AI for real estate: what it does well, and the rules it still has to follow

Which AI is good for real estate?

There isn’t one tool that does it all, because an agent’s week, an investor’s underwriting and a property manager’s inbox are different jobs. Real estate AI comes in three kinds:

• General assistants: ChatGPT, Claude, Gemini and Microsoft Copilot draft, summarize and rewrite. They know nothing about your listing or your client until you give them the material.
• AI inside the software you already pay for: drafting, summaries or lead scoring built into a CRM, transaction platform or property management system. Your data is already there, so nothing new has to be connected.
• Tools built for one job: virtual staging and photo editing, lead response, lease abstraction, rent analysis. Each does a narrow thing inside a workflow.

Four questions help you choose. Which job is slow today? Where does the data for it live? Does the plan use what you type to train its models, and can you turn that off? Who checks the output before a client, tenant or owner sees it? A tool without good answers to the last two isn’t ready for client work, however good its demo.

How do real estate agents use AI day to day?

For agents, AI earns its place on the writing and admin that fills the hours between showings:

• Listing descriptions: a first draft from your walkthrough notes, the MLS fields and the seller’s list of upgrades, checked line by line before it goes live.
• Lead follow-up: replies to portal inquiries, a summary of showing feedback for the seller and the note after an open house, each drafted for you to edit and send.
• Market updates: a plain-English summary of the month’s numbers for your farm area, written from the figures you export from your MLS.
• Contract dates: the deadlines and contingencies in a signed purchase agreement pulled into a checklist, which you confirm against the contract itself.
• Social posts and emails: one listing rewritten for each channel, in your own voice.

The pattern holds across all of them: the model drafts from material you supply, and you decide what goes out. Its memory is the weak point. Ask it for a school’s name, an HOA fee or a recent sale price without giving it a source, and it may produce something that reads well and is wrong.

Can an AI-written listing break fair housing rules?

Yes, and whoever makes or publishes the ad answers for it. The Fair Housing Act makes it unlawful to publish a notice, statement or advertisement for the sale or rental of a dwelling that indicates a preference, limitation or discrimination based on race, color, religion, sex, handicap, familial status or national origin (42 U.S.C. 3604(c)). Nothing in that rule changes when a model wrote the words.

An AI draft can drift toward trouble. Ask a model to make a condo appeal to young professionals and it may call the unit perfect for a couple without kids, which indicates a preference based on familial status. Ask it to describe a neighborhood and it may describe the people instead of the place. The fix belongs in both the prompt and the review: describe the property, its features and its location, never who should live there, and read each draft for words about people before it’s published.

Who sees the ad matters too. In June 2022, the Justice Department settled a Fair Housing Act case against Meta over how its system targeted and delivered housing ads. Meta denied liability and any wrongdoing. It agreed to stop delivering housing ads targeted with its Special Ad Audience tool by December 31, 2022, and to build a system to narrow the gap, by sex and estimated race or ethnicity, between the audience an ad was meant for and the people who actually saw it.

Do AI-edited listing photos need a label?

In California, yes, since January 1, 2026. AB 723 (Chapter 497, Statutes of 2025) added section 10140.8 to the Business and Professions Code. A broker or salesperson, or someone acting for them, who uses a digitally altered image to advertise a property for sale must say so in a statement on or next to the image, with a link, URL or QR code that leads to the original, unaltered image.

The law’s definition covers the edits AI tools make easy. An image counts as digitally altered when photo editing software or AI adds, removes or changes elements in it, and the list includes furniture, appliances, flooring, walls, paint color, landscaping and the view through a window, so virtual staging is in. An image with only common adjustments, such as lighting, exposure, color correction, straightening or cropping, that don’t change how the property is represented is left out. The useful question is what a buyer would find at the showing that the photo didn’t show them.

Outside California, check your state’s rules and your MLS’s photo policy before you publish an edited image. Even where no rule requires a label, a buyer who walks into a room that doesn’t match the photos has reason to doubt the rest of the listing.

How do I use AI for real estate lead generation?

In lead generation, AI works on response time and follow-through. A new inquiry that arrives in the evening or on a Sunday can get an answer with the listing’s facts, a few questions (timeline, financing, the areas they’re watching) and a showing booked from times the agent has opened. The agent picks up a conversation that has already started, with the answers in front of them.

Calling rules don’t relax for AI. In February 2024, the FCC ruled that AI-generated voices are “artificial” under the Telephone Consumer Protection Act. Outside emergencies and the rule’s exemptions, a call that uses one needs the called person’s prior express consent, and prior express written consent when the call introduces an advertisement or contains telemarketing. A synthetic voice working through a list of homeowners who never agreed to hear from you is the kind of call the ruling covers.

