Every EHS leader has sat through the demo. The lights dim, the vendor plays a curated clip, a neat green box snaps around a worker without a hard hat, and everyone in the room nods. It looks like magic. Three months and a signed contract later, the same system is drowning your team in false alerts, misses the incident that actually mattered, and quietly gets switched off.
The problem is rarely the technology. It is the buying process. Safety AI is now a crowded market, and the demos all look alike. What separates a platform that earns its place in your operation from one that becomes expensive shelfware is almost never visible in a forty-minute presentation. This guide is about how to see it anyway.
Start with the hazard, not the feature list
Most vendor conversations begin with a menu: PPE detection, people counting, fall detection, restricted-zone monitoring, speed, fire and smoke. It is tempting to tick the boxes and pick the platform with the longest list.
Resist this. A feature list tells you what a system claims to do. It tells you nothing about how well it does it in your environment.
Begin instead with your own incident data. What actually hurts people on your sites? For most heavy-industry operations, a handful of hazards account for the overwhelming majority of serious harm: vehicle–pedestrian conflicts, falls, and being struck by moving or dropped objects. Rank your hazards by consequence, then judge every platform on the two or three that matter most. A system that detects your top hazard reliably is worth more than one that detects fifteen hazards adequately.
The questions that actually separate vendors
Once you know what you need detected, the real evaluation begins. These are the questions that expose the difference between platforms.
"What is your false-positive rate in a live deployment?"
This is the single most important number, and the one vendors are most reluctant to give you plainly. An alert system that cries wolf trains your team to ignore it. Within weeks, a high false-alarm rate does not just waste time — it actively erodes safety culture, because your supervisors learn that the screen is usually wrong.
Ask for the error rate measured on real, uncurated footage, not on the vendor's demo reel. Ask what it means in practice: how many false alerts per camera, per day, will land on your control-room operator's desk. If the answer is vague, that is your answer.
"How does it perform in poor conditions?"
Rain, glare, low light, steam, crowding, unusual camera angles, workers partially hidden behind equipment. The demo was filmed on a clear day with good sightlines. Your site is not a clear day with good sightlines. Insist on seeing detection footage from conditions that resemble your worst shift, not your best.
"Where does our footage go?"
For many operations this is decisive. If a platform routes your camera feeds to an external cloud for processing, you have created a new data-security and privacy exposure — one your legal and IT teams will rightly scrutinise. An on-premise system that processes everything on site, where footage never leaves your network, removes that entire category of risk. Ask early. It shapes everything downstream, from procurement to works-council approval.
"What does this need from our infrastructure?"
The best safety AI adds a software layer to the CCTV you already own. Be wary of any proposal that requires ripping out working cameras and installing proprietary hardware. That is not a safety project any more; it is a capital works project, with the timeline and budget to match.
Proof beats promise
Anyone can build a model that works in a slide deck. Very few can show you a system that has held up under real operational pressure, over time, with lives genuinely at stake.
It is worth understanding where a platform was forged. Our own detection was built and hardened offshore, on oil and gas drill floors — heavy moving equipment, zero tolerance for error, no second chances. It runs today in national oil-major operations, a major international port and an international airport, with a sub-0.05% error rate and field-measured reductions of around 90% in unsafe behaviour. That kind of provenance is not a boast; it is a filter. A system proven in the most punishing environment there is will cope comfortably with a warehouse or a yard. The reverse is rarely true.
When you assess proof, ask for references you can actually speak to, in industries adjacent to yours. Ask how long those deployments have been live. A pilot that ran for six weeks is a promise. A system that has been running for two years is evidence.
Look past the pilot to the partnership
A safety platform is not a product you buy once. It is a relationship that has to last years. So evaluate the vendor, not just the software.
Who tunes the system to your site when the false-alarm rate needs dialling in? What happens when you add a new camera, or a new hazard, or open a new facility? How are model updates delivered, and do they risk breaking what already works? A responsive partner who understands operational safety is worth more than a marginally cleverer algorithm from a company that treats you as a ticket number.
A simple scorecard
Before your next demo, write down five things and score each vendor honestly against them:
- Top-hazard accuracy — how well it detects the two or three hazards that cause you the most serious harm.
- False-positive rate — measured on live footage, expressed as alerts per camera per day.
- Data handling — on-premise or cloud, and who can access your footage.
- Infrastructure fit — does it use your existing cameras, or demand new hardware.
- Proof and partnership — verifiable, long-running deployments in comparable settings, and a vendor who will stay in the room after the sale.
The platform that wins on these five will rarely be the one with the flashiest demo. It will be the one still working, quietly and accurately, on the day it prevents an incident you never hear about.
The buyer's checklist: what to verify before committing
Technical due diligence questions
The following questions should be answered with specific, verifiable data rather than general assurances, before any purchasing decision:
- What is the measured false positive rate in a deployment environment comparable to yours, over a period of at least six months?
- What happens to detection, alerting and recording when the site loses internet connectivity?
- How many concurrent cameras does a single edge node support, and what is the detection latency at that load?
- What is the model update process — how are updates delivered, who authorises them, and how is the model's performance verified post-update?
- What privacy and data protection documentation is provided as part of the deployment package?
Procurement process recommendations
Pilot before purchase: a genuine pilot — not a demo — runs the system on your cameras, in your environment, for at least 30 days, and delivers you data from your own site. Any vendor who cannot offer a genuine site-specific pilot should be viewed with caution.
Specify the handover requirement: before the pilot begins, define what successful completion looks like in measurable terms — a false positive rate below a defined threshold, a detection accuracy above a defined threshold, and a verified integration with your VMS or control room system.
Contractual provisions for ongoing performance: the service agreement should include provisions for false positive rate SLAs, model update notifications, and the process for addressing performance degradation over time. These provisions are as important as the initial accuracy claims.
Total cost of ownership: the numbers you need
The total cost of deploying and operating an AI safety monitoring system over a five-year period includes:
- Hardware: edge compute nodes, camera additions or replacements identified in Discovery
- Software: licence or subscription fees, including any per-camera pricing
- Implementation: Discovery phase, installation, commissioning, training
- Internal resource: the time required from IT, safety, and operations teams for ongoing management, alert review, and data analysis
- Upgrade costs: hardware refresh or software upgrade costs over the period
The TCO over five years is typically two to three times the first-year implementation cost. Understanding this fully before committing ensures that the investment is budgeted correctly and that the business case is based on realistic numbers.
See what your cameras have been missing — book a demo.
