Key takeaways
- Nearly half of REALTORS use AI daily or weekly[2]; without a written policy, each agent quietly sets their own rules.
- One page is enough if it covers five things: approved tools, client data, disclosure, human review and records.
- Three data classes (green, yellow, red) tell agents at a glance what may go into which tool.
- AI is no excuse: Utah's 2025 law says AI making a statement is not a defense to a consumer protection violation[3].
Your agents have already written a policy, one each
Nearly half of REALTORS now use AI daily (23%) or weekly (25%)[2]. In a brokerage with no written AI policy, that adds up to dozens of private ones: which tools, which client details go in, whether anyone checks the output before a client sees it. Plenty of organizations sit in that gap. Gallup found that 44% of US employees say their organization has begun integrating AI, but only 22% say it has communicated a clear plan for doing so[1].
Clients are watching too. Among sellers who hired an agent, 21% knew and another 36% suspected their agent was using AI tools[4]. A one-page policy will not settle every question, but it gives agents a default answer and gives the broker something specific to supervise against. The compliance risk overview maps the risks; this article is the document itself.
The uses show where the policy has to bite. Writing listing descriptions leads at 75% of AI-using REALTORS, ahead of social media posts (56%), emails and follow-up (52%) and document review or summarizing (27%)[2]. The first three are advertising and client communication, which already have rules of their own. The fourth puts client documents into a tool, which makes it a data question.
What fits on one page
Tools
An approved list, with an owner and a review date for each entry
Data
Green, yellow and red classes for what may go into which tool
Disclosure
When clients, callers and the public are told AI is involved
Review
A licensee checks anything that reaches a client or the public
Records
Which AI-assisted outputs are kept, where and for how long
Anything longer than a page belongs in the procedures manual, linked from the policy.
Keep the policy short enough to read aloud at a sales meeting and sign at onboarding. Vendor answers, the tool register and step-by-step procedures sit behind it, the way the policies and procedures manual holds the rest of your rules. A policy nobody reads in full is a policy nobody follows.
Approved tools and the register
The register is a short table the broker owns. A tool goes on it after someone has asked the vendor the awkward questions. It comes off when it changes owners, changes its terms or stops being used.
| Tool (Example) | Approved use | Data class allowed | Owner | Next review |
|---|---|---|---|---|
| Listing copy assistant | Draft descriptions and social captions | Green only | Marketing coordinator | March |
| Phone answering line | Main and listing lines, after hours and overflow | Yellow (caller details) | Office manager | June |
| Meeting note taker | Buyer and seller consultations, announced at the start | Yellow | Managing broker | June |
| Virtual staging tool | Staged images, labelled on the listing | Green (photos) | Listing agents | March |
| CRM email drafting | Follow-up drafts an agent edits and sends | Yellow | Team lead | September |
The phone line shows how much detail an entry needs. For an SI Employee answering the main line, the register would record that it answers from an approved listing sheet, takes showing requests for agents to confirm, records and transcribes every call, and emails a summary. The greeting text and the list of questions it must hand to a licensee live in the procedures manual. What it does lists the live capabilities to copy into the entry, and what an AI phone line may say under license law covers the boundaries.
- Does the vendor, or any subprocessor, train its own models on our data or our clients' data? Can we opt out in writing?
- Where is data stored, for how long by default, and can we delete a single record on request?
- Who at the vendor can see our data, and how and when are we told about a breach?
- Can we export everything if we leave, in a format we can open?
Keep the answers with the register entry. Vendor security questions for phone and AI tools has a fuller list for tools that touch calls.
Client data in three classes
Agents should not need a lawyer to decide whether a seller's payoff statement can go into a chatbot. Three classes do the job, and they fit on a laminated card.
- Green: public or published information. MLS listing facts, public records, published market statistics, your own marketing copy. Any approved tool.
- Yellow: ordinary client details. Names, contact details, property addresses, showing feedback, call summaries. Only tools the register approves for yellow data, used under a brokerage account.
- Red: never into an AI tool unless the register names the tool and the use. Social Security numbers, bank and wire details, ID documents, loan files, a seller's bottom line, other buyers' offer terms, and anything about a client's health, disability or family situation.
Red is where the real harm sits. Wire details and closing dates are what fraudsters hunt for, and a seller's bottom line or a competing buyer's terms are confidences the agent holds for a client. Convenience does not change the class. If an agent cannot do a task without putting red data into a tool, the task goes to the broker, not around the policy.
Disclosure: say it where a person would want to know
Pew found that 76% of Americans say it is extremely or very important to be able to tell whether pictures, videos and text were made by AI or by people[5]. Disclosure law is still patchy, so the policy should set a house standard that does not depend on which state a client happens to be in.
- Phone and chat: the greeting says an automated assistant is answering, and anyone who asks gets a straight answer. Utah requires disclosure when a person clearly asks[3]; disclosing AI to callers and leads has wording.
- Outbound AI voice: no calls without the consent the TCPA requires. The FCC treats AI-generated voices as artificial under that law[6].
- Images: AI-altered or virtually staged photos are labelled on the listing. See virtual staging disclosure.
- Meetings: if a note taker is running, say so at the start and ask whether everyone is comfortable with it.
- Questions from clients: when a client asks how the brokerage uses AI, answer plainly. Explaining your AI use to clients has scripts.
Human review: a licensee owns every output
The review rule is the heart of the policy: nothing AI-assisted reaches a client or the public until a licensee has read it and would sign it. Utah wrote the principle into law in 2025. It is not a defense to a consumer protection violation that generative AI made the statement or took the action[3], and licensed professionals who provide services through generative AI must still meet every requirement of their occupation[3].
