Near misses with AI — the warning the plant gives itself, captured before the hit.

The problem

The near miss nobody reported is next quarter's accident.

Heinrich's pyramid has been in every safety manual for decades: for every serious injury there are dozens of minor injuries and hundreds of near misses. What the manual does not say is that in most plants near misses are neither captured nor counted:

1. Reporting costs time. Filling in a form after the shift, on a screen you can only reach from the office, never fits into the operator's routine.
2. Fear that the report will backfire. "If I say I almost caught my hand, won't they open a file on me?" Result: silence.
3. The evidence is lost. Three hours after the event, nobody remembers which forklift went past, at what time, or whether the light curtain was bypassed.

The safety manager knows it — and still the only data available is the accident that did happen. The classical system (form + suggestion box + monthly meeting) captures maybe 5% of the real near misses. That invisible 95% is precisely the pyramid warning about the next hit.

How it fits the IRIS system

iLEAN drops the cost of reporting to zero — and brings the cost of silence into the open.

The classical OHS management system fails because it loads everything onto the operator and the line supervisor. iLEAN does not add a fourth form: it acts as the filler that seals the gap between what happens on the line and what reaches the safety committee. It captures the event when it happens, through the most natural channel available.

Edge sees the forklift that passed too close. Connect hears the operator say "the guard has opened again". The agent classifies and proposes. The person signs the corrective action.

The three iLEAN pieces applied to the near miss:

• Edge — a terminal with computer vision (CNN) over critical zones: forklift aisles, panels with residual energy, presses, robots. It classifies events as near misses (minimum distance breached, light-curtain bypass, entry into a zone with live equipment) and records the sequence. It works with no network: if the plant loses WiFi, it keeps capturing on cabinet power.
• Connect — the operator carries a phone or a full-duplex headset on the line. They report by voice at second zero: "lockout step 4, the padlock has moved". Voice transcribed, audio stored, photo optional. It also captures what arrives from outside: the forklift supplier's notice, the shift leader's message, the insurer's email.
• Agent — cross-checks the event against the ISO 45001 / OSHA taxonomy, classifies it by hazard type, potential severity and likelihood of recurrence, and proposes a prioritized corrective action. It assembles the live audit dossier without anyone filling in a spreadsheet after the fact. The committee validates and signs.

See the full IRIS architecture →

Before and after

Manual near miss capture vs. captured with iLEAN

Aspect	Form + box + monthly meeting	With iLEAN Edge + Connect + Agent
Moment of capture	Hours later, if the operator remembers	At second zero — voice, photo, video of the sequence
Actual reporting rate	~5% of real near misses (plant estimate)	Jumps an order of magnitude as the cost of reporting falls
Evidence	The operator's memory, a handwritten note	Audio + Edge imagery + the full time window
ISO 45001 classification	The safety manager, on Friday	The agent proposes, the manager validates
Corrective action	Reaches the next committee, two weeks later	Automatic prioritized proposal, signed within hours
Audit dossier	Rebuilt by hand before the visit	Live, exportable in one click, with traceable evidence

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the actual data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Plant with 100–250 operators, 3–5 critical zones (forklifts, presses, panels with residual energy), monthly OHS committee.
• Edge pilot in one critical zone + Connect across the shift team. First value expected in a few weeks: reporting rate up by at least an order of magnitude and a live dossier ready for audit.
• Indicative payback between 4 and 9 months, depending on historical incident frequency and the average cost of sick leave, penalty or line stop.
• The hard lever is a single serious accident avoided: days of sick leave, compensation, sanction proceedings, line stoppage, reputational damage. One event pays for the whole system.

And the safety manager's fair objection

"Isn't this about watching the operator?" — no. The Edge cameras look at the critical zone, not at faces: what they see is the forklift trajectory and the state of the guard, not who is driving. The system reports "near miss in aisle 3 at 14:32", not "so-and-so made a mistake". This is process verification, not people: fight the problem, not the person — and hallucination is a problem of free-form generation, not of anchored tasks such as classifying an event against a taxonomy: on anchored tasks the best models brought the error rate below 1.5% ^[1].

^[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.