AI for Property Management in 2026: What to Automate First, and What to Keep Away From a Model

AI for property management: what to automate first, and what to keep away from a model

If you run a property management company, your software vendor has probably added AI to something this year. So has every startup in your inbox. Some of it saves real hours, and some of it could put you on the wrong side of a housing regulator.

We're an AI engineering company. We build AI systems for clients and run our own product in production, a website assistant called ConGreeto, so read this knowing we sell in this space. This is the guide we'd want an operator to read before talking to us or anyone else: where AI pays off in property management, where it doesn't belong, and when to buy instead of build.

The gap: faster isn't the same as more profitable

IREM and AppFolio published their latest research on 5 November 2026, based on a survey of 3,662 real estate management professionals fielded from 7 May to 7 June 2026. AppFolio sells property management software, so bear that in mind, but the numbers aren't all flattering:

58% now use AI, up from 21% in 2023.
Among AI users, 85% report productivity gains, but only 48% report a positive impact on NOI or profitability.
Only 8% of organizations have a formal written AI policy.
74% of professionals have had no formal AI training.

(IREM and AppFolio release, 5 November 2026; AppFolio summary of the report. We read the release through a syndicated copy, and we haven't seen the full report's workflow tables.)

Two other surveys point the same way. Buildium's 2026 industry report says AI adoption jumped from 20% in 2024 to 58% in 2025, while only 8% of companies have been able to fully automate any process (Buildium). And in AppFolio's survey of 1,617 US residential property management professionals, 78% said they can't yet rely on the AI features in their legacy software (AppFolio, February 2026).

These surveys don't tell us why the gap exists. Our reading, from building these systems, is simple: saving someone 20 minutes on an email doesn't change your NOI unless the time goes somewhere that does. So tie every AI project to a number your business already tracks, such as days vacant, renewal rate, doors per manager, owner churn or maintenance cost per unit. If you can't name the number, you may get the productivity gain without the profit.

How to decide what to automate

AI is a good fit for a task in a property management company when the task is:

Repetitive. The same kind of request, many times a week.
Tied to a number. You can say what should change if it works.
Checkable. A person, or a hard rule, catches a wrong answer before it causes harm.
Not a decision about a person's housing. It doesn't decide who gets a home, who pays what or who gets evicted.
Clean on data. You know where tenant and owner data goes, and one owner's data can't leak to another.

Run every proposal through those five. Most of what goes wrong fails 3 or 4.

What to automate first

1. The first reply to enquiries

Renters compare several options, usually online. In Zillow's 2025 rentals report (more than 36,000 US renters, fielded March to July 2024), the typical recent renter took just one in-person tour (Zillow). So much of the shortlisting happens before anyone speaks to you.

Answering questions from your own listings, capturing the renter's budget and dates, and routing owner enquiries separately ties directly to days vacant and new doors. It's also often a buy, not a build. Our sister product's team has written the detailed version: what a property management website chatbot should answer, and what it must never say.

2. Maintenance intake and triage support

Buildium reports that 38% of rental owners name maintenance as their top source of stress, and that among renters unsure about renewing, 31% would renew if their manager were more responsive to maintenance requests (Buildium, December 2025; the article gives no sample sizes).

Turning "the thing under the sink is doing that noise again" into a structured ticket (location, trade, access, photos, likely urgency) is a good job for a model, because a coordinator checks the result. Set the safety rules outside the model. Certain words, such as gas, carbon monoxide, smoke or flooding, should trigger a fixed response and a human escalation every time. They should never depend on what a model decides that day.

3. Search over your own documents

Leases, management agreements, house rules, vendor contracts, past correspondence. A retrieval system (RAG) that finds the right clause and links to the source can save staff a lot of searching. The person still reads the clause before acting on it. The hard part is permissions, which we'll come to.

4. Drafts a person sends

Replies to online reviews, renewal outreach, routine tenant notices, explanations of an owner statement. In AppFolio's summary of the research, respondents were relatively comfortable with AI drafting routine messages and far less comfortable letting it near owner and investor communication (AppFolio). That instinct is sound. Draft for owners if it helps, but have a person send.

5. Reporting and exceptions

Summaries of the portfolio that highlight what needs attention: units vacant for longer than usual, work orders stuck for more than a week, renewals coming up with no conversation started. The figures should come from your system of record. The model only writes the summary, so it never has to calculate them.