- Listing copy: the facts match the listing sheet, and nothing describes who the home suits by race, color, national origin, religion, sex, familial status or disability, the classes the Fair Housing Act protects[7]. Drafting listing descriptions with AI has a checklist.
- Reviews and testimonials: never generated. The FTC's 2024 rule bans creating or selling fake reviews and testimonials, including AI-generated ones[8].
- Lead questions and phone scripts: approved by the broker, checked for fair housing bias risks, and kept to what an unlicensed assistant may say.
- Summaries and notes: the agent corrects anything wrong before the note is relied on or forwarded.
Records: keep what reached a client
Keep the final version of anything AI-assisted that went to a client or the public, in the same place as everything else in the file. Prompts and discarded drafts rarely need keeping, and keeping them adds clutter and exposure. Call recordings, transcripts and summaries follow the phone retention schedule; see retention rules for call recordings and texts.
Adding a new tool
01Your team
Agent requests a tool
Names the tool, the job it does and the data it would touch.
02Your team
Vendor answers collected
Training on data, storage, deletion, breach notice and export.
03Your team
Test with fictional data
Two weeks, no real client details, every output reviewed.
04Your team
Added to the register
Approved use, data class, owner and review date written down.
05Your team
Agents trained
Ten minutes at a sales meeting, with sign-off recorded.
Example. A tool used only with green data can take the short path: vendor answers, then a register entry.
The AI pilot guide covers the test period in more depth. If you want an outside framework to check the policy against, NIST's AI Risk Management Framework, released on January 26, 2023, is intended for voluntary use, and NIST added a generative AI profile on July 26, 2024[9].
Rolling it out without a revolt
Agents who already use AI every day will read a new policy as a ban unless it arrives with approved tools in hand. Launch the register and the policy together, so the first message is what agents can use, not only what they must stop.
- Week 1: ask agents which tools they already use and for which jobs. Expect more tools than you guessed, including free ones on personal accounts.
- Week 2: approve the common tools that pass the vendor questions, each with its data class, and say no to the rest with a reason.
- Week 3: walk through the policy at a sales meeting using three realistic samples: a listing description, a client email and a call summary.
- Week 4: collect signatures, add the policy to onboarding, and put the first register review on the calendar.
After launch, the managing broker spot-checks a handful of AI-assisted outputs each month, the same way they review advertising, and brings anything that went wrong to the next meeting without names. That habit does more than any clause to keep the policy alive.
The one-page policy
Example text to adapt. Swap the role names for the people who hold them, and have every licensee and staff member sign it.
- Scope. This policy covers every licensee and staff member who uses AI tools for brokerage business, on any device.
- Approved tools. Only tools on the AI register may be used with client information. The managing broker owns the register and reviews each entry at least twice a year.
- Client data. Green data may go into any approved tool. Yellow data only into tools the register approves for it. Red data never, unless the register names the tool and the use.
- Disclosure. Phone and chat greetings disclose automated assistants, and anyone who asks is told plainly. AI-altered images are labelled. Note takers are announced.
- Review. A licensee reads and approves anything AI-assisted before it reaches a client or the public, and is responsible for it as if they had written it.
- Never. No AI-generated reviews or testimonials, no outbound AI-voice calls without required consent, and no opinions of value or negotiation by an automated assistant.
- Records. Final client-facing outputs go into the transaction file. Calls, transcripts and texts follow the retention schedule.
- Problems. Report a wrong output, a data slip or a client complaint about AI to the managing broker the same day.
Example: the 40-agent front desk scenario shows clause 6 in practice. The brokerage writes its boundaries into the answering line's setup before the first call, so neighborhood questions beyond the listing facts go to an agent every time.
Frequently asked questions
Does a brokerage need a written AI policy?
The rules that matter most here, from fair housing to advertising, consent and privacy, apply to AI-assisted work just as they do to anything else. Utah's 2025 law spells it out: AI involvement is no defense to a consumer protection violation[3]. A written policy is how a broker shows that work is supervised.
Should agents disclose that a listing description was written with AI?
The policy's test for copy is accuracy and review: a licensee checks the facts and the fair housing language before anything publishes. Images that alter how a property looks are different and should be labelled.
Can agents use free AI tools?
With green data only, unless the register approves the tool for more. Free tools often carry the weakest terms on how your data may be used, so check before approving one for client details.
How often should the policy change?
Review the register twice a year and the policy once a year, or sooner when a tool changes its terms or a new state law takes effect. NIST's own framework is being revised[9].
Who enforces it?
The managing broker or broker of record, the same person who supervises advertising and transaction files. Name them in the policy so agents know where to report a problem.
References
Gallup (2025). AI Use at Work Has Nearly Doubled in Two Years (opens in a new tab) ↑
National Association of REALTORS® (2026). REALTORS® Adopt Technology to Save Time and Improve the Client Experience, NAR Report Finds (2026 REALTORS® Technology Report) (opens in a new tab) ↑
Utah State Legislature (2025). S.B. 226 Artificial Intelligence Consumer Protection Amendments (2025 General Session), enrolled copy (opens in a new tab) ↑
Zillow Research (2026). Sellers: Results from the Zillow Consumer Housing Trends Report 2026 (opens in a new tab) ↑
Pew Research Center (2025). How Americans View AI and Its Impact on People and Society (opens in a new tab) ↑
Federal Communications Commission (2024). FCC Makes AI-Generated Voices in Robocalls Illegal (opens in a new tab) ↑
U.S. Department of Housing and Urban Development (2026). Housing Discrimination Under the Fair Housing Act (opens in a new tab) ↑
Federal Trade Commission (2024). Federal Trade Commission Announces Final Rule Banning Fake Reviews and Testimonials (opens in a new tab) ↑
National Institute of Standards and Technology (2026). AI Risk Management Framework (opens in a new tab) ↑