Workplace safety with AI vision — anticipating the near miss before it becomes an accident.
Every plant has restricted areas, shared crossings and machines with a risk perimeter. The classical system manages them with signage, training and memory — and that fails exactly when it matters. iLEAN Edge sees near misses in real time and warns at second zero. Data is aggregated by zone, never by person. The person signs off; the system anticipates.
The accident everyone saw coming — and nobody had a way to record.
Every industrial plant has known risk zones — the shared pedestrian-forklift crossing, the press perimeter, the robot cage, the loading area. The classical prevention system does what it can:
- Signage and training — effective while attention holds, fragile on a heavy day.
- A supervisor who remembers — the shift leader who warns for the umpteenth time that day, until they get tired of it.
- Post-accident investigation — a meticulous report of what happened, once it has already happened.
- Unrecorded near misses — every day there are coincidences that "were close", but because they never became accidents they are never documented. And they are exactly the signal announcing the accident that will happen next month.
Modern prevention has known for decades that near misses are the gold mine of continuous improvement. But capturing them systematically, without depending on an operator writing them down after a scare, was until very recently impossible. That has just changed.
iLEAN doesn't add another CCTV — it adds a pair of eyes that understand what they see and warn in time.
The safety problem isn't a lack of cameras: most plants already have CCTV. The problem is that cameras record, they don't understand — and the footage gets watched afterwards, if it gets watched at all. iLEAN acts as the filler between the eye that looks and the warning that reaches the operator. See the full IRIS architecture →.
Edge looks at the crossing and understands what it sees. If the forklift's path and the pedestrian's path are about to meet, the signal turns amber — before the incident.
The iLEAN pieces applied to workplace safety:
- Edge with CNN in critical zones — machine perimeters, crossings, restricted-area entries, loading zones. Inference in milliseconds. It recognizes people, vehicles, postures, trajectories and risky proximities.
- Local warning at second zero — amber light, traffic signal at the crossing, a soft voice prompt in the affected operator's earpiece. The warning arrives in time to avoid — not after the photo in the report.
- Aggregated prevention agent — near misses are aggregated by zone, time of day and pattern, never by person. The safety manager sees where the structural risk sits (a crossing with poor visibility, a time of day with more fatigue) and acts on the place. Operators never appear by name in safety reports — that is the red line.
- ISO 45001 evidence — objective, continuous logging of the near misses that classical OHS depends on someone writing down. Auditors value it — because it shows the plant doesn't wait for the accident to act.
- Works without network — if connectivity drops, Edge keeps looking and warning. What is critical does not depend on the Wi-Fi.
See PPE management with vision → See cross-training and polyvalence →
Signage + recording CCTV vs. safety with iLEAN AI vision
| Aspect | Signage + CCTV + post-hoc investigation | With iLEAN Edge in the critical zone |
|---|---|---|
| Type of response | Reactive, after the incident | Anticipated, at second zero |
| Near-miss capture | Only the ones somebody writes down | All of them, aggregated by zone |
| Data handling | Video identifying individuals | Patterns by zone, not by person |
| ISO 45001 support | Selective memory | Continuous, objective evidence |
| System response time | Days (reviewing the recording) | Milliseconds |
| Resilience to network loss | Depends on the central NVR | Edge keeps running locally |
Impact estimate for your plant — to be validated with your numbers.
Block flagged as an estimate to be validated. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.
- Plant with 2-6 critical zones (crossings, perimeters, restricted areas), 60-200 operators, at least one near miss logged per month. First value (first zone with Edge live and near misses aggregated) within a few weeks.
- Indicative payback between 4 and 9 months. The hard lever is brutal: one serious accident avoided is worth far more than the plant's entire system. But the structural ROI sits in continuous improvement of OHS with real data, not selective memory.
- Expected reduction of incidents and near misses in instrumented zones of ≥ 30% versus the baseline — always conservative.
- Continuous evidence for ISO 45001, with no manual effort from the safety manager.
The underlying data and the reasonable doubt
"What if the system gets it wrong and warns when there is no risk?" — the reliability of AI in anchored tasks (recognizing a defined trajectory, a defined posture, a defined proximity) is very different from its reliability in free generation. In tasks anchored to the source, the best models pushed the error below 1.5% [1]. A false alert is far less of a problem than the alert that never arrived — and the person stays in command: the agent warns, the safety supervisor decides the structural changes, the operator corrects the path.
[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.
What people ask about workplace safety with AI vision
What exactly does "workplace safety with AI vision" cover?
Three main scenarios: (1) restricted-area intrusion — an operator or a vehicle enters a delimited zone (a press perimeter, a robot cage, a loading area) when they shouldn't; (2) near misses — gestures, trajectories or coincidences that didn't end in an accident this time, but were one second away from doing so (a pedestrian close to a moving forklift, a hand close to a machine's moving area); (3) vehicle-pedestrian contact in shared aisles. iLEAN Edge sees those patterns in real time and fires a local warning — light, voice, traffic signal — before anything happens.
How is this different from the security cameras we already have?
Traditional cameras record — someone will watch the footage later, if anyone watches it at all. Edge infers in real time with convolutional neural networks (CNN) trained to recognize risky trajectories, postures and proximities. The response is local and takes milliseconds: the warning reaches the operator in the moment, not a hard drive for the day there is an incident. And nobody is watching a monitor: Edge looks and warns; the safety manager later receives aggregated patterns, not video of each individual.
Are the detected near misses used to discipline operators?
No. That is the red line of the IRIS model: the system is designed to empower people, not to police them. Near misses are aggregated by zone, time of day and pattern — never by person. The safety manager sees where the risk is (a badly designed crossing, visibility blocked by a column, a time of day with more fatigue) and acts on the place, not on the individual. Detecting near misses is one of the most valuable findings in modern prevention — and for the first time it is affordable to capture all of them, not only the ones somebody remembers to write down.
How does it relate to ISO 45001 and to employer liability?
ISO 45001 requires systematic processes for hazard identification, risk assessment and continuous improvement. One of the hardest levers in the standard is documenting near misses, because humans forget them or normalize them. Edge provides objective, aggregated logging that feeds OHS processes with continuous evidence instead of selective memory. It is the kind of capability auditors value — because it shows the plant doesn't wait for the accident to act.
Does it work if the plant is old and the aisles are poorly marked?
Yes. In fact that is where it shows most. Edge doesn't need a new plant or a perfect layout: it needs to see. The model is trained on the real zones (their lighting, their angles, their floor markings if there are any) and learns to recognize the risk patterns of that specific plant. In old plants with shared aisles and poorly delimited zones, AI vision is usually the most realistic way to close the safety gap without heavy construction work — because the real work (redesigning the crossing, installing barriers) comes after knowing where the near misses are.
Tell us about your case and in 48h we'll send you the estimated ROI of your workplace safety with AI vision.
We work on the real critical zones of your plant and your documented accident record. Diagnostic with no commitment.
Request estimated ROI in 48h See Lean Manufacturing