RCM with AI — reliability centered maintenance stops being a consultant's PDF.

The problem

RCM does not fail because of the method — it fails because of the operating model.

Almost nobody disputes the RCM method: identify the asset's functions, its failure modes, their consequences, and decide the maintenance that applies. The problem is not the what — it is how it is sustained over time.

1. The analysis is done once. A consultancy comes in for three weeks, interviews the team, looks at the critical assets and delivers a 400-page document with the maintenance plan. The day the delivery is signed off is the last day the document is up to date.
2. The plant changes faster than the PDF. An asset is modernized, a spare-parts supplier is swapped, a new failure mode is discovered. None of that finds its way back into the analysis — because updating it by hand costs weeks nobody has.
3. The real history is never crossed with the analysis. The CMMS accumulates breakdown reports, the maintenance manager keeps a parallel spreadsheet, and the veteran knows the odd symptoms by heart. But that knowledge is never compared with the theoretical plan. The plant runs two maintenance programmes in parallel: the one the PDF describes and the one people actually do.

The outcome is predictable: the RCM plan becomes corporate folklore, real maintenance gets decided day by day, and critical assets keep breaking down through failure modes that were described in the analysis but that nobody ever crossed with what the CMMS was already seeing.

How it fits the IRIS system

iLEAN does not redo your RCM — it connects it to the reality of the asset.

A well-built RCM already exists in many plants. What is missing is the filler between that analysis and daily operations: Connect captures the history wherever it lives, the Agents cross that history with the RCM failure modes, and Edge watches for the early symptom before it becomes a breakdown. iLEAN does not ask you to change CMMS or redo the RCM. It connects it.

Connect captures the asset's history. The agent crosses it with the described failure mode. Edge sees the symptom on the line. The maintenance person decides and signs off — the plan stays alive.

The three iLEAN pieces applied to a living RCM:

• Connect — reads the modern CMMS, the legacy CMMS, the maintenance manager's parallel spreadsheet, the paper breakdown reports the supervisor photographs, and the veteran's knowledge dictated by voice into an earpiece. Everything enters at second zero as usable data.
• Agents — cross the failure modes of the RCM analysis with the real history. They flag the ones that behave as predicted, the ones that show up at a different frequency, and the new failure modes the original RCM never considered. They propose an update to the plan; the maintenance manager decides.
• Edge — a terminal with computer vision and sensor inputs on the critical asset. It detects the early symptom (out-of-pattern vibration, thermal halo, leak, abnormal noise), crosses it with the expected failure mode, and alerts through Connect. When it triggers, it does so in milliseconds and with no network needed.

See the full IRIS architecture →

Before and after

RCM in a PDF vs. living RCM with iLEAN

Aspect	Classical RCM (consultancy + PDF)	RCM with iLEAN Edge + Connect + Agents
Validity of the analysis	Valid on delivery day; goes stale on its own	Living plan, updated against real history
New failure modes	Do not enter the plan until the next project	The agent detects them and proposes their inclusion
Asset history	In the CMMS + parallel spreadsheet + the veteran's head	Unified and crossed with the expected failure mode
Early symptom	Operator's visual round; depends on who walks by	Edge watches the critical asset constantly
Operation without a network	n/a (a PDF needs no network)	Edge keeps watching as long as the panel has power
The veteran's knowledge	Walks out the door the day they retire	Captured by voice and turned into an asset pattern

Impact estimate

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

The block below is an estimate to be validated with your plant's data. We lay it out so the steering committee has an order of magnitude; we refine it during the diagnostic.

• Mid-sized industrial plant, RCM done at some point with a consultancy, outdated today. Critical assets identified: 10-15 according to the Pareto.
• Pilot on the 2-3 most critical assets: Connect against the CMMS + spreadsheet + voice, Agents crossing with the RCM failure modes, Edge on the most critical of the three. First value expected within a few weeks.
• Indicative payback between 4 and 9 months depending on the criticality and stoppage cost of the chosen asset. A single unplanned stoppage avoided on a critical asset pays for the pilot.
• Expected reduction in unplanned downtime hours on the pilot assets in the order of ≥30% after the first learning cycle.

And the maintenance manager's reasonable doubt

"What if the agent proposes a wrong update to my RCM plan?" — hallucination is a problem of free generation, not of anchored tasks. Here the agent does not invent the plan: it crosses the real history with the failure mode described in the analysis. In tasks of this type (recontextualizing a data point from one system to another, comparing against a known pattern), the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: the agent proposes, the person signs off. The three safety rings exist precisely for this.

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

Related methods: TPM (Total Productive Maintenance) · OEE · PDCA