Short answer: To triage tenant maintenance requests with AI, write down your own emergency rules first, then let an AI intake assistant ask the follow-up questions, collect photos, give safe first steps (shut off the water, flip the breaker) and sort each request into emergency, urgent or routine. True emergencies go straight to your on-call person or approved vendor; everything else becomes a clean work order for the morning. The AI does the questioning and summarising, while your rules and your people decide what gets a van out at 2am.
The problem
Anyone who manages rentals knows the pattern. A tenant calls at 11pm. It might be a burst pipe, or it might be a dripping tap that has been dripping for a week. You can't tell until you have asked four or five questions, and by then you are awake.
A recent BiggerPockets thread (October 2026) asked self-managing landlords exactly this: do you pick up at night, how do you decide what counts as an emergency, and what has gone wrong. The replies were practical and fairly consistent. People define emergencies in writing in advance: water that won't stop, no heat in freezing weather, whole-unit power loss, gas smells, lockouts. One reply noted that tenants often call back simply because nobody told them what happens next, so an automatic acknowledgement helps. Another asks for a photo before deciding, because a picture separates a slow drip from a flood in seconds. And a few set a not-to-exceed amount so a trusted plumber can stop active damage without a phone call to the owner.
None of that is complicated. The trouble is that it depends on a tired person running the same script, correctly, at any hour, and writing it all down afterwards.
What it costs
There are two costs. The first is staff time and on-call fatigue: every non-emergency call that wakes someone up is paid for in sleep, overtime or an answering-service fee. The second is tenant retention. In AppFolio's 2026 renter survey of 3,002 US renters, renters satisfied with maintenance were 81% more likely to renew, and maintenance had the strongest link with overall satisfaction of anything the survey measured. Among satisfied renters, 79% said their issues were resolved within a few days or less.
In the UK there is also a legal edge for social landlords. Under Awaab's Law, which applies to hazards reported on or after 27 October 2025, landlords must investigate and make safe emergency hazards within 24 hours. Shelter's summary notes that the clock runs from when the landlord becomes aware of the potential hazard, and that private rented homes are not covered. If a report sits in a voicemail box until Monday, that clock has already been running.
Why it keeps happening
Tenants are not the problem here. Most people have no idea whether a smell of damp, a tripped breaker or a warm fridge counts as urgent, so they call. They also call through whatever channel is closest: the office line, a personal mobile, a text, an email, the portal. Requests arrive in five places with five levels of detail.
On the management side, the emergency rules usually live in someone's head. A new hire or a weekend cover person applies them differently. Notes from a 2am call are scribbled or forgotten, so the morning team starts from scratch and calls the tenant back to ask the same questions.
Answering services help with coverage, but many of them are reading a short script and passing everything on, which means the on-call person still does the real triage. Portals help with structure, but tenants in a panic tend to phone rather than fill in a form.
How AI solves it
The useful version is an intake assistant that sits on your phone line, text number and portal, follows your rules, and hands a tidy summary to a person or a vendor. Here is the workflow step by step.
- Receive the request on any channel. A voice agent answers calls; the same logic handles texts, emails and portal messages. Everything lands in one queue.
- Ask the right follow-up questions. This is where a language model earns its place. It reads a free-text message like "kitchen ceiling is wet and the light is flickering" and asks what a good coordinator would: is water actively coming through, is it near the light fitting, can you turn the water off, can you send a photo.
- Give safe first steps. From a script you approve, it tells the tenant how to stop the immediate damage: where the stopcock or main shutoff usually is, to switch off the circuit at the breaker, to leave the building and call the gas emergency line if they smell gas.
- Classify against your rules. The model proposes a category (emergency, urgent, routine) and the reason. Then plain rules check it: certain keywords and answers (gas, sparks, no heat below a set temperature, water still flowing) force an emergency regardless of what the model thinks.
- Route. Emergencies trigger a call or text to the on-call person, or straight to an approved vendor within a spending limit. Urgent jobs go to the top of tomorrow's list. Routine requests become work orders in your property management system.
- Acknowledge and log. The tenant gets a message saying what happens next and roughly when. Your team gets a short summary with photos, the tenant's answers, the category and the time each step happened.
AppFolio now offers a version of this inside its own platform: its Realm-X Maintenance Performer walks residents through simple fixes before creating a work order, classifies emergencies "based on the rules your team sets", and routes vendors by issue type, severity and time of day. That is a good sign the approach works, and if you already use AppFolio it is the first thing to try.
Where the LLM helps and where rules are better
| Step | Best tool | Why |
|---|---|---|
| Understanding a messy description | LLM | Tenants describe the same leak a hundred ways |
| Asking follow-up questions | LLM, from your question list | Needs to adapt to the answers |
| Reading a photo (is that a drip or a flood?) | Vision model, with a person checking | Useful hint, not a final decision |
| Deciding "emergency" | Rules first, LLM second | Gas, fire and flooding must never depend on a guess |
| Safety advice | Fixed approved script | Never let the model improvise safety instructions |
| Vendor dispatch and spending limits | Plain automation | Simple, auditable logic |
| Writing the work order summary | LLM | Saves the morning team a call back |
The design principle is that the model can escalate on its own, but it should never be able to downgrade something your rules call an emergency. When it is unsure, it treats the request as urgent and wakes a person, because an unnecessary phone call is a far smaller cost than a missed gas leak.
Costs and timeline
These are estimates, and your numbers will differ. The assumptions are stated so you can swap in your own.
