Predictive Cleaning

The schedule was written before anyone measured anything

Every commercial cleaning contract contains a frequency table, and almost every one of them was produced the same way: someone walked the building once, applied a rule of thumb by room type, and wrote it down. It then stayed fixed for years, through a department moving floors, a side entrance opening, headcount changing twice.

The table is not wrong so much as blind. It cannot be right about a building it never measures, and it was never going to be — not because anyone was careless, but because until recently nothing was counting.

Frequency from usage

Predictive cleaning replaces the fixed row with a moving one. Each area has a measured baseline — what its usage normally looks like — and a threshold where its condition starts to matter. The visit is scheduled to land before the threshold, not after a complaint says it was crossed.

In practice this looks unremarkable, which is the point. Nobody sees a model. A supervisor opens the day and the list is already ordered by what the building did yesterday.

The trade nobody else says out loud

Here is the part most predictive-cleaning pitches skip: this reallocates effort, it does not create it.

If the third-floor restroom is getting an extra afternoon visit, something else is getting one fewer — the meeting room nobody booked, the corridor with a tenth of the traffic anyone assumed. Your contracted hours stay your contracted hours. What changes is that they stop being spread evenly across a building that is not used evenly.

A vendor whose predictive model only ever recommends more cleaning is not running a model. That is a sales funnel with sensors attached.

The floor that never moves

Reallocation without a floor is how this goes wrong, so there is a floor.

Every area carries a minimum frequency it cannot drop below, regardless of what usage says. For most spaces that is a contractual baseline. For medical rooms, food-handling areas and anything with a regulatory minimum, it is fixed and the model has no authority over it — it can schedule an additional visit, never remove a required one. This is set once, per account, and it is visible to you rather than buried in a configuration screen.

The rule is the same one that governs every routine with a requirement behind it: where a minimum exists, an algorithm does not get a vote.

When the supervisor is right and the model is not

Irregular spaces defeat this. An event hall, a training room booked at random, a lobby that empties for a fortnight — the model reads the last month and extrapolates, and the last month did not contain the thing about to happen.

Supervisors override in those cases, with the reason attached. Those overrides are read back into the forecast, so the recurring ones stop needing to be made. We would rather have a model that a person corrects than one that is right on average and wrong in the building you actually occupy.

What it runs on

Usage comes from IoT sensors and restroom monitoring; the forecast from AI prediction; the scheduling and the monthly view from predictive management. What was actually done, and when, is recorded the same way as the rest of the contract in proof of service.