AI Prediction

A forecast that changes nothing is a dashboard

Predictive maintenance in this industry usually stops one step short. Something measures the building, a chart appears, and the cleaning schedule stays exactly where it was. The prediction was real; it just had no consequences.

We treat a prediction as an instruction. If the model says the east-wing restroom will cross its condition threshold around 3 PM, the useful output is not a chart with a rising line — it is a task, assigned to a named person, timed for 2:30 PM.

That is the entire point of the proactive model. Everything else on this page is mechanism.

What the model reads

Three inputs, all of them things that already exist in the operation rather than a separate data-collection project:

Usage. Sensor counts on entries and fixtures give the actual traffic an area took — not the traffic the floor plan implies. Two identical restrooms on two floors routinely differ by a factor of three, and only one of them needs a mid-afternoon visit.

Condition. Readings from the same sensors — dispenser levels, waste volume, ambient signals — say how far the area has drifted from where it should be.

History. What inspections found the last hundred times an area looked like this. This is the input that makes the model specific to your building instead of generic to the industry.

The first two come from IoT sensors and the restroom monitoring layer. The third comes from smart inspections — which is why inspection scoring is not paperwork here. It is the training signal.

The four moves, in order

Prediction on its own is one quarter of the model, and the weakest quarter alone:

1. Risk detection — the pattern that precedes a failure is recognised while there is still time to act on it.
2. Early warnings — the person who can do something is told, before the condition is visible to anyone using the building.
3. Predictive cleaning — the visit is scheduled against real usage rather than a calendar.
4. Proactive response — when something still gets through, the intervention closes it before a complaint is filed.

Each of the four has its own page because each fails differently. A system that predicts well and notifies badly produces the same outcome as no system at all.

Where it is wrong, and how you find out

Every forecast is compared against what the inspection actually found, and the gap is kept. That comparison is the honest part of this page: over a month, some areas are predicted well and some are not, and the ones that are not are usually the ones with irregular use — an event space, a training room booked at random, a lobby that empties for a week.

For those, the model is worse than a supervisor who knows the building. So the supervisor overrides it, and the override is recorded with a reason. We would rather run a model that gets corrected in public than one that quietly averages its errors into a number nobody questions.

What it runs on

The forecast lives inside predictive management in the CleanVision platform — the same system your supervisors and ours open every day. The tasks it generates are ordinary tasks in the same queue as everything else, and the proof that they happened is the same proof of service that covers the rest of the contract. There is no separate AI product to buy, log into, or believe in.