ChatGPT for Realtors: Brokerage Rules for Emails, Disclosures, Follow-Ups

A brokerage owner can let agents use ChatGPT for client emails, follow-ups and plain-language summaries of disclosures, provided agents redact client data first, use the brokerage's approved account, never ask it to judge what a disclosure means, and run a fact-check pass against the source document before anything is sent.

The risk in a brokerage is not one agent making one mistake. It is twenty agents each inventing their own habits, with the brokerage liable for all of them. So the owner's job is to set the policy, provide the account, standardise the prompts and run the checks that turn a good intention into practice.

Follow me on Instagram@sagnikteaches

Summarise, redact, check: the rules that carry the policy

A policy agents will actually follow is short. These three cover most of the risk, and each can be tested.

Connect on LinkedInSagnik Bhattacharya

1. Summarise, never advise. ChatGPT may restate what a document says, list its items and tidy wording. It may not say whether something is acceptable, how serious it is, what a buyer or seller should do, or what the law requires. Those answers belong to a licensed person who has read the document.
2. Redact before pasting. No names, addresses, contact details, financial details, identity numbers or personal circumstances go into a prompt. Use placeholders.
3. Check before sending. Every sentence that states a fact, a date, a figure or a clause is checked against the source by the agent whose name is on the message.

Write them at the top of a one-page document, add the approved account and the five prompts below, and ask every agent to sign it. The signature matters less than the conversation it forces.

Subscribe on YouTube@sagnikteaches

Give agents an approved place to work

The most common leak is not malice. It is an agent using a personal free account because it is faster than asking. Remove the reason. Set up one brokerage workspace on a ChatGPT business plan: OpenAI says it does not train on business workspace data by default, and a single workspace lets you apply one policy. Individual accounts have a training switch each user must manage. Business plans need a minimum of two seats, which is no obstacle for a brokerage. Check OpenAI's current plan terms before committing, since plans and limits change.

Inside the workspace, create a Project named "Client communications" and put the brokerage's standing instructions there. Skip the idea of a brokerage GPT: custom GPTs are being retired as OpenAI shifts them to plugins, and a Project holds shared instructions well enough. For the plan question in more detail, see the comparison of free, Plus and business plans for realtors.

The agent's personal brand and your brokerage's risk meet in tooling choices, so it is worth reading which assistant fits a realtor's day before standardising on one.

Brokerage instructions with a built-in refusal

You help real estate agents at a brokerage draft client emails, follow-ups and plain-language summaries.
You are not a licensed agent and give no advice.
Never: say whether a disclosure item is serious, acceptable or a problem; interpret a contract; state what the law requires; predict a sale price, a market change or an outcome; recommend an offer amount; invent facts about a property.
If asked for any of these, reply: "That needs your supervising broker or a lawyer. I can summarise what the document says."
Use only text the user pastes. If a fact is missing, write [MISSING].
Use placeholders [BUYER], [SELLER], [ADDRESS], [PRICE], [DATE], [AGENT NAME]. If the user pastes personal data, tell them to redact and replace it.
British English unless the user asks otherwise. Plain words. Under 150 words for emails.

The refusal line is the visible alarm. When an agent sees "That needs your supervising broker", they have drifted into advice. The line is also useful in review: any sent message that sounds like advice did not come through this Project.

Redaction that agents can do in two minutes

Agents resist redaction when it feels like a research task. Give them a lookup table and a rule of thumb.

The rule of thumb: if the document would identify someone to a stranger who read the pasted text, redact more. Where redaction destroys the meaning, as with a long inspection report full of site-specific detail, skip the model and read it yourself. Signatures are the usual miss: an email thread pasted for a reply often carries the client's phone number and address in the footer, so redact from the bottom of the thread upwards.

Five reusable prompts, each with its limit

Prompt 1. Follow-up after a showing

Write a short follow-up email from [AGENT NAME] to [BUYER] after a showing of [ADDRESS]. Include two things I noticed, supplied below, ask one question about what they liked or disliked, and suggest a call on [DATE].
Things I noticed: [AGENT'S OWN OBSERVATIONS]
Do not describe the property beyond what I wrote. Do not predict price or demand.

Realistic output: a warm, brief email with the two observations and a question. What to fix: models like to praise a property ("a wonderful family home"), which can slide into a representation about the property that the agent did not verify. Delete praise adjectives and keep the agent's actual observations.

Prompt 2. Summary of a seller disclosure

Below is the text of a seller's disclosure with personal data removed. List each item the seller disclosed, grouped by area (structure, systems, environment, legal, other), using the document's own words where possible.
End with a list headed "Unclear or missing", containing anything that is ambiguous, blank or refers to another document.
Do not say whether any item is serious, normal or a concern. Do not recommend questions for the buyer.

