AI Policy for a Real Estate Brokerage: What Agents May and May Not Do

A brokerage AI policy should say which tools agents may use, what client information must never be entered, how fair housing is protected, when AI use is disclosed, how to verify identities against deepfakes, and that no wire instruction is ever trusted by email or by an AI-generated voice. Write each as a short do or don't.

NAR publishes an AI policy template for brokers as a starting point (member login required), and I could not read its text for this tutorial, so everything below is written as agent-level rules to compare against it. A policy that sits in a drawer protects nobody; one that names the tools, the data and the consequences gets followed.

Follow me on Instagram@sagnikteaches

Who the policy is for, and what it must change on Monday

A policy written for the principal broker's benefit reads like terms and conditions. One written for agents reads like instructions: this tool is approved, this one is not, these things never go in, this is who you ask. Judge each clause by one test: can an independent agent, working from a phone between showings, tell from the sentence what to do?

Connect on LinkedInSagnik Bhattacharya

Five outcomes should change on Monday. Agents know which AI tools are approved. They know what client data stays out of AI tools. Listing copy and ads pass a fair housing check before publication. Any money instruction is verified by a call to a known number. And someone is named to answer questions. The rest of the policy supports those five.

Subscribe on YouTube@sagnikteaches

The wider process of drafting an AI policy for any small firm is in how to write an AI usage policy. What follows is particular to a brokerage.

Approved tools: a short list with an owner

The most useful clause in the policy is a list. Name each approved tool, what it may be used for, and the data class allowed. Everything not on the list is not approved for brokerage work.

Give the list an owner and a review date. A sample clause: "Only tools on the approved list may be used for brokerage work. The list is held by [role], reviewed every quarter and after any incident. To request a tool, send [role] its name, its data terms and the task you want it for. Use of an unapproved tool for client matters is a breach of this policy." For the tools brokerages most often consider, see AI tools for real estate brokers, from transactions to agent support.

Client data: the rules an agent can remember

Agents handle names, finances, mortgage pre-approvals, identity documents, offers and personal circumstances. The simplest policy is a three-tier rule.

• Green, may be entered in an approved tool: public listing facts, neighbourhood information, your own marketing text, anonymised scenarios.
• Amber, only in a tool approved for the purpose: a client's first name, general preferences, the property address, notes already held in the CRM.
• Red, never entered: identity documents, bank or card details, mortgage and income documents, signed contracts, offers or counter-offers, and anything a client asked you to keep private.

Sample clause: "Do not enter client financial information, identity documents, or any offer or contract into any AI tool unless the principal broker has approved that tool in writing for that use." Before and after for one common task: an agent wants an email summarising an offer. Before: pastes the offer PDF into a chatbot. After: writes the offer terms from memory into a plain template, sends it, and asks the assistant only for tone edits on text that has the figures replaced with placeholders.

The agent's real test: would I be comfortable if the client watched me type this? If not, it goes in the red tier. If the agent hesitates, the answer is to leave it out and ask the person named in the policy.

Fair housing: rules for listings, ads and chat

AI can reproduce biased language fluently, and it can do so without any intent on the agent's part. The agent is still responsible for the listing. The policy needs two parts: what may be published, and who checks.

• Do: describe the property's features, condition and location facts. Use the same description rules for every listing.
• Do: run every AI-drafted listing, ad and social post through a fair housing check before publication, using a checklist in the policy.
• Don't: ask AI to describe the type of person a home or neighbourhood suits, or to describe a neighbourhood by who lives there.
• Don't: let a chatbot or assistant decide which listings or areas to show a client based on their family, religion, origin or other protected characteristics.

Sample rewrite. Draft from an assistant: "A perfect family home in a quiet, safe area, ideal for young professionals." Fixed: "Three bedrooms, a fenced garden, 12 minutes' walk to the station, on a no-through road." The first version implies who belongs and makes a safety claim; the second states checkable features. Pair the policy with fair-housing-safe prompts for listing descriptions, and run the same checklist over any chatbot script before it goes on your site.

Disclosure and photos: telling people when AI is involved

Three situations come up. First, a client talking to something that is not a person: a chat widget or voice agent must say so at the start. Second, altered images: virtual staging, sky replacement and object removal can mislead. Third, AI-written content that makes claims, such as market statistics, which an agent must check against a source before using.

REALTOR Magazine's article "Using AI to Enhance Listing Photos Can Be Legally Risky" ties the risk to the Code of Ethics duty not to exaggerate, conceal or misrepresent pertinent facts about a property, and says altered photos should not go on the MLS without disclosure. That gives a good basis for a clause: "Photos may be adjusted for exposure, colour and cleanliness. They may not be altered to add, remove or change permanent features, views, finishes or condition. Any virtual staging must be labelled as such on the listing." Check the wording against your MLS rules, since an MLS can set its own photo and staging requirements.

Sample disclosure line for a chat widget: "You are chatting with an automated assistant. A licensed agent will review anything involving your offer or your finances." It is short enough to show every time. If you use AI voice calls for follow-up, the rules on calling consent are separate; see AI voice agents for real estate follow-up and the calling rules.

