FR Book 30 min

Home / Client cases / Boiler room survey

Heating company · around 90 people

From boiler room photos to a priced maintenance contract.

A heating company takes over buildings for maintenance. The technician photographs the boiler room, the AI fills in the report, and the maintenance contract is priced. Validated on three real sites, go-live in progress.

0serial numbers, references or brands invented on the sites tested
76 to 81%accuracy against the expert's report, on what photos can show
to the centthe maintenance contract price reproduced on the replayed site
The flow
  1. Boiler room photosThe technician follows a photo-taking guideline.
  2. AI readingNameplates and components: boilers, pumps, vessels, valves, meters.
  3. ReportStructured takeover report; an unreadable field stays empty.
  4. Pricing enginePrice calculated without AI, from the company's rules.
  5. Maintenance contractStandard clauses assembled, editable document.
From the takeover visit to the maintenance contract.

The context

To take over a building for maintenance, a salesperson and a technician list the equipment on paper, then a maintenance contract is written from those notes. The team itself estimated a building takeover at one hour for sales and two to three hours for the technical side, and writing a contract at ten to twenty minutes, with double data entry.

Choosing the use case

An AI diagnostic produced a prioritisation matrix of eight use cases, ranked by technical feasibility and estimated impact. Management picked one: the takeover chain, from survey to contract.

What was built

  • The technician photographs the boiler room. The AI reads the nameplates and components (boilers, pumps, expansion vessels, valves, safety valves, meters) and fills in a structured report.
  • A pricing engine with no AI and standard clauses assemble the maintenance contract, delivered as an editable document.
  • A photo-taking guideline for technicians: photo coverage turned out to be the real limiting factor, ahead of the model.

How the prototype was tested

Blind generation: the AI does not see the report of the company's expert. Its output is frozen before that report is opened, then graded on a grid set in advance. On three real sites, accuracy ranges from 76 to 81% on what a photo can show, with no invented value; on the site replayed end to end, the contract price lands to the cent. The rest of the report (measurements, observations) relies on voice, planned for the next step.

Takeover reportFictional example
ComponentRead from photoStatus
Boiler 1BRAND A · 130 kW · 2014Read
Boiler 2BRAND A · 130 kW · 2014Read
Primary pumpBRAND B · twinRead
Expansion vessel80 LRead
Safety valveunreadable (glare)To complete
Heat meternot photographedMissing

Maintenance contract

Visits per year2
Annual fee excl. VAT€1,840.00
Fictional data. What cannot be read is flagged, never guessed.

Where the project stands

The prototype is validated end to end on real data. Go-live is in progress.

Further reading