November 1, 2026 · 7 min read
Client Data and AI: A One Page Policy for Your Brokerage
Agents are already pasting client details into AI tools. Here is a one page policy a broker can adopt this week, walked through section by section, with a worked example and the sources behind each rule.
A brokerage should give agents one page with five rules: use only approved AI tools on work accounts, never paste red-list client data (IDs, finances, contracts, offers, access codes), scrub names and addresses from everything else, read and verify every output before it reaches a client, and report a mistake the same day. Below is that page, section by section.
Most offices have no such page. Agents use AI anyway, and client details creep in without anyone deciding they should.
This post is part of our guide to AI for real estate marketing. It is a starting draft, not legal advice. Your state defines what counts as confidential, so have your attorney and your errors and omissions carrier read the final version.
Why a brokerage needs this page now
NAR's 2026 REALTORS Technology Report found that 23 percent of agents use AI daily and 25 percent weekly, and only 12 percent say they are not using it and have no plans to. Writing listing descriptions was the top use (75 percent of AI users), followed by social media posting (56 percent) and creating emails and follow-up (52 percent).
A follow-up email is written about a person. That is where client details get typed into a chat window.
Offices are also behind on rules. Metricool's 2026 State of AI in Social Media Study, based on more than 700 social media managers and similar roles, found that only 12 percent have a clearly defined AI policy and 35 percent report no restrictions or formal rules. Those respondents are social media professionals, not agents, so read this as a signal, not a brokerage benchmark.
The duty underneath is old. NAR's Code of Ethics, Standard of Practice 1-9, says the obligation of REALTORS to preserve confidential information (as defined by state law) provided by their clients continues after the relationship ends. A client who told you something in confidence did not agree to have it typed into a tool you picked.
Section 1: Who is covered and which tools are approved
Keep the first block short.
Who. Every agent, employee, assistant and contractor who touches the office's work, including remote helpers.
Tools. Name them: "Approved: the office's paid account with [tool name], on a work login." Everything else is off limits for client information, though agents may still use it for public material.
Enforce the account, not the brand. The same product can behave differently on a personal free login and on a business plan. Agents use several tools: among those who use AI, NAR's 2026 report found 91 percent use ChatGPT, 42 percent Gemini, 27 percent Copilot and 25 percent Claude. A rule that bans one name leaves the other three.
Give the list one owner, usually the broker. A request for a new tool gets an answer in a day.
Section 2: The three lists that do the real work
Sort every kind of information into green, yellow or red.
Green: fine in an approved tool
Facts on the live listing page: bedrooms, baths, square footage, features, the price shown.
The agent's own bio and the office voice guide.
General questions, such as "outline a first-time buyer checklist."
Yellow: only after you scrub it
Describe the situation, not the people. Replace names with "the seller" or "Buyer A," the address with "the listing," and remove dates that point to one person.
A follow-up after an open house, with no names or numbers.
Red: never goes in
Government ID numbers, birth dates, bank, card and loan details, pre-approval letters.
Contracts, offers, counteroffers and disclosures, pasted or uploaded.
What a client will accept, why they are selling, and any hardship such as a divorce, job loss or illness.
Lockbox, gate and alarm codes, passwords and one-time codes.
Recordings and transcripts of client calls, unless the client agreed and your attorney approved. Recording consent rules vary by state.
Add one line under all three: describe the home, never who should live there. Fair housing applies to a prompt the way it applies to a flyer. Our post on AI listing descriptions and the fair housing check shows the method.
Section 3: Account settings, and what vendors say
Settings decide where your text goes after you hit enter.
First, approved accounts are business accounts, with the training and sharing options reviewed by the owner. Read the vendor's page for the plan you hold, and write the date you checked on the policy.
Second, "no training" is not "no risk." Take Anthropic, whose pages are clear. Its privacy center says that by default it does not use inputs or outputs from its commercial products for model training. It lists exceptions: if you explicitly report feedback or bugs, or otherwise choose to allow it. It covers its consumer plans, Claude Free, Pro and Max, in a separate article. Its Claude for Small Business announcement says it does not train on your data by default on its Team and Enterprise plans.
