AI agents for property management: leasing to repairs.

02Which property management jobs suit an AI agent first?

The work is a stream of small requests, each needing a lookup in the PMS and an action in it. That suits agents, provided the rules sit outside the model. Not every job needs a language model, though:

The first three rows suit a conversational agent. The rest are better served by other kinds of AI, and a good partner will say so.

03How should an agent handle leasing inquiries and tours?

An inquiry arrives from a listing portal, your website or WhatsApp at nine in the evening. The agent should create or update a guest card in the PMS, answer from the PMS’s own availability, and offer tour slots from the real calendars of your leasing staff.

Three rules belong in code, not in the prompt:

• Price and availability come from the PMS only. The model never types a rent figure that a tool did not return. Where the market bans bidding, the agent must also refuse offers above the advertised rent. In England, from 1 May 2026, the Renters’ Rights Act roadmap says landlords and letting agents cannot ask for, encourage or accept an offer above the advertised rent, or request more than one month’s rent in advance.
• Every prospect gets the same information. The model does not decide who is worth a tour, and it asks everyone the same questions (move-in date, bedrooms, budget against your published criteria).
• Neighborhood questions get facts, not opinions. Walking time to a station is a fact. “A good area for families” creates steering risk.

Commercial leasing has the same shape. The agent answers questions about suites, floor plans and building rules, books viewings with the broker and sends information packs, while terms, incentives and letters of intent stay with a person.

04How do you triage repairs when some are emergencies?

This is the part that should worry you most. A tenant writes “there’s a smell in the kitchen and the boiler is making a noise.” A language model might file it as routine. So the design principle is simple: the model may raise urgency, but only deterministic rules may lower it.

1. A rules layer reads every message first. It matches gas smell, carbon monoxide alarm, flooding, no heat in cold weather, no water, sparking wires, fire, a door that will not lock, a person trapped in a lift, or anyone saying they feel unwell. Any single match makes it an emergency, whatever the model thinks, in every language your tenants write in.
2. The model classifies what the rules did not catch, and extracts the unit, the trade, access details and whether the tenant gives permission to enter. When it is unsure, the request goes to a person as urgent. “Unsure” is never routine.
3. The safety reply is fixed text your team approved, not generated: leave the property, don’t switch anything on, call the emergency number. It goes out within seconds, before anyone is paged.
4. On-call paging has an acknowledgement timer. If nobody confirms within the window you set, it pages the next person and then rings a phone. This is a fallback chain, the same idea as in our piece on designing AI that fails gracefully.

Your own policy defines what an emergency is, and some of it is law. In England, social landlords must investigate suspected emergency hazards and make them safe within 24 hours under Awaab’s Law, and the government’s guidance lists gas or carbon monoxide leaks, broken boilers, total loss of water, exposed wiring and major leaks as examples. It says an emergency investigation can be done remotely and that triage should take the tenant’s circumstances into account. Phase 2 of those rules is due to start on 30 November 2026. In October 2026 the government also opened a consultation on extending Awaab’s Law to private rented homes, open until 18 December 2026, so the rules and timing for private landlords are not yet set.

Gas is its own case in England. Under the Gas Safety (Installation and Use) Regulations 1998, landlords must arrange an annual gas safety check by a Gas Safe registered engineer, and HSE says that where the management contract makes the agent responsible, the agent carries the same duties. An agent can watch certificate expiry dates in the PMS and raise the booking early, but the engineer does the check.

Emergency handling needs its own test set: real past requests, translated and badly spelled, with the safe answer for each. We treat a missed emergency in that set as a failed release, as described in evals before features.

05What approval limits should be in code?

Your management agreement with each owner usually sets how much you can spend without asking. Those numbers must live in data the agent reads, not in a paragraph of instructions it might paraphrase.

• A per-property approval limit. Under it, the agent can dispatch a vendor. Over it, it requests an estimate and sends an owner approval request with the photos and the quote.
• A rolling total, not just a per-job check. Otherwise one large job can become three small ones on the same unit in a week. Count spend per unit and per vendor over a window.
• An owner reserve check. No dispatch if the owner’s balance cannot cover it, unless it is an emergency your agreement allows you to act on.
• Vendor eligibility. Insurance and contract expiry sit on the vendor record in systems such as AppFolio, so the agent should only offer jobs to vendors who are currently compliant.
• Permission to enter, recorded on the work order from the tenant’s own answer.

Every action goes in a log with the rule that allowed it, so when an owner asks why a vendor was sent on a Sunday, the answer takes one lookup.

06How does an agent write back to AppFolio, Yardi, Buildium or Entrata?

The agent is only as good as its integration, and each platform has its own access model. As of October 2026:

Yardi says its maintenance interface lets third parties pull and push work orders, and its guest card interface lets them push leads and update guest cards. AppFolio’s work-order fields include priority, permission to enter and vendor.

A few engineering habits make write-back safe:

• Write structured records, never free text alone. The work order has a category, priority, access instructions and the tenant’s own words attached.
• Read back after writing. Confirm the record exists with the values you sent before telling the tenant it is booked.
• Make writes idempotent. A retried call must not create a second work order and send a second vendor. This is where a lot of first versions fail.
• Verify inbound webhooks. AppFolio signs its webhook notifications and says any that fail signature verification must not be processed.