This is the prompt most in need of the owner's rule. The output is a tidy list of what the form says, and the "Unclear or missing" section is the most useful part: blank answers and "see attached" references are where issues hide. A typical failure is a model merging two items into one summary line, softening an issue. The agent checks the list against the form line by line before it goes to the client, and tells the client plainly that it is a summary, that the document governs and that questions about meaning go to the supervising broker or a lawyer.

Prompt 3. Plain-language explainer of a process step

Explain in plain English, in under 200 words, what happens at the "[STEP]" stage of buying or selling a home, using the process steps I paste below. Do not add steps. Do not mention timescales, legal requirements or costs.
Our process: [PASTE BROKERAGE PROCESS]

This one is almost risk-free because the content is the brokerage's own process, not a client's data. Build a small set for the steps agents explain most (offers, inspections, appraisal, closing) and store them in the Project.

Prompt 4. Re-engaging an old lead

Write a short, friendly check-in to [FIRST NAME], who contacted us in [MONTH] about [BUYING/SELLING]. Offer to answer questions and let them know they can reply STOP to end messages.
Do not claim market changes or give prices. Under 70 words.

Follow-up rules about consent and contact preferences vary and carry penalties in some places, so the brokerage should define who may be contacted and how, and the CRM, not the chat, should enforce opt-outs. For automation of the follow-up routine see how to automate buyer and vendor follow-up with AI.

Prompt 5. Clean-up of the agent's own notes

Turn my notes below into a file note with headings: Client's goals, Properties viewed, Feedback, Next steps. Mark any statement of uncertainty as "agent unsure". Do not add information that is not in the notes.
Notes: [NOTES, names replaced]

Check that uncertainty survives. "Buyer might offer this week" must not become "buyer will offer this week". The "agent unsure" tag helps.

Matching every sentence to the disclosure or contract

Every message that states a fact gets the same short check before it leaves. Make it a stamp on the message in your system: "Checked by [AGENT] against source on [DATE]".

1. Locate the source. Open the disclosure, listing sheet or contract the statement came from.
2. Tick each fact. Every date, price, address, deadline, feature and item in the draft is matched to the source.
3. Look for additions. Anything in the draft that is not in the source, however plausible, is deleted.
4. Search for verbs of certainty. "Will", "guaranteed", "definitely", "no issue", "perfectly fine". Soften or delete.
5. Read as the recipient. Could a client act on this as if it were advice? If yes, rewrite it or send it to the supervising broker first.

A worked illustration of why it matters: a draft email to a buyer says "the seller has disclosed that the roof was replaced in 2019". The disclosure says "roof repaired, date unknown". The model has resolved an ambiguity in the seller's favour, and a buyer who relies on the email may make an offer on a false premise. Step two catches it.

Timing routine writing in a six-agent brokerage

Illustrative. A small brokerage with six agents and a part-time administrator adopts the policy. In the first month the owner measures one thing: minutes spent writing routine messages per agent per week, before and after.

• Follow-ups after showings: eight a week at ten minutes each, falling to four minutes with prompt 1 and the check. Saving: 48 minutes.
• Disclosure summaries: two a week at thirty minutes each, falling to fifteen with prompt 2 and the check. Saving: 30 minutes.
• Re-engagement notes: ten a week at five minutes each, falling to two. Saving: 30 minutes.

Roughly two hours per agent per week in this illustration, of which about a third is spent on checking that was not done before. The honest gain is the consistency and the audit trail, not just the minutes. These are made-up figures; time your own team for a fortnight and see.

A disclosure summary, marked up

To make the owner's review concrete, here is an illustrative excerpt of what prompt 2 returns for a seller's form (invented content), and what a careful agent changes.

Structure: roof repaired (date not stated); no known foundation movement; basement has had moisture "in the past". Systems: furnace serviced annually; electrical panel upgraded; plumbing "no known problems". Environment: no known contamination. Legal: boundary fence shared with neighbour; no known disputes. Unclear or missing: roof repair date; what "in the past" means for the basement moisture; whether the panel upgrade was permitted work; no reference to the attached inspection report mentioned in the form.

This is a good draft. Three things the agent does. First, they open the form and confirm "no known foundation movement" is what the seller ticked, not what the model inferred from a blank. Second, they check the quotation marks: "in the past" is the seller's phrase, and the summary rightly keeps it. Third, they notice that the model did not say the moisture was minor, which is correct, and they do not add that reassurance themselves. The client email that follows says only: "Here is a summary of what the seller has disclosed. Four items are unclear and I will ask about them. This summary is not advice, and the disclosure document is the record." Questions about whether the moisture matters go to the inspector or the supervising broker.

