In most businesses, a bad AI output is embarrassing. In real estate it’s a licence problem.
That’s the difference nobody puts in the tool roundups. The same listing description that would be a harmless typo on a blog can be a Fair Housing complaint. The same photo edit that’s routine on Instagram can be a disclosure violation. The same dialler that saves a marketer an afternoon can generate statutory damages per call.
So this is less a list of tools and more a map of three exposures — what the AI says, what it shows, and who it calls — followed by where AI genuinely makes real estate agents money. In all three exposures the liability sits with the licensed agent, never the software.
What the AI says: Fair Housing
Listing descriptions are the most common first use of AI in this business, and the most quietly dangerous.
A model writing marketing copy optimises for appeal. Fair Housing law restricts language that indicates a preference based on protected classes — race, colour, religion, sex, national origin, familial status and disability federally, with additional classes in many states and under NAR’s Code. A model has no instinct for that boundary, and its most persuasive instincts run straight into it.
The phrases that get generated without anyone intending anything:
- Familial status — “perfect for a young family,” “ideal for empty nesters,” “not suitable for children.”
- Religion — “walk to church,” “near the synagogue,” “in the heart of the [religious] community.”
- Disability — “no wheelchair access,” or the inverse framing that presents accessibility as unusual.
- National origin and race by proxy — describing an area by the people in it rather than the property.
- Steering by implication — “safe neighbourhood,” “good schools,” “the right kind of buyer.” Claims about crime and school quality carry their own scrutiny in fair-housing guidance and are best left to sources the buyer consults directly.
The workable rule is the same one that governs AI in hiring, laid out in AI for HR: the model drafts, a person decides. Every generated description gets a deliberate Fair Housing read before it reaches the MLS — not a skim, a specific pass looking for the categories above.
And a second pass for accuracy, because the other failure mode is invention. Square footage, lot size, school district, year built, permit status — a model will supply confident numbers it doesn’t have. The verification habits in fact-checking AI-generated content apply directly, and in a listing the consequence isn’t a correction, it’s a misrepresentation claim.
What the AI shows: the image disclosure line
This is where the rules changed most recently, and where most agents are behind.
California’s AB 723 took effect on 1 January 2026. It added Section 10140.8 to the Business and Professions Code, requiring a broker or salesperson using a digitally altered image in advertising or promotional material to include a disclosure statement on or immediately adjacent to the image, plus a link, URL or QR code leading to the original, unaltered photo. Reporting indicates undisclosed use can be treated as a misdemeanour.
Wisconsin’s Act 69 follows on 1 January 2027. As of 2026 those are the only two dedicated state statutes on AI-altered real estate imagery. Everywhere else the duty comes from MLS rules, state advertising law, and NAR’s Code of Ethics — several large MLSs already require “virtually staged” labelling under their own rules (state-by-state map).
The practical question is which edits cross the line:
| Generally requires disclosure | Generally doesn’t |
|---|---|
| Virtual staging — adding furniture to an empty room | Colour correction and white balance |
| Sky replacement and twilight conversion | Exposure adjustment |
| Removing objects — power lines, cars, clutter | Lens distortion correction |
| Changing landscaping — green grass over brown | Cropping |
| Adding or enhancing fixtures — fireplaces, water features | Sharpening |
| AI-generated or AI-animated video, including simulated drone moves and walkthroughs built from stills | — |
The organising principle: edits that change what the property is require disclosure; edits that change how the photograph looks generally don’t. A brown lawn made green is a claim about the property. A slightly brighter kitchen is a claim about the camera.
Video deserves its own note. A smooth drone-style approach assembled from still photos, or a walkthrough where the camera never moved, sits exactly on the line between marketing and misrepresentation — and trade coverage has been pushing agents toward a formal disclosure test for precisely this (HousingWire).
Two things agents get wrong:
A blanket disclaimer doesn’t satisfy it. “Images are illustrative” in the listing footer is not a label on or beside the altered image, and it doesn’t provide the original. Generic catch-alls generally fail the standard.
You have to keep the originals. California’s requirement includes access to the unaltered photo. That’s a workflow obligation, and most editing tools don’t preserve or organise originals by default. Build the archive on day one; you can’t reconstruct it later.
One correction worth making, because several guides get it wrong: Colorado’s AI Act is not a listing-photo rule. It governs high-risk AI decision systems in areas like hiring and lending. Citing it as an image-disclosure requirement is simply an error.
If you’re choosing image tools, the licensing and ownership questions sit on top of all this — Midjourney vs DALL·E vs Firefly covers what you’re actually permitted to do with generated imagery commercially.
Who the AI calls: the TCPA problem
AI voice diallers are marketed hard to real estate agents, for the obvious reason that prospecting calls are the industry’s daily grind. This is the exposure with the largest downside, because TCPA damages are statutory and accrue per call.
The governing fact: the FCC ruled in February 2024 that AI-generated voices are “artificial voices” under the TCPA. That isn’t a new rule so much as a clarification that the existing one already covers you. Consequences:
- Marketing calls to mobile numbers need prior express written consent.
- The National Do Not Call Registry applies to prospecting.
- Artificial-voice calls must identify the calling entity by name and provide a contact number or address at the start of the call — a requirement on the books since 1991, long before anyone was building voice agents.