What to keep behind a person

Screening decisions. Using a model to approve or reject applicants is where the legal risk is highest. In Louis v. SafeRent Solutions (D. Mass.), applicants alleged that a tenant screening score disproportionately harmed Black and Hispanic applicants using housing vouchers. The court gave final approval to a $2.275 million settlement on 20 November 2024, with injunctive relief (Cohen Milstein, case page). Those were allegations, and the settlement wasn't a ruling on them.

The regulatory picture has moved, but not in a way that removes the risk. HUD withdrew a set of fair housing guidance documents effective 17 September 2025, and its notice says actions that don't comply with the text of the Fair Housing Act "continue to be subject to enforcement" (HUD, 6 April 2026). In January 2026 HUD proposed removing its disparate-impact regulations, which trade group NAA notes wouldn't eliminate the liability itself (NAA). Check where that rulemaking stands, along with your state's laws. If you use consumer reports, the FTC also spells out what you must tell an applicant after an adverse decision, including when the report played only a small part (FTC). "The model decided" won't satisfy any of this.

Rent set from competitors' private data. In November 2025 the Justice Department filed a proposed settlement with RealPage. It would stop the software using competitors' nonpublic data to set rents in live operation, limit training to nonpublic data at least 12 months old, and put a court-appointed monitor in place (DOJ, 24 November 2025). Pricing software isn't banned, but ask any vendor exactly what data trains its model and whose it is. Get legal advice before you pool data with other operators.

Prices that don't match your total. The FTC's $24 million Greystar settlement concerned advertised rents that allegedly left out mandatory fees, and a week later the FTC warned 13 property management software providers about tools that may hinder accurate pricing in listings (FTC, 2 December 2025; FTC, 9 December 2025). If a model writes listings or answers rent questions, the total price should come from one system of record, not from the model's reading of a page.

Fraud checks as the only check. In AppFolio's February survey, 56% of managers reported facing application fraud in the past year. In the November research, 93% said they were concerned about AI-generated fraudulent applicant documents. AI can flag documents for review, but a flag isn't proof. Falsely accusing a real applicant is a problem of its own, so a person decides.

Legal notices and anything to do with eviction. Notices, deadlines and dispute replies can be drafted from approved templates with legal review. Don't let a system send them on its own.

Agents that act without approval. Issuing work orders above a spending limit, messaging owners, changing rent, adjusting a ledger. Draft, yes. Act, only inside limits you set, with a person approving anything above them.

If you build: what the system must do

This is the part generic tools tend to get wrong in property management, and it's where a custom build can be worth the money.

Permissions checked at the moment of use. Owner A must never see Owner B's statement, and a tenant must never see another tenant's file, not even as a snippet inside a generated answer. Every record carries who may see it, and that check runs each time it's retrieved, before the model sees it. A permission recorded once goes stale the day a property changes manager or a tenancy ends. We've found this class of bug in our own product more than once (a permission checked when something was saved, not when it was used), so we now design for it from the first day.

Safety rules outside the model. Emergency keywords trigger a fixed message and a human alert through a channel that wakes someone up, such as a phone call or an on-call service. Email won't do. Our own product's lead emails, for example, arrive after a conversation ends, which is fine for a sales lead and wrong for a gas leak. Test what happens when the model provider is down, too. The emergency path shouldn't depend on it.

A paired test set for fair treatment. Before launch and after every model change, send pairs of prompts that differ only in a protected trait ("a family with two children" and "two adults", for example), and compare the answers. Add the questions that invite steering: "Is this a safe area?" "Who lives there?" A test like this won't prove compliance, but it will catch failures you can fix. We use a similar approach when replacing models: when we tested replacements for GPT-4o on our own product, we ran 22 cases built from real prompts.

Tests for silent failures. We've written up five bugs from our own product that raised no error at all. In one, some lead notification emails arrived without the visitor's email address, even though the visitor had typed it into the chat. In maintenance, the equivalent is a ticket created with no address on it. So test the output a person actually receives, not just whether the job ran.

An audit trail. What came in, what the system retrieved, what the model returned, who approved it and what was sent. When an owner, tenant or regulator asks how a decision was made, you can show them.

Integration through the systems you already use. AI that can't read or write your property management software creates a second place to keep up to date. Before you commit, check what your platform's API actually allows. Reading data is often easy, and writing it back is where projects stall. Ask for this to be proven with your own data early in the project.

