A model can read eighty thousand tags a second. It cannot smell.
That sentence sounds like a joke until you have stood next to a switchgear cabinet and known, before any instrument agreed with you, that something inside it was cooking.
What the walk actually is
Gemba is Japanese for the actual place. The walk is not an inspection and not an audit. It is an engineer moving through the plant with no particular agenda, letting twenty years of accumulated pattern-matching run in the background.
What that engineer is doing, described technically, is multi-modal sensing with an unwritten prior. Vision, hearing, smell, and the vestibular sense of a floor that vibrates differently than it did last month. No instrument on the plant covers that range. More importantly, no instrument carries the prior.
Three things only the walk finds
Overheating insulation
The smell arrives long before the thermal signature crosses a threshold. By the time a temperature tag moves enough to trip an alarm, the failure is under way. A person walking past a panel catches it at the point where it is still an inspection item, not an outage.
The motor that sounds wrong
Not louder. Not faster. Wrong. A bearing entering the early stages of spalling changes the character of the noise before it changes its amplitude, and character is precisely what a threshold cannot encode. Vibration analysis will find it eventually — the walk finds it first, and for free.
The site that has become unsafe
A pallet left in a walkway. A guard rail removed for maintenance and not replaced. A hose across a stairwell. None of this appears in the CMMS, the historian, or the digital twin. It appears only to a person who is physically present and looking.
Why this is not nostalgia
The temptation, once a plant has a predictive platform, is to treat the walk as ceremony. Data covers the assets; the walk covers nothing the data does not. This is exactly backwards.
The walk is the only mechanism you have for validating the sensors themselves. When the historian says the pump is healthy and the pump sounds terrible, one of the two is lying. The engineer who has never walked the floor has no way to know which.
That is the loop. The model watches the asset. The human watches the model — but only by watching the asset directly, with instruments the model does not have.
What to do about it
- Schedule the walk as work, not as slack. If it is optional, it is the first thing cut in a busy quarter, and the sensor-validation loop quietly opens.
- Log what the walk finds, in the same system as everything else. A finding that lives in an engineer’s head is a finding the organisation does not have.
- Compare walk findings against model output, deliberately. Every disagreement is either a sensor problem or a model problem. Both are worth knowing about.
- Send the junior engineer. The prior is not transferable by document. It is built by walking.
None of this slows the AI programme down. It is what keeps the programme honest.
The reliability engineer of the next decade will not compete with the model on arithmetic. They will do the one thing the model cannot: go and look.