Risk Detection

The failure was visible for three days

Walk back from almost any cleaning complaint and you find a trail. The restroom that generated the email had been running low on supplies since Monday. The lobby floor that looked bad on Thursday had taken a fortnight of construction traffic through the side entrance. The smell someone finally reported had been mentioned, in passing, twice.

In the reactive model none of that is information, because there is no mechanism that collects it. The complaint is the first data point. Everything before it happened in a building nobody was reading.

Risk detection is the mechanism that reads it.

Four kinds of signal

Drift against an area’s own baseline. Not against a target — against itself. A dispenser that normally lasts nine days and is on pace for five has told you something, even though nothing has crossed a threshold and nothing is wrong yet.

Recurrence in inspection history. The same finding, in the same area, three inspections running. Individually each is minor and gets closed. Together they are a cause that nobody has addressed, and treating them as separate incidents is how a building stays broken for a year.

Consumption anomalies. Supplies emptying far faster or slower than the area’s traffic explains. Faster usually means traffic nobody re-scoped for; slower sometimes means a dispenser that is jammed, or an area people have stopped using.

Known load ahead. Term start at a school, a season change, an event on the calendar, a floor being renovated. This is the one that requires a person to enter what they know, and it is worth the small effort — predictable spikes are the cheapest risk to get right.

Ranked by consequence, not by count

Any system that scores risk purely on likelihood will bury you in high-probability irrelevance. The dispenser in a back corridor used by four people will reliably run low, and reliably matter less than the one in the main entrance where your clients arrive.

So the ranking asks a second question: if this goes wrong, what does it cost? A medical facility’s exam room, a school restroom during term, a lobby the day of a client visit — these carry weight that a raw probability score cannot see. The weighting is set per account, with the supervisor who runs it, because the answer is different in a warehouse and a clinic.

Detection ends in an owned task

A risk that is detected and not assigned is a liability. It creates a timestamped record that the condition was known, with nothing attached showing anyone acted — which is materially worse than not having looked.

So detection terminates in exactly one place: a task in the corrective actions queue, with an owner and a deadline, or a visit scheduled through predictive cleaning. What reaches a human as an interruption is the narrow subset covered by early warnings; the rest is absorbed by the schedule without anyone’s day being broken.

What it reads from

The readings come from IoT sensors and restroom monitoring. The history comes from smart inspections and the scored checklists in digital checklists. The ranking and the task creation happen in predictive management.