Deepfakes and impersonation: verify the person, not the voice

REALTOR Magazine's piece "Prevent Deepfakes from Hijacking Your Transactions" describes fake rental listings, criminals impersonating clients to extract transaction details, deepfake video used to lure agents into meetings, and wire fraud that is harder to detect. It advises contacting the source through a known email address or phone number before acting, adding a wire fraud warning to email signatures, and treating deepfake detection tools as a supplement, because they are still early and not always effective. The policy response is procedural: identity is verified by a channel the fraudster does not control, never by how convincing a voice or video seems.

• Verify a new seller or landlord's identity in person or by a method your broker has approved before listing, and compare against the public ownership record.
• Any request that changes how, where or when money moves is confirmed by calling a number you already held before the request arrived.
• A video call alone is not verification. If a face or voice asks you to act urgently on money or access, stop and use the call-back rule.
• Agents report suspected impersonation to the broker the same day, with a copy of what was received.

Sample clause: "No agent will act on a request to change payment details, release funds, or give access to a property based on email, text or any call or video where the requester has not been verified through a known number. Urgency is a reason for more checking." A call-back script is in real estate wire fraud and voice-clone calls.

Wire fraud rules: the ones that never bend

This is the section to word most firmly. Agents do not send wiring instructions, and never edit one. Title or settlement providers issue them. The agent's job is to tell clients, in writing and early, how instructions will arrive and that any change will be confirmed by phone.

1. Send every buyer, at the start of the transaction, a short written notice: instructions come only from the settlement provider, and they must phone a known number before sending anything.
2. Never forward wiring instructions yourself, or retype them.
3. Never ask an AI tool to read, summarise or draft anything containing account details.
4. If a client reports an instruction that has changed, stop, call the known number, and tell the broker immediately.

Illustrative incident: a buyer receives an email that looks like it comes from the agent, with a "corrected" account number. The buyer replies to the agent, who sees the instructions differ from the settlement provider's. Because the policy told the buyer to call first, the transfer is stopped. The policy mattered because the notice was sent on day one, not because anyone was clever on the day.

Examples of agent decisions, ruled by the policy

Policies earn trust when they settle real cases. Here are five, with the ruling and the reason.

• "Can I paste my buyer's pre-approval letter into the chatbot to draft an offer cover note?" No. It is a financial document, in the red tier. Write the note from a template with placeholders.
• "Can I use AI to write the listing for the Oak Street house?" Yes, from the public facts and photos, then run the fair housing checklist before publishing.
• "A seller I have never met wants to list a house and will only text. Can I use an AI service to check the ID photo?" No. Use the approved verification method and the ownership record. An AI check on an image is not verification.
• "Can I virtually stage a vacant living room?" Yes, if labelled as staged and not altered beyond furniture and decoration.
• "My personal AI account writes nicer emails. Can I use it for client emails?" Only if the email contains no client information beyond a first name and no transaction details; otherwise use the approved tool.

Add a "Rulings" page to the policy and extend it every time an agent asks something new. The questions are the policy's training data.

Marketing, leads and outreach: where AI use becomes a legal risk

Agents use AI most in marketing, so the policy needs a short section there. Three rules cover most of it. Lists: no bought lists loaded into an AI tool, and no scraping of personal data for outreach, because the permission question comes before the tool question (see AI lead generation without bought lists). Calls and texts: automated or AI-voiced contact needs the same consent as any other, and an agent should not set up an AI caller without the broker's sign-off. Claims: statistics, market predictions and "sold" figures in AI-written posts are checked against a source and dated, because a confident invented number is the most common AI error in real estate content.

Sample clause: "AI-written marketing may not state prices, sales figures, school ratings or crime information unless the agent has checked the figure against its source and recorded the source. Do not describe residents or the type of people who live in or would suit an area."

Before and after: an assistant writes "Prices in this street are up 14 percent this year, so buy now." The agent cannot source the figure, so the clause removes it. The post says "Homes on this street have sold between [low] and [high] in the past 12 months, according to [source and date]." It is less exciting, and every word of it can be defended.

Contracts, disclosures and legal wording

Agents are not lawyers, and AI makes it too easy to produce legal-looking text. The policy should say that AI never drafts or edits terms in a listing agreement, purchase contract, addendum or disclosure form, and that approved forms are used as issued. AI may help with an email explaining a form in plain words, as long as the agent checks the explanation against the form and does not present it as legal advice.

Illustrative: a seller asks what a clause on inspection means. The agent asks an assistant for a plain-English version of the clause (the clause is public form text, not client data), reads it against the form, and edits it. The reply to the seller says the explanation is general and to ask a lawyer about their own position. If the assistant had been asked to write a new clause, the answer under the policy is no.

When agents use assistants, teams or contractors

AI policy often fails at the edges. A virtual assistant, a photographer, a transaction coordinator or a team member working under an agent may use AI tools the broker never approved. The policy therefore says that it binds anyone who works on the brokerage's files, and that agents are responsible for the tools used by those they engage. Ask each contractor to confirm in writing which AI tools they use on client work, and to stop using unapproved ones. A short clause in the agreement does the job: "Contractor will use only AI tools listed in the brokerage's approved list for brokerage files and will not enter client financial or identity information into any AI tool."

