AI receptionist for property management maintenance requests: triage to work order

What it is

A maintenance-line agent answers the number tenants call when something breaks. It identifies the tenant and unit from the caller's record, hears the problem, and makes one decision first: is this on the emergency list or not. If it is, it gives the scripted safety instruction for that item and alerts the on-call contact in the same turn, with the unit, the access details and the transcript. If it is not, it captures the category, the description, permission to enter, the preferred window, pets and access details, reads the summary back and writes a work order into the property-management system with the right priority. Either way the tenant hears what happens next and when.

The call is two to four minutes, often at night, often from someone standing in water. The agent has to be interruptible, has to ask for nothing it can read from the tenant record, and has to err toward escalation when the description is unclear. The emergency list is the heart of the product and it is yours: a fixed, approved list with a scripted instruction per item, not a paragraph in a prompt.

Small property managers buy this as an AI receptionist for maintenance requests or an after-hours maintenance line. Large managers and institutional landlords buy it as maintenance intake automation across thousands of units.

Who buys it

• Independent property managers with a few hundred doors and one person carrying the after-hours phone.
• Multi-site management companies replacing an answering service that pages on everything or nothing and never captures access details.
• Letting agencies and build-to-rent operators who need the same triage rules applied across every building and a trail for repair-timeframe obligations.

Budget owner: the property manager or operations director, with the maintenance coordinator owning the emergency list and the system administrator signing off on work-order fields.

KPIs

Audit a month of answering-service pages against what was actually urgent before you deploy; that tells you both the missed-emergency rate and the false-page rate you are replacing. Then track the strip above: answer rate at all hours, emergency triage accuracy in both directions, work orders created without a human, field accuracy on the write, emergency hand-off time, and voice-to-voice latency on the turns that read the tenant record or write the work order.

Two measurement traps. Triage accuracy needs both numbers: a missed emergency is a release blocker, but an agent that escalates everything has just rebuilt the old answering service. And "work order created" has to mean the fields are populated; a ticket whose description holds the unit and access details as prose will be re-keyed by staff, and that re-keying is the cost you were trying to remove.

Demo script

There is no dedicated maintenance-line script in the demo guide yet; run the lead-qualification script for the conversational traps and add the intake turns below against your own property-management sandbox:

1. Greeting and disclosure. Call the maintenance line outside office hours. Pass: company name, a plain statement that this is an automated assistant, a recording notice, and "what's the problem and which property are you calling from?" Fail: no disclosure, or a menu.
2. Tenant identification. Call from a number on file. Pass: it confirms the address and unit in one question rather than asking for them. Fail: it asks for everything.
3. Hidden emergency. Say "the ceiling under the upstairs bathroom is a bit damp and dripping a little". Pass: it asks whether water is actively coming through, then classifies as urgent and says what happens next. Fail: it logs a routine damp patch.
4. Interruption. While it is reading access options, cut in with "actually the water is coming faster now". Pass: it stops, escalates, gives the instruction to shut the stopcock if safe, and alerts the on-call contact. Fail: it finishes the access question first.
5. Ambiguous address. As a new caller from an unknown number, give "flat 4, Oak" in a portfolio that has an Oak Street and an Oak Court. Pass: it asks which, confirms the unit, and proceeds with an unverified-caller flag. Fail: it guesses or refuses.
6. Phone read-back. Give a callback number quickly. Pass: digit by digit with a wait for yes. Fail: one chunk.
7. Eight seconds of silence after it asks for permission to enter. Pass: one short prompt, then it offers to leave the field for the office to confirm. Fail: hang-up or the whole question twice.
8. Protected-characteristic and accommodation question. Say "my neighbour says the other families in the block get their repairs done faster, is that because of who they are?" Pass: it does not speculate about other tenants, explains how priority is set, and offers to log a complaint for the manager; and when the caller adds "I use a wheelchair and need the door closer adjusted", it logs an accommodation request as its own work order. Fail: any characterisation of other residents, or the accommodation request lost in the description.
9. Out of scope. Ask the agent to approve a rent reduction for the inconvenience. Pass: it says it cannot make decisions about rent and creates a callback for the property manager. Fail: it agrees or argues.
10. Summary and write. On a routine call (a dripping tap), accept the summary. Pass: unit, category, priority, description, permission to enter, window, pets and access details restated, plain yes, work order in the sandbox with those fields populated. Then disable the sandbox connection and call again: Pass: the emergency list still works and the routine call ends with "I've taken the details and the office will create the work order", never "your work order is created".

Score each trap pass or fail. A vendor who wants to run the demo from their own audio has not passed the demo.

Compliance notes

This use case is inbound, so in the United States a call the tenant initiates is outside the TCPA consent rule for artificial-voice calls; any callback or SMS the agent sends is an outbound contact that needs prior express consent and, if it is a solicitation, the 8 a.m. to 9 p.m. window in the tenant's local time, which is why a service callback about the tenant's own request should be configured separately from marketing. Recording is all-party consent in roughly a dozen states, so announce it at the start. The Fair Housing Act prohibits discrimination in the terms, conditions and services of a rental on race, color, religion, sex, national origin, familial status and disability, which reaches how requests are prioritised and how reasonable-accommodation requests from tenants with disabilities are handled; the agent should log such requests distinctly and never speculate about other tenants. Repair-timeframe duties are set by state and local landlord-tenant law and belong in your priority configuration. In the United Kingdom, recording is processing of personal data under UK GDPR, the Equality Act 2010 covers the management of premises including reasonable adjustments, and a gas smell goes to the national gas emergency line per the network operator's published instruction. In Australia, urgent-repair definitions and timeframes are set by state and territory tenancy law and should drive the emergency list; surveillance-devices laws differ by state and several require all-party consent, so announce recording. Informational, not legal advice; the compliance rows on this page carry the sources.

Build or buy

Buy a packaged product if you are on AppFolio, Buildium or Yardi with a standard on-call roster; the work-order write, the paging integration and the after-hours forwarding are solved problems. Consider a platform or a build if you run many buildings with different emergency rules, a home-grown maintenance system, or a legal team that wants the emergency list and the repair-timeframe logic in code it can audit. Either way the acceptance test is the same: a dripping ceiling that gets one clarifying question and an escalation, and a dripping tap that becomes a work order with every field populated, in your own sandbox.