One of the projects we get asked about most often, especially by lettings firms and property management companies, is some version of a maintenance triage agent. The business problem is universal: tenants report problems, the property team gets buried, the right contractor takes too long to dispatch, and everyone ends up frustrated.
Rather than another generic "what could AI do for property" article, this is a detailed walkthrough of what we would actually build for a typical UK property firm. End to end.
The starting point
Picture a fairly standard UK property management business. Around twelve hundred properties under management. A team of eight. Maintenance reports arrive through email, the website portal, WhatsApp, and the occasional Saturday afternoon phone call. Most weeks, the team handles roughly a hundred and fifty maintenance reports.
The current process: a property manager reads each report, decides if it is urgent, looks up the building, identifies the right contractor, drafts a job order, sends the contractor an email or WhatsApp, and replies to the tenant with a timeline. About thirty to forty minutes per report, including the inevitable back and forth.
Total: roughly seventy to a hundred staff hours a week on maintenance triage alone.
What the agent does
The build is a triage agent that watches the inbound channels, processes the report, and produces a queued action for human approval.
For each maintenance report, the agent:
- Reads the message, in whatever channel it arrived.
- Identifies the property, the tenant, and the issue. Pulls the property record from the management system.
- Categorises the issue by type and urgency. Boiler in winter is urgent. A flickering bulb is not.
- Looks up the right contractor for the type of work and the location, with availability.
- Drafts a job order with the property details, access notes, contractor brief, and target completion window.
- Drafts a tenant reply confirming receipt and a realistic timeline.
- Queues both for human approval before anything goes out.
A property manager opens the queue, scans the agent's work, approves or edits, and presses send. What used to be thirty to forty minutes is two minutes.
The build, week by week
For a project like this, the typical timeline is six weeks.
Weeks 1 and 2: Discovery
Sit down with the property managers. Watch them work. Capture the rules they apply that nobody has written down. Understand the contractor list, the urgency definitions, the typical access arrangements. Get realistic examples of the last two months of maintenance reports.
Output: a clear specification, agreed in writing, of what the agent will do and what success looks like. Plus the test data the agent will be evaluated against.
Weeks 3 and 4: Build
Connect to the property management system, the email channels, WhatsApp Business and the contractor database. Build the categorisation logic, the contractor matching, the draft generation. Set up the human approval queue. Wire in logging.
Output: an agent that can process a real report end to end, in a test environment, with a human-facing approval queue ready to use.
Week 5: Internal testing
Run the agent against the last month of real reports, in shadow mode (no actions taken, just outputs reviewed). Compare the agent's decisions against what humans actually did. Tune the categorisation, tighten the prompts, fix the edge cases.
Output: confidence that the agent is making sensible decisions on real data.
Week 6: Soft launch
Go live for one property manager's portfolio first. Every output reviewed by the human before it goes out. Daily check-ins on the queue. Weekly tuning. Once that property manager is comfortable, roll to the rest of the team.
Output: a live agent, in use, with real metrics on time saved and decision quality.
What gets integrated with
For a typical UK property firm, the integrations look something like this.
- Property management system (Reapit, Alto, Jupix, Goodlord or similar).
- Inbound email channels.
- WhatsApp Business via the official cloud API.
- Contractor database (often a spreadsheet for smaller firms, a proper system for larger ones).
- Calendar for follow-up scheduling.
None of this is exotic. All of it has working APIs in 2026.
The point of a maintenance agent is not to take humans out of the loop. It is to give the property team the next two minutes of work already done, so they can focus on the conversations that need a person.
What it costs and what it saves
For the kind of project described above, the build cost typically lands in the £10,000 to £15,000 range. Running cost is a few hundred pounds a month, mostly in AI usage. Light maintenance is another few hundred a month if you want it.
The time saving for a firm doing a hundred and fifty reports a week, at thirty minutes saved per report, is around seventy-five hours a week. Even if you discount that aggressively, the build typically pays for itself within two to three months.
The harder-to-measure benefits are the ones that matter as much: tenants get faster, more consistent responses; contractors get better-briefed jobs; property managers stop being buried in admin and have time for the calls that need them.
What we will not pretend
The agent will get things wrong sometimes. It will misclassify the occasional report. It will pick the wrong contractor on rare occasions. The whole point of the human approval queue is that those mistakes get caught before they reach a tenant or a contractor.
Over time, with weekly tuning, the error rate drops to a level where the human review for the easy cases can be relaxed. The harder cases keep human approval. This is normal. This is the design.
If this sounds like your business
If you run a UK property firm and the description above sounds painfully familiar, this is exactly the kind of build we love doing. Tell us about your current maintenance process and we will come back with a real estimate. Or, if you would like a wider look at where AI fits across your operation, our two-week strategy audit is the right starting point.
This is one of several anatomy pieces. The rest are in the blog.