What happens when an agent leaves

Closing the loop matters. On departure, the agent confirms that no brokerage data is held in personal AI accounts, deletes any saved chats containing client details, and hands over any shared prompt libraries. Work this into the offboarding checklist the same way you collect a key and revoke a login. If an agent used an approved tool through a brokerage account, the broker disables the seat the same day.

Roles, breaches and incidents

Name a person responsible for AI questions, such as the managing broker or compliance lead, and give agents a way to ask that returns an answer the same day. A policy without a named person turns into guessing.

Say what happens on a breach, in proportion. First, a conversation and retraining. Second, a written warning. Serious breaches, such as entering client financial documents into an unapproved tool, may lead to suspension of AI privileges or more. And say what an agent does after a mistake: report it the same day without blame for the report. Hiding is the real danger. An incident log with date, tool, data involved and action taken gives the broker evidence and gives the next audit something to read.

Sample clause: "If you think you have entered client information into an unapproved tool, or acted on a fraudulent instruction, tell [role] immediately. Reporting promptly will be treated as a mitigating factor."

Training, sign-off and review

Do not issue the policy as a PDF and hope. Run a 30-minute session using the five rulings above as quiz questions, then have each agent sign a one-page acknowledgement listing the approved tools and the red-tier data. New agents sign at onboarding. Review twice a year, and add a line at the top of the policy giving the version date, so agents can tell they have the current one.

Check that it works with three tests each quarter. Ask five agents to name two approved tools and one red-tier item. Sample ten recent listings for fair housing issues. And run a call-back drill: have someone send a mock instruction email and see who calls the known number. The drill tells you more than a signature does.

One page for the agents' office wall

• Use only approved tools for client work. Ask [role] if in doubt.
• No client finances, IDs, contracts or offers in any AI tool.
• Check every listing and ad for fair housing before it goes out.
• Say when a chat or voice is automated. Label staged images.
• Verify identities and money instructions by calling a number you already had.
• Report mistakes the same day.

Questions brokers ask when writing the policy

Does NAR provide an AI policy template for brokers?

Yes. NAR's artificial intelligence page lists an AI Policy Template for Brokers as a starting point for clear standards on responsible AI use in a brokerage, with member login required. There is a separate template for associations. Adapt the template to your office rather than adopting it unchanged.

Should agents be allowed to use their own AI accounts?

Only for tasks that involve no client or transaction information, and only if the policy says so. Anything touching client names, offers, finances or documents should go through an approved tool the brokerage has reviewed. Personal accounts cannot be audited or switched off when an agent leaves.

Do agents have to disclose AI use to clients?

Disclose when AI interacts with a client directly, such as a chatbot or a synthetic voice, and when AI-altered images or text could mislead. Check local licensing rules and your MLS rules, since requirements differ. A plain line in your listing and marketing standards is a sensible default.

How often should the policy be reviewed?

Review it at least twice a year and after any incident, a new tool approval or a rule change from your association or MLS. AI products change monthly, so the approved-tools list needs a named owner and a review date, even when the rest of the policy stays put.

Further reads

• Can ChatGPT Write Real Estate Listings That Pass Fair Housing? — Test listing copy against fair housing rules.
• ChatGPT Ads for Realtors: Access, Cost and Fair-Housing Rules — Ad rules for agents using ChatGPT.
• AI Real Estate SEO: Neighbourhood Pages That Avoid Steering — Write area pages that avoid steering.
• AI Real Estate Cold Calling: Scripts, Lists and Do-Not-Call Rules — Calling rules before agents use AI scripts.
• AI Marketing for Real Estate Agents: A Four-Week Content Plan — A four-week content plan to apply the marketing rules.
• AI Acceptable Use Policy for a Small Professional Firm — Compare a policy written for another professional firm.
• Rechat AI for Brokerages: What Lucy Does and What It Costs — Cost a brokerage platform before approving it.
• How to Spot Deepfake Voice and Video Scams Aimed at Your Business — Spot cloned voices and fake video calls early.
• AI for Real Estate Brokers: Compliance File Review Step by Step — A checklist-driven way to review transaction files with AI: exception report, human sign-off and the data-handling rules that keep agent files safe.
• AI for Rental Properties: Screening Enquiries, Viewings and Tenant Replies — One landlord with six rentals, walked through enquiry replies, viewing scheduling and tenant messages with AI, with numbers and fair-housing guardrails.
• Places to Use AI in Real Estate: Eight Jobs for a Small Brokerage — Prioritise eight practical AI jobs by volume, source quality and risk, then pilot one workflow with a review owner.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: NAR's artificial intelligence page, which lists the AI Policy Template for Brokers (member login required), and the two REALTOR Magazine articles it links on AI-enhanced listing photos and on deepfakes in transactions, all read on 4 November 2026. The template's text was not available to read, so the rules and clauses here are my own and should be checked against it.