On-the-job training (OJT) with AI — the veteran teaches once, the plant learns forever.

The problem

The most expensive knowledge in the plant lives in a conversation nobody records.

OJT — the veteran teaching the new hire with the machine running — is the most effective way a factory has to pass on real know-how. And it is, at the same time, the most fragile:

1. It depends on who is on shift. If the talkative veteran is on, the new hire learns in three days; if the quiet one is on, it takes three months.
2. It depends on the day. If the line is on fire, there is no time to teach — and the new hire learns whenever they can, watching out of the corner of their eye.
3. There is no objective record of what was taught. The supervisor signs off «trained», but nobody knows which parts landed and which did not.
4. The day the veteran retires, half the plant walks out. The veteran's eye, the trick that keeps the dispenser from jamming, the exact recipe for the difficult SKU changeover — all of it leaves with them.

The result: the factory accumulates operators marked as «trained» who are in fact only half-trained, and it relies on luck (who does OJT with whom) for the critical know-how to stay inside. It is an invisible cost — until the day it shows, and it shows hard.

How it fits the IRIS system

iLEAN does not replace the veteran — it captures what the veteran knows and gives the new operator a safety net.

OJT is not digitized by making video courses: it is digitized where it happens, at the workstation, while it happens. iLEAN acts as the filler between the veteran's head, the gesture being taught and the screen where the new operator is alone at five in the morning on the following shift. See the full IRIS architecture →.

Connect records the voice, Edge sees the gesture, the agent stays with the new hire when they are left alone. A warm bed for the operator.

The iLEAN pieces applied to OJT:

• Connect — phone + earpiece (or a tablet where phones are banned) per operator. When the veteran teaches, Connect transcribes the session, links it to the workstation and files it under the new hire's record. The veteran does nothing extra — they talk as always, and the session stays.
• Edge — the vision terminal on the line sees the critical gesture (the selector adjustment, the pressure on the button, the coupling manoeuvre) and stores it as a visual reference. The day the new hire tries it alone, Edge will know whether it is being done the same way.
• Companion agent — the new operator, already on shift, wears the earpiece. If in doubt, they ask under their breath; the agent answers with the veteran's own answer. If the gesture is not the one they were taught, the agent warns them before — not after. Critical operations are always signed off by the person; the agent has no hands on the critical work order.

See skills matrix with AI → See cross-training polyvalence →

Before and after

Classical OJT vs. OJT captured with iLEAN

Aspect	Classical OJT (veteran + new hire)	With iLEAN Connect + Edge + Agent
Session capture	Evaporates at the end of the shift	Transcribed, linked to the workstation, reusable
Critical gesture	«You'll see it eventually»	Visual reference recorded by Edge
New operator alone on shift	«I'm going to mess this up»	Agent alongside through the earpiece
Objective record of what was learned	Signature on a checklist	Session + gesture verification on the line
Veteran's retirement	Their know-how walks out	Their session stays as a plant capability
Time to autonomous operator	Variable, down to luck	Shorter and more predictable

Impact estimate

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 50-150 operators, rotating shifts, several key workstations depending on a handful of veterans. First value (first OJT session captured and one new operator assisted by the agent) within a few weeks.
• Indicative payback between 4 and 9 months. Hard levers: less scrap from the operator in training, fewer veteran hours spent re-explaining the same thing, less overtime caused by slow newcomers during SKU changeovers.
• Expected reduction in time to autonomous operator of ≥ 30% versus the baseline — always on the conservative side.
• The heavy hit: when a veteran retires, what they know no longer leaves with them.

And the plant manager's fair objection

«What if the AI makes up what the veteran taught?» — hallucination is a problem of free generation, not of anchored tasks. Here the AI merely transcribes the session, links the image of the gesture and returns to the new operator exactly what the veteran said. On tasks anchored to the source, the best models brought error below 1.5% ^[1]. And even so, nothing critical is decided alone: the agent warns, the person signs off.

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