A one-page brokerage rule you can adapt

If you want the policy as text, this is a workable starting draft. It is illustrative, and your supervising broker or lawyer should adjust it to your licence terms and local rules before it is issued.

Agents may use the brokerage's ChatGPT workspace to draft emails, follow-ups and plain-language summaries. Agents must not use personal AI accounts for any client matter. Before pasting any text, agents remove names, addresses, contact details, financial and identity details, and personal circumstances, and use the brokerage placeholders. Agents must not ask AI to interpret contracts, judge disclosures, state legal requirements, predict prices or recommend offers. Every message drafted with AI is read in full and checked against its source by the sending agent, who is responsible for its content. Slips are reported to the broker the same day and are treated as a learning matter unless repeated. The broker reviews a sample monthly.

Note the last two sentences. Without a reporting culture, you hear about a slip from a complaint instead of from the agent who made it.

A first-week sign-off before agents get workspace access

New agents get access to the brokerage workspace only after a short sign-off, run by the broker or a senior agent in their first week. It has three parts, each with a pass mark.

• Redaction test. The agent redacts an invented email thread containing six identifiers: two names, a street address, a phone number in the signature, a loan figure and a mention of a divorce. Pass: all six replaced with brokerage placeholders. The signature phone number is the one most often missed.
• Summary check. The agent runs prompt 2 on an invented disclosure with three planted ambiguities, then finds each one in the form itself. Pass: all three located and listed for follow-up, none described as minor.
• Refusal drill. The agent asks the Project whether an inspection finding is serious, reads the refusal line, and says who they would ask instead. Pass: they name the supervising broker or the brokerage's lawyer, not a colleague's opinion.

Record the date and result next to the agent's name in the policy file. The test materials are invented on purpose: a training exercise is exactly where pasting a live client file feels harmless, and it is not.

Listings, photos and marketing copy

Listing descriptions are a separate risk, because a statement about a property is a representation to the public. A model will happily call a kitchen "newly renovated" or a street "quiet", neither of which it knows. If agents use ChatGPT for listing text, require that they supply the verified features as a list, forbid adjectives about condition, and check every claim against the listing sheet. The related tutorial on AI listing mistakes that can mislead buyers lists failures to look for. If agents use AI image editing, the virtual staging tutorial covers what to disclose.

When a mistake gets out

It will happen eventually: a message goes to a buyer with an invented fact. A good response is quick and calm.

• Correct in writing, soon. Send the client a short correction with the right fact and the source, before they rely on it.
• Tell the supervising broker. They decide whether anything further is needed, and some situations need a lawyer's view.
• Find the step that failed. Was the check skipped, the prompt weak, or the source ambiguous? Fix that step, not the person first.
• Update the Project instructions if the failure suggests a new "never" line. An example: after a model stated a school catchment as fact, add "never mention schools or catchments".
• Note it for the monthly review. Patterns matter more than single events.

Keep the policy to one page you can audit

Resist the urge to write a forty-page document or ban everything. Policies that are too tight are ignored, and policies that are too loose are not policies. Leave out model names and feature lists, which change monthly, and refer instead to "the brokerage's approved workspace". Leave out anything you cannot audit. And leave the legal interpretation to a qualified adviser: if you are unsure whether a rule fits your local requirements, ask a lawyer or your regulator and adjust the page. A short document that agents have actually read protects the brokerage more than a long one in a drawer.

Transaction coordinators and shared inboxes

Larger brokerages often have a coordinator who writes many routine messages on behalf of several agents: reminders that an inspection report is due, requests for signed forms, confirmations of appointment times. This is the best place to standardise, because one person writes the same ten messages for dozens of files. Give the coordinator a Project with those ten templates, and require a fact block for each file (dates and documents supplied from the transaction system, not from memory). The audit for coordinators is the same as for agents but with a higher sample, since one slip repeats across files. Illustrative sum: a coordinator handling twelve open files sends about thirty routine messages a week at four minutes each, two hours; with templates and the check at one and a half minutes each, forty-five minutes, a saving of roughly an hour and a quarter. Use that time for the messages that need judgement.

What the owner audits each month

Policy without audit decays in weeks. Fifteen minutes a month is enough.

1. Pull five sent messages per agent that were drafted with ChatGPT. Tick: placeholders replaced, no unsupplied facts, no advice language, check stamp present.
2. Ask each agent for one prompt that did not work well. Add the lesson to the Project instructions.
3. Check workspace membership: leavers removed, no unknown members. OpenAI's workspace documentation says a removed member's chats on a Business workspace are kept indefinitely, and reassigning their projects does not open their private conversations to the owner, so make sure final messages already sit in the transaction file before an agent leaves.
4. Review incidents: any redaction slip, any message corrected after sending. Treat them as system problems first, not people problems.

