The knowledge that predicts failure is spoken, and today it is lost

The problem

The knowledge that anticipates a breakdown is spoken, and it dies at the end of the shift.

The most valuable knowledge in an industrial plant is the knowledge that anticipates downtime. And it is tacit, spoken and perishable: it dies at the end of the shift. The maintenance management system is only fed once something has broken. The result is an asset history that records breakdowns perfectly but none of the signals that preceded them. Without those signals, predictive maintenance has nothing to learn from. And the person holding those signals is precisely the one least able to type: standing, on the floor, hands busy, with no surface to rest anything on. There is a longer horizon problem too: when a veteran technician retires, two decades of know-how leave with them, having never reached any system.

• On the round somebody notices the bearing that sings differently, the leak in a hydraulic cylinder, the electrode more worn than it should be. None of it gets typed.
• The maintenance management system is only fed once there is a breakdown. The asset history tells the failures perfectly and records none of the signals that preceded them.
• Without those signals, predictive maintenance has nothing to learn from: it is asked to predict from a history that only contains endings.
• And the person holding the signals is precisely the one least able to type: on their feet, in gloves, in hearing protection, with no surface to write on.
• There is an uncomfortable horizon too: when a veteran technician retires, twenty years of knowledge that never reached any system leaves with them.

How it fits the IRIS system

Connect in hands-free voice mode — the AI hears what the manager chooses to tell it.

Connect in hands-free voice mode.

This is the fastest payback in the series for a simple reason: it needs no plant hardware. An earpiece compatible with mandatory hearing protection, and the round that already happens every day.

• A Bluetooth earpiece compatible with mandatory hearing protection.
• The manager speaks while walking, in natural language and in the vocabulary of the trade.
• iLEAN listens and structures: asset, symptom, criticality, context. It stores this in central memory and opens the ticket in the maintenance system.
• It returns support in seconds with what the person does not carry in their head: prior reports on that asset, the date of the last component change, the preventive due date, and a proposed action.
• Nobody is under scrutiny: the AI hears what the manager chooses to tell it.

See the full IRIS architecture →

Before and after

Today's round versus the captured round

Aspect	Today	With iLEAN Connect
Where the observation ends up	In the head of whoever saw it	In the asset history
What the CMMS holds	The breakdown, once it happened	The prior signals as well
Effort for the manager	Type it later, if there is time	Speak while walking
What comes back on the spot	Nothing	Prior reports, last component change, preventive due date
Basis for predictive	Failures only	A symptom trail per asset
A veteran retiring	The knowledge is lost	It stays in the history

Impact estimate

Estimated impact — to validate with your own numbers.

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

• Estimated payback 3-8 months — the fastest in the series, because it requires no plant hardware.
• Improved mean time between failures on the assets that accumulate the most prior reports.
• Reduced unplanned downtime at the bottlenecks, which in this plant are the press and the welder.
• Preventive work reordered by observed symptom instead of by calendar.

estimated payback 3-8 months — the fastest in the series, because it requires no plant hardware. It shows up as improved mean time between failures on the assets with the most prior reports, and as less unplanned downtime at the bottlenecks. *Estimate to validate.*

And the fair question from the production manager

«Is this here to check up on my team?» — no, and framed that way the case would fail within a month. The AI hears what the manager chooses to tell it, when they choose to tell it. There is no continuous listening and no record of where anyone is. The gain runs the other way: the technician no longer has to remember at the end of the shift what they saw at seven in the morning.

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