Be careful with what you read next, because this is where guidance gets overstated. The FCC’s 2024 proposal to require an explicit in-call “this call uses AI” disclosure has not been finalised. Federally, the AI-specific disclosure is proposed, not law. Several states have moved ahead independently — California and Utah with AI-disclosure duties, a set of states with their own stricter “mini-TCPA” consent standards and private rights of action, and others with telemarketer registration regimes (state overview).
The honest summary for an agent: an AI dialler doesn’t relax any consent rule you already have, and it adds a layer that varies by the state you’re calling into — not the state you’re sitting in. If you can’t document written consent for a number, an AI voice doesn’t make the call safer. It makes it faster, which is worse.
Inbound is a different matter and much safer ground. Answering your own phone and responding to your own web enquiries has no consent problem, and it’s where the missed-revenue maths actually favours automation — the same argument made for hospitality in AI for restaurants. Build it with a clear disclosure that it’s an assistant and a hard escalation rule, as in building an AI chatbot without code.
Where AI genuinely earns an agent money
| Use | Payback | Risk |
|---|---|---|
| Inbound lead response | Highest — speed to first reply is the whole game, and leads arrive at 10pm | Low, with disclosure and escalation |
| Transaction coordination | High — summarising documents, tracking deadlines, drafting reminders | Low; nothing is published |
| CMA and market-data summarising | High — turning data you already hold into readable client-facing narrative | Low, if the numbers come from your data and not the model |
| Follow-up and nurture drafting | Solid — drafts you send, not sends you discover | Low; consent rules still apply |
| Listing descriptions | Moderate — real time saved | Fair Housing review is mandatory |
| Photo enhancement and staging | Moderate | Disclosure plus original-file archive |
| Outbound AI voice prospecting | Lowest, relative to exposure | Highest — TCPA damages are per call |
The ranking is deliberately the inverse of how these tools are marketed. The AI products sold hardest to agents are the prospecting ones; the returns sit in the unglamorous middle — answering the phone, chasing the paperwork, turning your own data into something a client can read.
Notice what the top three have in common: none of them publish anything. Every high-risk use in this article involves output going to the public. Every low-risk one is internal or one-to-one with someone who already contacted you.
A workflow that keeps you licensed
- Nothing AI-produced reaches the public without a human review pass. Not a skim — a specific check for Fair Housing language and for invented facts.
- Archive every original photo before any edit, named so you can produce it on request. Your disclosure obligation may require it, and no editing tool does this for you by default.
- Label altered images individually, on or beside the image, not in a footer disclaimer.
- Never let an AI voice dial a number you can’t document consent for. Check the rules of the state you’re calling into.
- Write down your brokerage’s AI policy and which tools are approved. When something goes wrong, the question will be what your process was.
- Re-check the rules quarterly. Two states have dedicated statutes today, more are moving, and MLS rules change faster than legislatures.
For the marketing side that sits around all this, email marketing automation covers the deliverability rules, writing SEO content with AI covers the neighbourhood-guide content that actually brings enquiries, and the best AI tools for small business owners is the wider stack.
This article is general information, not legal advice. Disclosure requirements vary by state, by MLS and over time. Confirm the rules that apply to you with your broker and your own counsel before relying on any summary, including this one.
Frequently asked questions
Do I have to disclose AI-edited listing photos?
In California, yes — AB 723 has required since 1 January 2026 that a digitally altered image carry a disclosure on or immediately adjacent to it, plus a link, URL or QR code to the original unaltered photo. Wisconsin’s Act 69 follows in 2027. Elsewhere the duty typically comes from MLS rules, state advertising law and the NAR Code of Ethics rather than a dedicated statute, and several large MLSs already require virtual-staging labels.
Which photo edits actually trigger disclosure?
Edits that change what the property is: virtual staging, sky replacement, twilight conversion, removing objects, altering landscaping, adding or enhancing fixtures, and AI-generated or animated video including simulated drone moves. Standard corrections — colour, exposure, lens distortion, cropping, sharpening — generally don’t. Confirm with your own MLS.
Is a general “images are illustrative” disclaimer enough?
Generally no. The standard under AB 723 and most MLS rules is a clear label on or beside the specific altered image, together with access to the original unaltered photo. A catch-all line elsewhere on the listing doesn’t meet that, which is why keeping an organised archive of originals matters from day one.
Can AI-written listing descriptions violate Fair Housing law?
Yes. A model optimising for appeal readily produces phrasing that indicates a preference based on protected class — familial status, religion and disability are the most common — or that steers through implication about neighbourhoods, schools or crime. The description is your advertisement regardless of what drafted it, so every generated draft needs a deliberate Fair Housing review before it’s published.
Are AI voice calls allowed for prospecting?
Only within the existing rules. The FCC confirmed in February 2024 that AI-generated voices count as artificial voices under the TCPA, so marketing calls to mobile numbers require prior express written consent, the Do Not Call Registry applies, and the call must identify the caller at the outset. A federal proposal to mandate an explicit in-call AI disclosure has not been finalised, though some states have their own rules. Damages are statutory and accrue per call.
What’s the safest high-return AI use for a real estate agent?
Responding to inbound enquiries quickly, followed by transaction coordination and summarising your own market data for clients. None of those publish anything to the public, which is where nearly all of the compliance exposure lives, and speed to first reply on an inbound lead is one of the few things that reliably converts.
