ChatGPT for Estate Agents: Rules for Vendor Updates and Feedback

An estate agent can use ChatGPT for vendor updates, viewing feedback and chasers without errors if the owner sets three rules: facts come only from the CRM record pasted in, no buyer or vendor personal data goes into the chat, and a named negotiator checks every message before it is sent. ChatGPT drafts; people verify and send.

Those rules are the owner's job, not the negotiator's. If ten staff each invent their own prompting habits, the branch carries ten sets of risks and has no way to audit them. The job is to set the rules once, give staff a few approved templates, and check a sample every month.

Follow me on Instagram@sagnikteaches

Why rules matter more than prompts in an agency

A vendor update is a small document with a large consequence. Vendors read it closely, forward it to their partner, and quote it back when a price drop is discussed. A single wrong number (a viewing count, an offer level, a buyer's position) damages trust far more than a late update would have. ChatGPT makes wrong numbers more likely, not less, because it writes fluent sentences even when it has no facts, and fluent sentences do not get questioned.

Connect on LinkedInSagnik Bhattacharya

The risks fall into four groups, and each gets a rule below:

Subscribe on YouTube@sagnikteaches

• Invented facts: interest levels, "strong feedback", market claims.
• Leaked data: buyers' names and numbers, vendors' financial situation, sensitive reasons for selling such as divorce or probate.
• Tone: messages that sound like a different person, or over-promise ("we expect an offer shortly").
• Accountability: nobody quite owns the message because "the AI wrote it".

Choose the account before the policy

Staff using personal accounts is the first problem to solve. You cannot set data settings, see what was pasted, or end access when a negotiator leaves. A business plan gives you one workspace and admin controls, and OpenAI's business-data page says Business, Enterprise and Edu content is kept out of model training by default. Personal Free, Go, Plus and Pro accounts work the other way: OpenAI's pricing table shows their chats feeding model training until each person opts out. Seat prices change, so read the current figure on OpenAI's pricing page; the same page says Business plans start at two users, so even a two-desk branch qualifies.

If you do not buy seats yet, the rule must be explicit: no client data in personal accounts, and anyone who uses one switches off model improvement in data controls. Whichever route you choose, put it in writing and have staff initial it.

Rule 1: facts only from the record

Every prompt gets the same opening fence. Give it to staff as a snippet they paste before the CRM notes:

Use ONLY the facts below. Do not add any statement about buyer interest, market conditions, likely price or what the vendor should do. If something is missing, write [NEED: item] instead of guessing. Do not include any buyer's name, phone number or email.

Then the facts: number of viewings, dates, the anonymised feedback lines, any offers received (amount and buyer position, not name), and the date of the last contact. The model turns them into prose. The negotiator reads the result against the facts. Any sentence that cannot be traced to a line is deleted.

Example: a vendor update that holds only what is true

Say the CRM shows four viewings this fortnight on a three-bedroom house, two from first-time buyers and two from couples upsizing, one repeat viewing booked, no offers. Anonymised feedback: "kitchen dated", "garden loved", "price slightly high for condition". Prompt: "Write a vendor update of under 150 words, warm and honest, covering these facts, with a closing line suggesting a call on Thursday."

A good draft (illustrative) reads: "Hello [first name], quick update after a busy fortnight. We held four viewings, including one second viewing booked for Saturday. Buyers liked the garden. Two mentioned the kitchen feels dated, and one felt the price is slightly high for the condition. No offers yet. Could we speak on Thursday about whether to adjust anything?" Everything in it traces to the record. A weaker draft would add "we are confident of an offer soon", which has no source, so it is deleted.

Rule 2: what never goes into the chat

Print this as a one-page list and keep it by every desk. It is the owner's most important document.

• Buyers' and vendors' full names, phone numbers, email addresses and home addresses (use "the buyer", "[first name]").
• Anything about a vendor's finances: mortgage balance, arrears, a need to sell quickly.
• Sensitive reasons for a sale: divorce, bereavement, illness, redundancy.
• Identification documents, proof of funds, anti-money-laundering records.
• Offer details that could identify a buyer, such as "the couple from number 12".
• Photographs of people, or interior photographs that show security details.
• Login credentials and portal passwords.

The property address itself is usually public from the listing, but treat it with care when combined with personal circumstances. If staff are unsure, the default is to leave it out. "A three-bedroom semi" is enough for the model to write the message.

Rule 3: the sending check

The negotiator who sends the message is the author, whatever drafted it. Give them a five-item checklist before pressing send:

1. Every number matches the CRM (viewings, dates, prices).
2. No buyer is identifiable.
3. No promise or prediction about price, timing or buyer intent.
4. The tone sounds like our branch and like this negotiator, not a brochure.
5. The next step has a date.

The fifth item is the one that improves vendor relations. A vendor update that ends with "a call on Thursday at 10" is more useful than ten paragraphs of detail.

