ChatGPT for Mortgage Brokers: Emails and Case Summaries, Not Advice

A mortgage broker can use ChatGPT for client emails, case summaries, document chasers and referral partner updates, as long as nothing in the draft recommends a product, judges affordability or quotes lender criteria. Keep client identifiers out, let an adviser read every message, and treat the output as wording, never as a recommendation.

The easy way to stay on the right side of that line is to pair every prompt with a guardrail sentence that names what the draft must not contain. Each of the ten prompts below carries one, with a realistic output and the edit an adviser would still make before sending.

Follow me on Instagram@sagnikteaches

The line between wording and advice

Advice, in practice, is anything that tells a particular client what to do with a particular product, or implies that they will be accepted for it. Wording is everything that surrounds that: scheduling, document requests, explaining a stage, summarising what was said. The test for any draft is a single question: could a client read this and act on it as if it were a recommendation? If yes, it is for the adviser to write.

Connect on LinkedInSagnik Bhattacharya

Three kinds of sentences cross the line most often when a model drafts for you:

Subscribe on YouTube@sagnikteaches

• Reassurance about outcomes: "You should have no problem getting approved."
• Soft product steers: "A fixed rate would suit you given your circumstances."
• Criteria statements: "Lenders usually accept income of at least four times the loan."

Models produce all three readily, because they are common in the text they learned from. Your prompts need to forbid them in words, and your review needs to search for them. For what else sits on each side of the line across a whole case, see what mortgage brokers can automate with AI and what stays advice.

A case-communications Project with the guardrails built in

Retyping a guardrail in every prompt guarantees someone will forget it. Create a ChatGPT Project called "Case communications" and paste the standing rules into its instructions. With custom GPTs on their way out (OpenAI is moving them to plugins), the Project is where this belongs. Illustrative instructions:

You draft messages for a mortgage brokerage. You are not an adviser.
Never: recommend, rank or compare products or lenders; say or imply that a client will be accepted, declined or offered any rate; comment on affordability or borrowing capacity; state lender criteria; give rates, fees or dates unless the user supplied them in this chat.
If a request needs any of those, reply only: "This needs an adviser's judgement. I can draft the surrounding wording."
Use [CLIENT FIRST NAME], [ADVISER NAME] and [FIRM] placeholders. British English, plain words, under 150 words unless asked.

Use a business plan where the whole team works, since OpenAI's plan page promises no training on business data by default, whereas individual plans rely on each person keeping the training switch off. The placeholder rule for a brokerage is short: paste no names, addresses, employers, account numbers or dates of birth, and give income as a band if you need it at all.

Prompts for client emails

1. The document chaser

Write a short email asking a client for these documents: [LIST]. Say we need them by [DATE] so the application is not delayed.
Explain in one sentence each why we need them, in plain English.
Guardrail: do not say whether the documents will be enough for approval, and do not mention any lender's criteria.

Realistic output: a friendly four-paragraph note with a bulleted list and one-line reasons such as "so we can confirm your income". What to fix: models often add "this will help us secure the best rate for you", which is a promise about outcome. Cut it. Keep the reasons about the process, not the result.

2. The stage explainer

Write a plain-English email explaining what happens at the [STAGE] stage of a mortgage application and what the client needs to do next.
Use the steps I give you: [STEPS]. Do not add steps.
Guardrail: do not give timescales unless I have given them; do not predict the lender's decision.

Use it for stages such as valuation, offer and completion. The "do not add steps" line matters because models fill gaps with generic steps that may not match your process. Before: "The lender will review your file and typically respond within a few days." After: "Once your documents are with the lender, I will tell you as soon as I hear anything." The second does not promise a timescale.

3. The delay update

Write a calm, honest update telling a client their application has been delayed because [REASON, supplied by me]. Say what we are doing and when I will next be in touch: [DATE].
Guardrail: do not speculate on how the delay will affect approval, rate or completion date.

Delay messages are where reassurance creeps in. Read for "don't worry" and "this is quite common". A client mid-purchase hears these as promises. Keep the message factual: what happened, what you are doing, when you will next update. For more on keeping clients informed during an application, see whether AI can keep mortgage clients updated.

4. The post-meeting recap

Turn my notes into a recap email to the client: what we discussed, what each of us will do next, and any dates.
Notes: [YOUR NOTES, names removed].
Guardrail: use only what is in my notes; do not add recommendations, opinions on affordability or product names I did not mention.

