Every safety manager who has sat through a vendor demo knows the feeling. The screen shows a tidy warehouse aisle, good light, a single forklift, a single person in a hi-vis vest. A neat box appears around each. A confidence score ticks up. The salesperson smiles. It all looks the same as the last three demos you watched.
And that is precisely the problem. In a controlled clip, on a clean camera, in flat daylight, almost any modern computer-vision system can draw a box around a person. The gap between vendors does not show up in the demo. It shows up six months later, on your site, at three in the morning, in the rain, when the alert either fires or it doesn't.
So the real buyer's question is not "can it detect a person?" Nearly all of them can. The question is: what happens when conditions get difficult — and how do you tell that in advance, before you have signed?
Why every demo looks the same
Object detection has become a commodity. The underlying models are widely available, the tooling is mature, and a competent team can stand up something that recognises people and vehicles in a weekend. That is genuinely good news for the industry. It is also why the demo has stopped being a useful way to choose.
A demo rewards the wrong things. It rewards good lighting, a cooperative subject, and a scene the vendor has quietly tuned for. It hides everything that actually determines whether a safety system earns its keep: how it behaves at a bad angle, in glare, in fog, in a crowd, when two workers overlap, when a load swings between the camera and the hazard, when the same alert has fired forty times today and nobody is listening any more.
If you want to buy well, you have to look past the polished surface and interrogate three things that a demo will never show you.
The three things that actually separate safety AI
1. Depth — proven where lesser systems break
Depth is the difference between a system that works in a car park and one that works on a drill floor. It is accuracy under the conditions that real industrial sites throw at a camera every single day.
Ask a vendor to show you the hard cases, not the easy ones. Partial occlusion — a person half-hidden behind machinery. Heavy weather — rain on the lens, spray, low sun straight into the sensor. Density — twenty people in a muster area, not one. Unusual poses — a worker on the ground, which is the exact posture your fall detection exists to catch and the exact posture generic "person" models are worst at. Odd mounting angles, because your cameras were installed for security, not for AI, and nobody is going to remount them.
A shallow system degrades quietly. It keeps drawing confident boxes while quietly missing the case that matters, or it floods you with false alarms until your team switches the alerts off — which is the same as having no system at all. Depth is what keeps detection honest when the scene stops cooperating.
2. Proof — deployed, measured, and still running
The second separator is evidence. Not a benchmark on a public dataset, not a lab figure, but proof that the system has run on real sites, under real conditions, and produced a measurable change in behaviour that a customer was willing to stand behind.
Ask three specific questions. Where has this run in production, in an environment like mine? What was measured — detection error rate, false-alarm rate, reduction in unsafe acts — and over what period? And is it still deployed today, or was it a pilot that quietly ended? A vendor with genuine proof answers plainly. A vendor without it reaches for adjectives.
Here is where pedigree earns its place, and only here. Our detection was not born in a warehouse. It was forged offshore, in oil and gas, on drill floors — an environment with heavy moving equipment, zero tolerance for error, and lives at stake in every shift. It has since been deployed in the operations of a national oil major, a major international port and an international airport, running at a sub-0.05% error rate and delivering field-measured reductions of around 90% in unsafe behaviour. We lead with that not to boast, but because a system hardened in the worst conditions tends to be unremarkable in easier ones. The reverse is rarely true.
3. Privacy and control — where your footage goes
The third separator is quieter but increasingly decisive: where does the video actually go? A great many safety-AI products stream your footage to a cloud you do not control, to be processed on infrastructure you cannot see. For a lot of industrial sites — offshore, defence-adjacent, critical national infrastructure, anywhere with a serious works council or a GDPR-conscious workforce — that is a non-starter before the accuracy conversation even begins.
An on-premise system, where analysis happens on site and footage never leaves the perimeter, changes the entire procurement conversation. It shortens the security review, it satisfies the data-protection impact assessment, and it means the answer to "who can see our workers?" is simply: you, and no one else. Ask early. It is far easier to design in than to bolt on.
How to run a demo that actually tells you something
You cannot un-invent the demo, but you can make it earn its keep. A few practical moves:
- Bring your own footage. Hand the vendor a clip from your worst camera — the one with the glare, the awkward angle, the crowded shift change — and ask them to run it live.
- Ask for the misses. Request the false positives and false negatives, not just the hits. How a vendor talks about their errors tells you more than their accuracy figure.
- Ask about alert fatigue. A system that cries wolf is worse than none. Ask how they tune sensitivity and what the real-world false-alarm rate looks like after a month on site.
- Insist on a reference. Speak to a customer running the system in an environment like yours, still using it a year on.
None of this is exotic. It simply moves the conversation from what looks good on a screen to what holds up on a site.
The bottom line
Safety AI has quietly split into two categories that look identical in a slide deck. One draws confident boxes in easy conditions and falls apart in the ones that matter. The other has been proven where failure is not an option, can show you the measured evidence, and keeps your footage under your own roof. The demo will not tell them apart. Depth, proof and privacy will.
Choose on those three, and you are no longer buying a demo. You are buying a system that will still be watching when the weather turns.
The questions that separate real safety AI from safety AI theatre
Deployment history questions that matter
The single most useful question a buyer can ask a safety AI vendor is not about features or accuracy claims — it is about deployment history. How many live industrial deployments is the platform currently running in, and for how long have they been operating? A platform that has been running in 100 sites for six months is a different risk profile from one that has been running in 13 sites for eight years. Longevity of deployment is the proxy for everything that demos and pilots cannot show: whether the system handles the long tail of unusual events, whether it remains accurate as lighting and environment change across seasons, and whether the operational model is sustainable for a client to manage.
The error rate under operational conditions
An error rate quoted in marketing material is a claim that requires verification. The questions to ask are: over what time period was this measured, on what environment, and what is the denominator — how many total events does 0.05% represent? A 0.05% error rate on 100 events per day is not the same as 0.05% on 18,000 events per day. Ask for the methodology behind the number.
What happens when the network goes down
A safety AI system that stops functioning when the internet connection drops is not a safety system you can rely on. Ask directly: what happens to detection, alerting and recording when the site loses connectivity? The answer defines whether the system is a tool for your safety programme or a dependency on your connectivity provider's uptime.
How false positives are managed and what recourse you have
Every system generates false positives. What distinguishes vendors is what happens next: is there a systematic calibration process to address false positive patterns, is there a mechanism for you to flag false positives that feeds into model improvement, and is there a defined SLA for reducing false positive rates that are above an agreed threshold?
Implementation checklist for evaluating safety AI vendors
- Request live deployment references in your sector: case studies and demos are vendor-controlled; a reference call with a safety manager at a site in your industry running the system for more than two years is the most reliable evidence
- Ask for a data-specific pilot proposal: a vendor who offers only a demo is not offering a pilot; a genuine pilot runs the system on your cameras, in your environment, and delivers you the data from your own site
- Assess the Discovery phase scope and methodology: the quality of the deployment process is the strongest predictor of the deployment outcome; a vendor whose Discovery phase is a sales conversation rather than a technical assessment is unlikely to deliver a calibrated, effective system
- Review the ongoing support and calibration model: understand what ongoing technical support is included, how model updates are delivered, who manages the system post-deployment, and what the process is for reporting performance issues
See what your cameras have been missing — book a demo.
