AI for facilities management

AI for facilities management: triage the work orders, then predict the breach

Facilities management is a business of thousands of small decisions a day, and almost all of them are made with information that already exists somewhere in the system.

Work order triage is the obvious start

Requests arrive as half a sentence from a phone: no heating in the east wing, water coming through the ceiling near reception, door not closing. A helpdesk operator reads each one, guesses the trade, guesses the priority, picks a contractor and hopes. Misrouting is expensive twice, once in the wasted visit and once in the SLA clock that kept running.

A model trained on your own closed work orders does this better than a new starter, because it has seen how similar wording resolved thousands of times. It proposes trade, priority and likely parts, with the historic jobs it is reasoning from shown alongside. The operator confirms. Keep the confirmation step: it is what stops a wrong category becoming a wrong contractor on site.

Predicting the breach before it happens

Most FM contracts pay penalties on SLA breaches that were visible days earlier. The signals are already in the system: a job reassigned twice, a subcontractor whose average response time in that region has been drifting, a part with no stock, a site where the same asset has failed three times this quarter. None of these is hard to spot individually, and nobody has time to spot all of them across a portfolio.

A daily scoring pass over open jobs, surfacing the twenty most likely to breach with the reason attached, is a small build with a directly measurable outcome. You know your breach numbers today, so you will know within two months whether it worked.

The asset register is the real project

Every ambitious FM AI project runs into the same wall: the asset data is wrong. Duplicated records, assets that were decommissioned years ago, serial numbers in a notes field, three spellings of the same manufacturer. No model fixes a decision made from bad asset data.

Cleaning it is unglamorous and it is the highest-leverage work available. A model helps here too, matching duplicates, extracting model and serial numbers from historic job text, and reconciling what engineers actually reported against what the register claims. Treat it as a project with its own scope, not as something that will happen incidentally.

Subcontractor invoice matching

Invoices come in against work orders, and checking them line by line against the agreed schedule of rates is nobody's favourite job, so it is often sampled rather than done. A matching pass compares each line against the work order, the rate card and the time recorded on site, and flags only the exceptions. Recovered value here tends to pay for the rest of the programme, which makes it a useful second build once triage has earned trust.

The sequence that works

1. Work order triage, because it is high volume and immediately visible.
2. Asset data cleanup, because everything downstream depends on it.
3. SLA breach prediction, once the data is trustworthy enough to act on.
4. Invoice matching, which funds the next phase.

Working out which of these your current systems can actually support is what an AI audit settles before anyone commits budget.

Frequently asked questions

Related