Two habits keep AI lead work out of trouble:

• Record consent where the outreach runs: how and when each lead agreed to be contacted, stored in the same system that does the contacting.
• Keep a person on anything that advises or commits: price opinions, terms, representation agreements and anything a client might later rely on.

How do I use AI for real estate investing?

For an investor, AI is a fast first reader. Give it an offering memorandum, a rent roll, a trailing twelve months of operating statements or a stack of leases, and it can summarize them, pull the numbers into a table and flag what doesn’t line up: a rent roll that disagrees with the leases, an expense line that jumps, a lease that ends before the loan does.

It shouldn’t be the source of your numbers. NIST’s July 2024 profile of generative AI risks names confabulation, “the production of confidently stated but erroneous or false content”, as one of them. In underwriting, that looks like a cap rate computed from a rent the model assumed, or a comparable sale that never happened.

One way to split the work: the model reads, extracts and flags; your spreadsheet does the math from figures you’ve checked against the documents; the appraisal, the inspection and the title work stay with the people who sign them.

Can AI help with property management?

Yes. Property management AI suits the steady, repeated work:

• Maintenance requests: reading the tenant’s message and photos, sorting by urgency and trade, and drafting a work order for a person to approve and send to a vendor.
• Tenant questions: answering the routine ones (when rent is due, parking, trash days, how to submit a request) from your own documents, and passing anything else to staff.
• Lease abstraction: dates, rent steps, renewal options and who pays for what, pulled from each lease into one table.
• Owner reporting: a first draft of the monthly owner letter from the figures in your accounting system.
• Lease ends: a list of leases ending in the next 90 days, with draft notices for staff to review and send.

Keep the approvals with people: anything that spends an owner’s money, changes a lease or turns down an applicant.

Can AI set rents or screen tenants?

Both drew new rules and enforcement in 2025.

Rent pricing. In November 2025, the Justice Department announced a proposed settlement with RealPage, whose revenue management software landlords use to set rents. Under it, RealPage would stop having its software use competitors’ nonpublic, competitively sensitive information to set rental prices, and would limit the nonpublic data it trains its models on to data at least 12 months old. The settlement needed the court’s approval when it was announced. In 2025, California also passed AB 325 (Chapter 338, Statutes of 2025), which prohibits using or distributing a common pricing algorithm as part of a contract, combination or conspiracy to restrain trade, or coercing someone to adopt a price or term it recommends. The question to put to any pricing tool is whose data it learns from.

Tenant decisions. California’s privacy agency counts housing among the significant decisions covered by its rules on automated decisionmaking technology. Businesses covered by the California Consumer Privacy Act that use the technology to make those decisions must comply by January 1, 2027, and the agency lists what consumers get: notice that it’s used, an opt-out where one applies, and access to information about how it worked and affected them. The Fair Housing Act’s bar on discrimination in renting applies as well, and the statute makes no exception for decisions made with software.

A safer line: let AI gather an application, summarize it and check it against criteria you’ve written down, and keep the decision with a person who can explain it.

Where does real estate AI go wrong?

These are the failures to plan for:

• Invented facts: square footage, HOA dues or a sale that never happened, stated with the same confidence as the true ones.
• Old numbers: a model’s general knowledge stops at a cutoff date, so market figures should come from your MLS export.
• Words about people: neighborhood and buyer descriptions that drift into fair housing trouble.
• Altered photos with no label: in California a legal problem, and everywhere a trust problem.
• Missing consent: AI-voiced calls to people with no consent on record.
• Client documents in the wrong tool: a buyer’s pre-approval letter or a tenant’s application pasted into a tool whose plan trains on what you type.

Where should you start?

Start small, and write your rules before your prompts, so the tool works to your standards:

1. Pick one slow, repeated job: listing drafts, inquiry replies or maintenance intake, whichever takes the most hours each week.
2. Write the rules down first: your fair housing word list, your photo policy, where consent is recorded and who approves what.
3. Look inside the software you already pay for: check what your CRM, MLS and property management system include before buying another tool.
4. Give it the source material: the MLS fields, the lease, the export. Don’t ask a model to remember facts it should be reading.
5. Review everything at first, and time it: keep a person on every draft a client, tenant or owner will see, and track the hours saved against the hours spent checking.

If more than one job is slow, our guide to what to automate first weighs them on how often each runs, how fixed its rules are and the hours it takes.

If you’d rather have help with it, our real estate page lays out what we would build for real estate businesses. We look for where AI can take work off your team, then build it, run it and keep it working, and we don’t build AI that decides who gets a lease. Every engagement starts with a 15-minute call.