Software has an old saying: garbage in, garbage out. In industrial reliability it reads differently.
Garbage in, danger out.
The difference is not stylistic. A bad model in marketing produces an inaccurate sales forecast. A bad reliability model tells you the bearing has three months left.
What actually happens
There is a widespread assumption that artificial intelligence is a filter — that it will look at dirty data, recognise the noise, and somehow discard it.
It is not a filter. It is an amplifier.
AI imitates every error contained in its input, and does so with a confidence no human possesses. An engineer looking at a suspicious record frowns. The model treats it as truth and builds a forecast on it.
Four sources of garbage
Miscalibrated sensors
A vibration sensor that has drifted fifteen per cent will not tell you so. It keeps sending numbers. The model accepts them as reality and learns from them. It will then explain, at length, why the pump is healthy — right up to the day it stops.
Missing history
Equipment without failure history does not exist as an object of prediction. The model will not say “I don’t know”. It will find a pattern in what it has — and that pattern will be a pattern in noise.
Sloppy CMMS coding
Work order closed, failure cause: “other”. Or blank. Or “fixed it”. A thousand such records, and the NLP module will faithfully identify your root cause: work was performed.
Blank failure records
The failure happened. Nobody logged it, because the shift was ending. In the model’s view that failure never occurred. MTBF rises. Everyone is pleased.
How it ends
The system issues corrective and preventive actions based on patterns that do not exist — or that exist, but were produced by a data-collection fault rather than by physics.
The fishbone diagram will be drawn immaculately. The RCA report will be written in a minute, in careful prose, with conclusions. The recommendation will be unambiguous.
And it will be wrong.
What the human does
The engineer, unlike the machine, understands context. There are four things they know about their plant that appear in no table:
- Which records can be trusted. They remember that the crew on line three fills in work orders honestly, and the crew on line five fills them in when they can. This knowledge is documented nowhere.
- When a sensor is lying. Because they have seen that sensor, know when it was last verified, and remember that it has read high since the March overhaul.
- When the readings contradict physics. The model does not know that this cannot happen. The engineer does.
- When the machine must be overridden. The right of override is not a technical function. It is a professional obligation.
This is what we mean by the judgment layer — a layer that sits above the computation, not beside it. Validating models and deliberately hunting for hallucinations is a direct duty of the human in the pairing, not an optional discipline for the paranoid.
The order of work
From which follows a conclusion that is inconvenient for anyone selling software.
An AI programme does not begin with AI.
It begins with data quality. Calibration audit. CMMS entry discipline. A mandatory failure-cause field. A single failure-mode taxonomy. Training the workforce on why they fill in those fields at all — because a technician who does not understand the purpose of a record will write “other”.
Only then, models.
A company that has bought a predictive analytics platform and laid it over a dirty CMMS has bought a mechanism for producing wrong decisions quickly and confidently. Previously, incorrect conclusions were reached slowly, one at a time. Now they arrive by the thousand, on handsome dashboards.
What this means for you
If you are a director: the budget for data and for training people is not preparation for deploying AI. It is deploying AI. Everything else is a licence fee.
If you are an engineer: your value in the age of models grows not because you can calculate. The machine calculates. Your value is that you are the only person who knows where the bodies are buried in that data.
Do not give that knowledge away. It does not digitise.