Brokerage-level rules about contracts and contingencies deserve their own treatment; AI real estate contract review, deadlines and contingencies goes into what a model can and cannot do there.

Where this policy will be tested

Expect three pressure points. First, a busy Friday when an agent pastes an unredacted email to save two minutes. The prevention is making redaction easy and normalising reports of slips. Second, an agent who asks ChatGPT "is this a problem?" about an inspection finding and relays the answer. The prevention is the refusal line and a clear route to the supervising broker. Third, a client who asks directly whether the agent uses AI. Decide the honest answer in advance: that drafts may be prepared with software, that the agent reads and takes responsibility for everything sent, and that advice comes from licensed people.

Keep the whole thing humble. ChatGPT will not make agents better at the parts of the job that earn trust, such as judging a neighbourhood, reading a seller's mood or negotiating. It can free time from the writing, as long as the writing stays honest. The brokerage's job is to make that the easy path.

Brokerage owners' questions about agents using ChatGPT

Can an agent ask ChatGPT whether a disclosure is a problem?

No. A model can summarise what a disclosure says and list its items, but judging whether an item matters, what must be disclosed or what it means for a transaction is for the agent's supervising broker or a lawyer. Make that line explicit in the brokerage policy and in the saved instructions.

What should agents redact before pasting a document?

Names, addresses, contact details, ID numbers, financial details, loan information, and anything about health, family or legal disputes. Replace each with a consistent placeholder. If a document cannot be redacted without losing the point, do not paste it; read it yourself or ask the supervising broker.

Should the brokerage ban personal ChatGPT accounts?

Prefer a brokerage-managed workspace on a business plan, where business content is not used for training by default, and say in writing that client information stays out of personal accounts. Bans are hard to police, so pair the rule with an approved alternative that is easy to use.

Do clients need to be told that agents use AI?

Check your regulator's expectations, your listing and buyer agreements, and local practice. Some brokerages choose plain disclosure in their terms that drafts may be prepared with software and are reviewed by the agent. Ask your supervising broker or lawyer before settling the wording.

Further reads

• AI Lead Generation for Real Estate Agents: No Bought Lists Needed — Lead generation without bought lists.
• Best AI CRM for Real Estate Agents: Follow Up Boss, Lofty, BoldTrail — Where follow-ups live once drafts leave ChatGPT.
• AI Receptionist for Realtors: Cost, Call Routing and Showing Cover — Cover for showing calls without hiring.
• Virtual Assistant or AI for a Real Estate Agent: What Goes Where — What goes to a person and what goes to AI.
• AI Voice Agent for Real Estate Follow-Up: Setup, Scripts, TCPA — Rules and scripts for AI voice follow-up calls.
• Shadow AI: How to Stop Staff Pasting Client Data Into Free Tools — Stop staff using unapproved tools with client data.
• How to Anonymise Client Data Before You Paste It Into AI — A fuller redaction method for long documents.
• Rechat AI for Brokerages: What Lucy Does and What It Costs — Lucy's features from Rechat's own help pages, the pricing Rechat does not publish, and a 30-day pilot plan for a ten-agent office.
• AI Transaction Coordinator: Which Real Estate Tasks to Automate — Sort transaction paperwork, reminders and chasing into safe AI drafts and human decisions before automating a deal.
• Real Estate Agent Recruiting With AI: Outreach and Onboarding Plan — A four-touch recruiting sequence, an onboarding checklist and a first-30-days plan for a brokerage, with the compliance lines AI must not cross.
• Real Estate Newsletter With AI: A Market Update You Can Verify — A monthly market-update newsletter drafted with AI, with a figure register that gives every number a source, a date and a definition before you send.
• Real Estate Sphere of Influence: A Past-Client Touch Plan With AI — A three-tier contact list, a 12-month touch calendar and AI-drafted templates for anniversaries and market notes that keep past clients warm without spam.
• Real Estate Wire Fraud and Voice-Clone Calls: A Call-Back Script — An office call-back procedure for wire instructions, with a staff script, a client notice and a written policy that holds up even against a cloned voice.
• ChatGPT Ads for Realtors: Access, Cost and Fair-Housing Rules — What a realtor may and may not advertise on ChatGPT, what a first month costs, and the fair-housing checks for ad copy and targeting.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI's ChatGPT pricing page, its workspace user-lifecycle documentation and its custom GPT migration guide. Disclosure duties, licensing rules and contract content come from your regulator, your supervising broker and your lawyer, not from a chatbot.