Model change without surprises. Models get retired. OpenAI is shutting down GPT-4o and GPT-4.1 in its API on 14 January 2027, for example, and a system built around one model needs its test set ready before that happens.

A policy before a pilot

With only 8% of organizations reporting a formal written AI policy, this is the cheapest thing on the list. One page is enough:

Which tools are approved, and what data may go into them (tenant and applicant data especially).
Which decisions AI may never make alone.
Who checks output, and who owns a mistake.
How staff report a bad answer.

Your staff are probably using AI already. In the same research, 74% had no formal training, so write the page down and spend an hour walking people through it.

Buy or build?

Buy for most things. If you're on AppFolio, Buildium, Yardi or a similar platform, try its AI features first, because they already sit next to your data. Use established screening services (with the checks above), and buy a website assistant for enquiries.

Build, or have something built, when one of these is true:

Your systems don't talk to each other. Several platforms, acquired portfolios, or an accounting system nothing integrates with.
Your documents are the asset. Years of agreements and correspondence across many owners, with permissions a generic tool can't enforce.
Your way of working is the advantage. Owner reporting, maintenance routing or vendor rules that off-the-shelf tools would force you to give up.
AI is part of what you sell. An owner portal, or a product for other operators.

If you do talk to developers, us included, these 12 questions will tell you quickly whether they've built anything in production. For property management, add three more: How do you keep one owner's data from reaching another? What happens to an emergency report when the model is down? How do you test for fair treatment?

Frequently asked questions

How are property managers using AI in 2026?

According to IREM and AppFolio's 2026 research, 58% of real estate management professionals now use AI. AppFolio's reporting on the earlier edition lists uses including resident and prospect communication, review replies, lease and vendor paperwork, reports, renewals and maintenance. Full automation is still rare: Buildium's report says only 8% of companies have fully automated any process.

What should a property management company automate first?

The first reply to enquiries, then maintenance intake with a person checking the result, then search over your own documents, then drafts a person sends, then exception reporting. Tie each one to a number you already track, such as days vacant or renewal rate.

Is it legal to use AI for tenant screening?

It isn't banned, but you remain responsible for the outcome. The Fair Housing Act still applies, state and local laws may add more, and if you use consumer reports you have notice duties after an adverse decision. The SafeRent case shows that screening scores can be challenged in court. Keep the final decision with a person, and get legal advice.

Why isn't AI improving our profit if it saves time?

The IREM and AppFolio research shows the gap (85% report productivity gains and 48% report an NOI or profitability impact) without saying what causes it. In our experience, time saved only reaches profit if it's redirected to something measurable. Choose projects tied to a number, and measure that number before and after.

Should a property management company build its own AI system?

Usually not for enquiries, screening or anything your platform already does well. A custom build makes sense when your systems need connecting, your documents and permissions are too complex for generic tools, your processes are your advantage, or AI is part of a product you sell.

Sources

IREM and AppFolio, New research on AI in property management, 5 November 2026. 3,662 respondents, 7 May to 7 June 2026. Read via a syndicated copy of the release.
AppFolio, AI adoption in real estate 2026, 18 September 2026, and 2026 Property Management Benchmark Report release, 18 February 2026 (1,617 respondents, 26 September to 3 November 2025).
Buildium, 2026 Property Management Industry Report, and maintenance article, 4 December 2025. Sample sizes not given on these pages.
Zillow, Rentals Consumer Housing Trends Report 2025, fielded March to July 2024.
Cohen Milstein, Louis et al. v. SafeRent Solutions et al., No. 1:22-cv-10800 (D. Mass.). Cohen Milstein was plaintiffs' counsel.
US Department of Housing and Urban Development, Notification of Withdrawal of Fair Housing and Equal Opportunity Guidance Documents, Federal Register, 6 April 2026.
National Apartment Association, Federal regulatory changes seek to limit disparate impact liability, 2026.
Federal Trade Commission, Greystar settlement, 2 December 2025; warning letters to software providers, 9 December 2025; Using consumer reports: what landlords need to know.
US Department of Justice, RealPage proposed settlement, 24 November 2025.
OpenAI, API deprecations, notice of 14 October 2026.

All sources checked 5–6 November 2026. This article is general information, not legal advice.

Related reading

Brillnex Systems builds AI agents, assistants and retrieval systems for businesses, and runs its own AI product in production. If one of the four custom-build conditions above fits your company, tell us what you're trying to fix, or see our AI integration and RAG implementation services.