The lesson for the page: a feedback click can make a conversation eligible for training, and a personal login is a different product with different terms. So the red list holds even on a paid plan. Each approved tool gets one line: plan, date checked, who checked.
Section 4: The check before anything leaves the office
AI drafts. A person sends. Then list what the person checks.
Facts. Every figure, date and name matches the listing page or the file. Delete anything the tool added that nobody supplied.
Voice. It reads like the agent. A written voice guide makes this fast.
Fair housing. Nothing describes who should or should not live in the home.
Disclosures. The office's required line is present on listing material.
Promises. No outcome, price or timeline stated as certain.
If a message would reach a client unread, it does not go out.
Section 5: Same-day reporting
An agent will eventually paste an offer summary into the wrong window. The page decides if that agent reports it or hides it. Make reporting cheap:
Tell the owner the same day, with no blame for a prompt report.
Write down the tool, the account, the date and what was pasted.
Delete the conversation if the tool allows it.
The broker and the attorney decide if the client must be told, and how.
Add what you learned to the page.
Agents who fear punishment stop reporting.
A worked example: one seller follow-up
Take an invented boutique brokerage with 14 agents. An agent wants help writing to a seller after a slow weekend of showings.
The risky prompt pastes in the seller's name, the address, the offer that came in, the number the seller said they would take, and a note that the seller is relocating for a spouse's job. Every one of those is on the red list.
The compliant prompt reads: "Write a short, calm email from a listing agent to a seller after a weekend with several showings and no offers. Tone: direct and warm. Ask for a 15-minute call. Do not promise a result." No names, no address, no numbers.
The agent adds the real details by hand after the draft arrives. The tool supplies structure and tone. The agent supplies the facts.
The whole page, ready to copy
Put this on one sheet with the office name, date and owner at the top.
Scope: all agents, staff and contractors.
Approved tools and accounts, each with a date checked.
Green, yellow and red lists.
Red-list information never goes into any AI tool.
Yellow information is scrubbed first.
No AI output goes to a client or the public unread.
Fair housing: describe homes, never people.
Report mistakes to the owner the same day, with no blame.
Review every quarter and after any vendor change.
How Urban Marketing Edge handles this
We use AI to move faster on drafts and variations. Strategy and taste stay human. Our drafting starts from public material, the live listing page and the client's approved voice guide, not from private files, and every piece goes through our post approval system so nothing is published unseen.
Our short answer on AI for real estate marketing is a quick read for your agents. Plans are month to month, billed in advance, no setup fees. See the plans page for what each includes.
Questions brokers ask
Can agents put client information into ChatGPT or Claude?
Not red-list information, on any tool. Scrubbed, general situations can go into an approved business account. Anthropic's privacy center says its commercial products are not used for training by default, with exceptions such as feedback, and it treats consumer plans separately. Read the page for your plan.
Do we need a policy if only a few agents use AI?
Yes. NAR's 2026 report found 23 percent of agents use AI daily and 25 percent weekly. One page is cheap, and it gives you something to point to if a client asks what your office does with their information.
Is a written AI policy a legal document?
It is an office rule, not a contract, and this post is not legal advice. Your state defines confidential information. Have your attorney and insurance carrier read the final page.
Should we ban AI instead?
A ban mostly moves use to personal accounts, where you cannot see it. Approved tools with clear limits give you more control than a rule nobody follows.
How often should we update the policy?
Every quarter, and any time a vendor changes its plan or terms. Put the date you last checked each tool's page on the sheet.
Book a 30-minute strategy call
If your agents are using AI with no written rules, bring what you have. On the call we map the simplest plan for your content, themes and monthly cadence, and mark which drafting jobs are safe to hand to AI first, with a person checking, and which stay by hand. No pitch deck, no pressure. Book a 30-minute strategy call, or look at what each plan covers on the plans page first.
