by definition, not seeing zones B through F.
What changes when software watches
Continuous computer vision flips the model: instead of a human searching footage for violations, the violations come to the human. The safety manager's job stops being "watch everything", an impossible task, and becomes "act on what's surfaced", which is exactly the judgement-heavy work humans are good at.
The bar the software has to clear is trust, not omniscience. An alert that arrives with a clip, a camera ID, and a timestamp can be verified in ten seconds. An alert that cries wolf gets the whole system muted by Friday. That's why persistence thresholds (alerting on sustained violations rather than single frames) matter more than raw model accuracy.
The honest version of this pitch
Software watching cameras will not replace a safety program. It won't attend the toolbox talk, coach a crew, or make the judgement call about whether work continues in high wind. What it does is narrow: it makes the existing camera investment forward-looking, so the pattern that precedes the injury gets seen while there's still time to interrupt it.
That narrow thing happens to be the thing nobody on site has time to do.