Viewing feedback: from scrappy notes to an honest summary

Feedback is the part vendors care about most and agents write least well, because it can sting. ChatGPT helps by organising a pile of comments and suggesting kinder wording, but it must not soften the substance. A vendor told "buyers are loving it" while no offers arrive will lose confidence in you.

Example: sorting ten feedback comments

Say ten viewings produced these anonymised notes: "lovely light" (3 times), "kitchen needs updating" (4), "price too high" (3), "road noisy" (2), "great school catchment" (2), "needs decorating" (3). Prompt: "Group these comments by theme, count each, and present them as a table for the vendor with the heading Positive, Neutral and Concern. Do not interpret. Then suggest three neutral phrasings for the concern items." The table is factual; the phrasings let you say "four of ten viewers mentioned the kitchen" instead of "the kitchen is a problem".

Example: the awkward message about price

When feedback says price is the barrier, the message needs care. Ask for two versions of a paragraph linking the feedback to a price conversation: one gentle, one direct. You choose. The model must not recommend a figure, because your valuation judgement and the local sales evidence sit with you. A rule worth adding: ChatGPT never suggests a price.

Example: tone check on a draft you wrote yourself

Not every message starts with the model. Paste your own draft (anonymised) and ask: "Does this contain anything that sounds like a promise or prediction? Does it sound defensive? List each issue with the sentence." Used as a reviewer, ChatGPT is a second pair of eyes, and the risk of invented facts is lower because the facts are yours.

Chasers that stay polite

There are three common chasers in a branch: the vendor who has gone quiet, the buyer who viewed and did not respond, and the solicitor's office with a late document. Each gets a template that staff fill in, rather than a free-form prompt.

Write a short chaser (under 70 words) to a [vendor / buyer / solicitor's office]. Purpose: [one line]. What I need: [one item]. By when: [date]. Tone: friendly, direct, no pressure language. Do not mention any other party. Sign off as [negotiator first name].

A realistic output for a buyer who has viewed: "Hello [first name], thanks for coming to see the house on Saturday. I wondered whether you had any thoughts, or any questions I can answer. If it is not for you, that is no problem and I would like to know so I can keep looking for something suitable. [negotiator]." The phrase "no problem" is useful: it lowers the cost of replying. Staff edit; they do not accept the draft as written.

For chasers to a buyer, check that your own marketing consent and any local rules on contacting buyers are respected. ChatGPT does not know your jurisdiction's rules, and the agency owner remains responsible for them.

Example: reporting an offer without naming the buyer

Offers are where accuracy and confidentiality collide. Say an offer of 4 per cent under the asking price arrives from a buyer who is selling their own home, which is under offer already. The negotiator pastes: "Offer received Tuesday, 4 per cent below asking, buyer has a sale agreed, mortgage in principle in place, wants to exchange within 10 weeks." The prompt asks for a vendor message of under 120 words that states the offer, the buyer's position and a proposed call, with no recommendation to accept or reject.

The draft (illustrative) says the offer is "reasonable" and that the vendor "should consider it carefully". Both phrases are opinions the model added. The negotiator deletes "reasonable", because whether it is reasonable is the negotiator's advice and belongs in the phone call. Add to the rules: ChatGPT reports offers; it never characterises them.

Example: a message after a price reduction

Another common update follows a reduction. Provide the old and new price, the date of the change, the portals updated and the number of enquiries in the week before. Ask for a short note that explains what changes and what happens next. The model should not explain why the price was reduced beyond the agreed reason you give it, such as "following feedback and the valuation review on Monday". A frequent error is a sentence like "this should attract more interest", a prediction the rules ban. Replace it with the next factual step: "I will review enquiry numbers again next Friday."

Leavers, new starters and the rule document

Two moments test a policy: someone joins and someone leaves. For starters, give them the one-page no-paste list, a login to the shared Project, and a first week where the owner or a senior negotiator reads every drafted message before it is sent. For leavers, remove access to the workspace on the same day as email, export or archive anything the branch needs, and know what happens to their history. On ChatGPT Business, OpenAI's help centre says a removed member's chats and files are kept indefinitely and are not visible to the workspace owner, while their projects pass to an owner. Anything the branch needs must therefore live in a shared Project or the CRM before the leaver goes, and nobody should keep client details in private chats.

Keep the rule document to one page. A suitable outline, filled in for a branch (illustrative): purpose in one sentence; the approved account; the three rules (facts from the record, no personal data, named sender checks); the approved templates and where they live; what to do if something wrong is sent; who to ask. A policy nobody reads protects nobody, so keep it short enough to be pinned to a wall.

Using ChatGPT to improve the process itself

The owner can also use ChatGPT on the system rather than on individual messages. Paste ten anonymised sent updates and ask: "Which phrases appear in more than half of these? Which sentences make claims about interest, price or timing? Which updates end without a next step?" The answers show where the branch's habits are weak. A second use is training. Ask for five realistic but fictional feedback scenarios, including a difficult vendor who disputes the price, and use them in a monthly ten-minute exercise in which staff draft, then check, a reply. Because the scenarios are invented, there is no personal data at all.