Treat inbound messages as untrusted input, since anyone can write anything. Our note on prompt injection explains why that matters doubly when the agent can dispatch vendors.

07What can an agent tell tenants, and on which channel?

• Balance and ledger questions need an identity check against the lease on file, so one household never sees another’s details.
• Say it is an AI. The EU’s transparency rules for chatbots have applied since 2 August 2026, and clear disclosure is good practice everywhere.
• WhatsApp has its own rules. Meta’s business policy allows automation inside the 24-hour customer service window as long as there is a clear way to reach a person, and outside that window only approved templates. It also bars wrongful discrimination, including on family status. Meta’s business terms bar general-purpose AI assistants from being the main thing offered through the platform and, where an AI company acts as your service provider, bar using your message data to train AI models other than ones for your own exclusive use. Confirm with your builder how both are handled.
• Voice is different. The FCC ruled in February 2024 that AI-generated voices count as “artificial” under the TCPA, so US calls using one require the called party’s prior express consent (with exceptions for emergencies and exempt calls). Start with text and add voice once consent is recorded properly.

Rent reminders, notices to enter and anything that starts a legal clock go out from your approved templates, not fresh model text.

08What do fair housing and renting rules mean for the agent?

A leasing agent talks to hundreds of strangers, so equal treatment has to be built in. The rules differ by market, and a qualified professional decides how they apply to you. As of October 2026:

• United States. The Fair Housing Act covers race or color, religion, sex, national origin, familial status and disability, and the Justice Department describes false information about availability and steering to certain areas as ways it happens. HUD withdrew its 2024 guidance on digital advertising effective September 17, 2025, but said conduct that does not comply with the Act’s text remains subject to enforcement. In January 2026 it also proposed removing its discriminatory effects rule, so check where that stands, and many states add their own protections.
• Screening. If a decision to deny an applicant, require a co-signer, or raise the deposit or rent is based partly or completely on a consumer report, the FTC says the applicant must get an adverse action notice.
• England. Since 1 May 2026, for assured tenancies in the private rented sector, it is illegal to make a tenant less likely to rent because they have children or receive benefits, and the roadmap says that includes withholding information about a property, with people acting on a landlord’s behalf also liable.
• Gulf. Registration is part of the workflow. Dubai’s Ejari system registers tenancy contracts, including by licensed property management companies, and in Saudi Arabia a residential lease not registered on Ejar is not treated as a valid contract for administrative and judicial effects. The agent can collect the documents and prepare the details, and a person registers. Our guide to building AI in the UAE covers data law and Arabic.

In practice: the same questions for every prospect, no characterizing neighborhoods or residents, no accept or reject message from the agent, and a log of what each prospect was told.

09Where is AI the wrong tool, and where does a person stay in charge?

Some decisions should never be delegated, however good the demo looks:

• Screening outcomes and lease decisions. The agent can collect documents and check completeness. A person applies the criteria and signs off.
• Accommodation requests from disabled applicants or tenants, such as an assistance animal or an accessible unit. The agent logs them with a timestamp and passes them straight to a trained person.
• Arrears, notices, evictions and deposit disputes. These start legal clocks and carry real harm. The agent can fetch the ledger and draft a summary.
• Hazards affecting vulnerable occupants, such as no heat for an elderly tenant. A person triages with their circumstances in view.
• Rent setting. Use a pricing model with a person approving changes, not a chat agent.

Let the model do the reading and the typing. Let your rules, your owners’ limits and your people make the decisions.

10How do you start, and what do you measure?

For most mid-size managers, maintenance intake is the best first project: high volume, easy to measure, and its riskiest case can be tested. Our services page shows how we run such projects. The plan looks like this:

1. Collect 200 to 500 past maintenance requests and label the right category, priority and action. Include the awkward ones: other languages, photos only, vague complaints.
2. Write down your emergency, approval and access rules as data.
3. Connect to the PMS read-only first, then run in shadow mode: the agent proposes, a coordinator confirms, and you compare.
4. Turn on write-back for routine categories once triage matches your coordinators, keeping emergencies and over-limit work with people.
5. Add leasing inquiries on the same foundations, then tenant status messages.

Track a few numbers from the first week:

• Time to first response on requests and on inquiries, including evenings and weekends.
• Emergency recall on your test set, where the target is no misses.
• Time from request to vendor dispatch, and the share of work orders that need a second visit because information was missing.
• Share handled without a person, alongside the rate at which coordinators override the agent.
• Inquiry-to-tour rate, compared with the period before.

11Questions property managers ask

Do we need to switch property management software to use an AI agent?

No. The agent sits beside your PMS and reads and writes through its API or partner interface. Some platforms need an application or agreement first, so check that early.

Can the agent decide what counts as an emergency?

It should help, not decide alone. Emergency rules are written in code from your policy and the law that applies to you, and the model can only raise urgency. Anything unclear goes to a person as urgent.

Can it approve repairs for owners?

Only within limits your management agreement already sets. Below the limit it can dispatch a compliant vendor, and above it the agent sends the owner a request with the quote and photos. Emergency work follows what your agreement allows.

Does this work for commercial portfolios as well as apartments?

Yes, with different content. Commercial agents handle suite availability, viewings, building rules, after-hours access and maintenance requests under lease service levels, while terms and incentives stay with your brokers.