Running cost
Model usage is the cheap part. Assume each request involves about 3,000 input tokens and 500 output tokens across the conversation. At Anthropic's published rate for Haiku 5.5 ($0.10 per million input tokens and $0.50 per million output tokens for prompts under 100K tokens), that is roughly $0.0005 per request. Even at 1,000 requests a month, model costs would be well under $5. A larger model for harder cases, such as Sonnet 5.5 at $2 and $10 per million, would come to roughly $11 for the same 1,000 requests. Voice calls add telephony and speech costs on top, which usually outweigh the model cost.
Time saved (estimate)
Assume a 400-unit portfolio produces 120 after-hours requests a month, that 15% are true emergencies, and that each non-emergency call currently takes 10 minutes of on-call time plus 5 minutes the next morning to re-contact the tenant. That is about 102 non-emergency calls, or roughly 25 hours a month of interrupted time. If the assistant handles intake and the morning follow-up disappears because the summary is complete, most of that time comes back. The emergency share in your own logs is the number to check first, since it drives the whole calculation.
Build, buy or do nothing
| Option | Rough cost (estimate) | Time to live | Good fit when |
|---|---|---|---|
| Do nothing | Existing on-call pay and answering-service fees | None | Very small portfolio, few calls |
| Buy: feature in your PM software or an off-the-shelf AI answering tool | Usually a monthly subscription, priced per unit or per call | Days to a few weeks | Your system has a native option and your rules are standard |
| Build: custom intake agent connected to your systems | Typically a few weeks of build, then hosting and model costs | 4 to 8 weeks for a first version | Mixed portfolios, several systems, unusual rules, or UK compliance logging |
A custom build makes sense when you run more than one property system, manage both residential and commercial units with different rules, need multilingual intake, or want an audit trail shaped around obligations like Awaab's Law. Otherwise, buy first.
When not to use AI
If you manage a handful of units and get two night calls a month, a written emergency policy, an automatic text reply and a named on-call person will do most of the job for free. That advice came up repeatedly in the forum thread above, and it is right.
Don't let AI make life-safety decisions on its own. Gas, carbon monoxide alarms, fire, sparking electrics and serious flooding should be hard-coded rules with fixed instructions and an immediate human call. The model's job in those cases is to recognise the situation quickly and get out of the way.
Be careful with vulnerable tenants. Elderly residents, people with disabilities and anyone who struggles with a voice bot should always have a simple route to a human. Offer "press 0" or "reply HUMAN" and make sure it works at 3am.
And if your current problem is that vendors don't turn up, AI intake won't fix it. Better triage gets the right job to the right vendor faster; it can't make a contractor answer the phone.
How to start
- Pull three months of maintenance requests and tag each one: emergency, urgent, routine. Note the hour it arrived and the channel.
- Write your emergency rules as a list a stranger could follow, including temperature thresholds and the safe first step for each case.
- Set spending limits for approved after-hours vendors and confirm their insurance is current.
- Check whether your property management software already has an AI intake feature. If it does, pilot it on one building.
- If it doesn't, or your rules are unusual, map the workflow and test it on past requests before any tenant talks to it.
Our AI workflow automation service builds this kind of intake agent, with rules, escalation and logging connected to the systems you already run. If you are weighing a voice or chat assistant more broadly, our guides on what an AI chatbot costs to build and the difference between AI agents and chatbots cover the background.
FAQ
Can AI answer after-hours maintenance calls for property managers?
Yes. Voice and text assistants can answer calls, ask follow-up questions, collect photos and create work orders at any hour. The safe setup is one where your written emergency rules decide what gets escalated, and any doubt is passed to a person on call.
How does AI decide if a maintenance request is an emergency?
A language model reads the tenant's description and answers, then proposes a category with a reason. Fixed rules sit on top: anything involving gas, fire, sparking electrics, water that won't stop or no heat in freezing weather is treated as an emergency automatically. The model can raise a request's priority but should not be allowed to lower one your rules flag.
What counts as an emergency maintenance request?
Most landlords include active water leaks or flooding, gas smells, fire or smoke, electrical hazards, loss of heat in cold weather, sewage backups, total loss of power or water, and a home that cannot be secured. In England, social landlords also have to meet the 24-hour emergency hazard deadline under Awaab's Law. Write your own list down and share it with tenants at move-in.
How much does AI maintenance triage cost?
Model usage is small, often a fraction of a cent per request with an efficient model. The larger costs are the software subscription if you buy, or the build and integration work if you have it made, plus telephony for voice calls. Compare those against your current answering-service fees and on-call hours.
Does AppFolio have AI for maintenance requests?
AppFolio offers the Realm-X Maintenance Performer, which handles maintenance intake, suggests simple fixes, classifies emergencies using rules your team sets and texts vendors from the work order. If you already use AppFolio, it is worth testing before building anything custom.
Sources
- BiggerPockets forum, "How do you handle after-hours maintenance calls from tenants?" (October 2026): https://www.biggerpockets.com/posts/7280332
- KEYT / Stacker, "What 3,002 renters say matters when free rent offers end" (AppFolio 2026 renter survey, September 2026): https://keyt.com/stacker-real-estate/2026/09/12/what-3002-renters-say-matters-when-free-rent-offers-end/
- GOV.UK, "Awaab's Law: guidance for tenants in social housing": https://www.gov.uk/government/publications/awaabs-law-guidance-for-tenants-in-social-housing/awaabs-law-guidance-for-tenants-in-social-housing
- Shelter England, "Hazards in social housing": https://england.shelter.org.uk/professional_resources/legal/housing_conditions/responsibility_for_repairs/hazards_in_social_housing
- AppFolio, "Realm-X Performers" product page: https://www.appfolio.com/product-demos/realm-x-performers
- Claude API pricing: https://claude.com/pricing
If you'd like to talk through how maintenance requests reach your team today and whether automating intake would pay off, get in touch at https://sintraflow.com/contact.