Tone rules that keep the branch's voice

A branch's voice is a competitive asset, and unedited AI drafts flatten it. Write down five tone rules and add them to a Project (ChatGPT's workspace for shared instructions and files) so every draft starts closer to your style. Examples: first names, short sentences, "we" for the branch and "I" for the negotiator, no exclamation marks, never "stunning", "must-see" or "deceptively spacious". Ask staff to paste two of their own past updates into the Project as style samples, with details removed, so the model imitates the branch rather than a generic agent.

Project sharing is on every ChatGPT plan, Business included, per OpenAI's help article. On a business plan, the owner can set up one Project per task (vendor updates, viewing feedback, chasers) and keep the approved templates inside it. Staff then open the Project rather than writing prompts from scratch, and the owner updates the instructions in one place when a rule changes. This is the single most useful setup step for a branch. It turns good practice into default behaviour.

A month-one rollout for a six-person branch

An illustrative plan for a branch with a manager, four negotiators and one administrator.

The pilot matters because errors show up in the first twenty messages, when they are cheap to fix. Time saved tends to be modest; as a rough illustration, an update that took 15 minutes to write might take five, plus two to check. Time a few yourself rather than assuming.

What owners find when they sample updates

Here is a short list of problems that a monthly sample typically reveals, with the fix for each.

• The phrase "strong interest" appearing with two viewings. Fix: add "do not describe interest levels; give counts only" to the instructions.
• Buyer positions mixed up: a chain-free buyer described as needing to sell. Fix: staff paste the position in a fixed format, "FTB, cash, needs to sell, chain-free".
• Dates wrong after a bank holiday. Fix: include the weekday with every date.
• Identical wording across vendors, which vendors notice if neighbours compare. Fix: ask for one variation in structure and keep the facts different.
• Over-long messages. Fix: set a word limit in the prompt.

Where marketing copy for a listing is also in play, material facts about a property (tenure, planning restrictions, known defects) must come from verified records, and this tutorial on material information checks covers that side. Descriptions are a separate job from updates; how agents use AI for property descriptions is the place for that, and preparing valuation appointments covers the earlier stage.

Correcting a wrong viewing count after it reaches a vendor

Plan for it. An illustrative case: an update tells a vendor that three viewings happened when the CRM shows two, because a cancelled slot was counted. The vendor tells a neighbour. The owner's procedure should say: correct it the same day by phone, then in writing; record what went wrong; and check whether the prompt or the checking step failed. Usually it is the checking step, in which case the fix is a reminder to the negotiator, not a new tool. Keep a simple log of corrections. If the log stays empty for a quarter, the sampling is probably too light.

Finally, remember that AI use is part of your professional duty to clients. If a vendor asks whether you use AI to draft messages, answer honestly, and explain that a named person checks everything that is sent. For the wider question of where else in the branch AI fits, this stack by job is a reasonable next read.

Questions branch owners ask before allowing ChatGPT

Should I let negotiators use their own ChatGPT accounts?

It is safer to provide one business account and ask staff to use only that. Personal accounts sit outside your control, so you cannot set data settings, share prompts or remove a leaver's access. If you cannot fund seats yet, write a rule that no client data goes into personal accounts.

Can ChatGPT write a vendor update from a CRM export?

Yes, if the export holds the facts and you tell it to use only those. The model should never fill gaps with guesses about interest levels or market conditions. Any sentence without a source line in the export gets deleted before sending.

What about buyers' names in viewing feedback?

Leave them out. A vendor update needs the substance of the feedback and the buyer's position (for example, first-time buyer or needs to sell first), not the buyer's name, phone number or email. Paste anonymised notes only.

How do I check staff are following the rules?

Sample five sent updates a month. For each, ask the negotiator to point to the CRM record behind every fact. A short monthly check does more than a long policy document, because staff know it will happen.

Further reads

• AI Lead Generation for Real Estate Agents: No Bought Lists Needed — Lead generation that respects the same data rules.
• Best AI CRM for Real Estate Agents: Follow Up Boss, Lofty, BoldTrail — Choosing the CRM that holds the facts behind each update.
• Virtual Assistant or AI for a Real Estate Agent: What Goes Where — Deciding what goes to a person and what goes to AI.
• AI Receptionist for Estate Agents: Cost, Call Flow and Handover — Call handling, once your written rules are settled.
• AI Marketing for Real Estate Agents: A Four-Week Content Plan — A four-week content plan that uses the same rules.
• Material Information in Listings: What Estate Agents Must Capture — The Part A, B and C material information checklist for listings, and an AI workflow that chases the gaps without guessing.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI business data privacy page, ChatGPT pricing page, OpenAI help-centre articles on projects and on data retention when a member is removed (checked 4 November 2026).