Realistic output is a neat list of actions. The failure is quiet: your note says "client considering fixed vs tracker" and the draft says "we agreed a fixed rate was the best option". Check every sentence against your notes, and have the adviser who ran the meeting approve it.

5. The polite "we cannot help with this case"

Write a kind, brief email telling an enquirer we are unable to take on their case at the moment, without giving reasons about their finances. Suggest they speak to another adviser.
Guardrail: do not comment on their credit, income or eligibility.

This stops two risks at once: a model inventing a reason ("because of your credit history") and an enquirer treating a decline as a judgement on their eligibility.

A full sample, marked up

To show what review looks like, here is an illustrative first draft from prompt 1 (invented, for a purchase case), followed by what the adviser changed.

Hi [CLIENT FIRST NAME], I hope you are well. To keep your application moving, could you send me your last three months of bank statements and a copy of your signed tenancy ending notice by [DATE]? We need the statements to confirm your income and spending, and the notice so we know your moving timeline. Once these are with us we will be in a strong position to secure the best rate for you. Many lenders also like to see a clean statement history, so avoid large unusual payments in the meantime. Kind regards, [ADVISER NAME]

Three edits. "Strong position to secure the best rate" is an outcome promise and a rate claim, so it goes. "Many lenders also like to see a clean statement history" states lender criteria from nowhere and hands the client behavioural advice, so it goes too (if the adviser wants to tell the client to avoid large transfers, that is the adviser's own instruction, written deliberately). "Signed tenancy ending notice" was invented: the adviser asked for two documents and the model added a third. The cleaned version keeps the first two sentences, adds one line about the deadline, and sends. The lesson is that the guardrail reduces these slips but never removes them, which is why the review is part of the process.

Prompts for case summaries

Case summaries are the strongest use, because they condense your own material rather than inventing anything. Feed the model your notes, never the documents with identifiers. A summary is for internal handover, compliance review or checking that the file is complete.

6. The handover summary

Summarise these case notes for a colleague taking over. Use headings: Client's goal, Current stage, Documents received, Documents outstanding, Next action and owner, Open questions.
Notes: [NOTES, anonymised].
Guardrail: do not add anything not in the notes; if a heading has no information, write "Not in notes". Do not suggest products or comment on suitability.

The "Not in notes" rule is the important part. Without it, a model fills empty headings with plausible text. A realistic output has two or three "Not in notes" lines, which is the point: they show where the file is thin.

7. The file-completeness check

Here is my checklist of what a complete file needs: [CHECKLIST]. Here are my notes of what we hold: [NOTES].
List, in a table, each checklist item with: Held / Not mentioned. Quote the line from my notes that supports each "Held".
Guardrail: do not judge whether any document is acceptable to a lender.

A "Held" without a supporting quote is a red flag; the model has assumed. This works as a second pair of eyes before submission, though not as a substitute for the actual document check. Our enquiry to fact-find tutorial shows where this sits in a whole case.

8. The call-note tidy

Tidy these rough call notes into a file note with: Date, Who was on the call (roles only), Topics, Client's stated intentions, Actions.
Guardrail: keep client statements marked as "client said"; do not rephrase uncertain statements as facts.

Uncertainty is the thing to protect. "Client thinks they may get a bonus in spring" must not become "client expects a spring bonus". The marked "client said" convention keeps provenance clear.

Prompts for referral partners

Estate agents, solicitors and accountants who refer clients want to know where a case stands without receiving private details. Draft status updates at the level the referrer is entitled to, which is usually the stage and the next date, not the finances. The dedicated tutorial on referral partner updates drafted with AI goes deeper; here are two prompts.

9. The stage-only update

Write a two-sentence update to a referral partner saying a shared client's application has reached [STAGE] and the next step is [NEXT STEP] by [DATE].
Guardrail: no financial details, no names beyond [CLIENT FIRST NAME], no prediction of the outcome.

10. The thank-you

Write a short thank-you note to a referral partner for sending a client to us, mentioning that we will keep them updated at each stage.
Guardrail: do not mention any commission, fee or payment, and do not promise how quickly cases will complete.

Check the thank-you against whatever your terms say about disclosing referral arrangements; that is a compliance question for you, not a drafting one.

A realistic slip with partner updates, and how it showed up: an adviser pasted fuller notes than usual into prompt 9, and the illustrative draft to the estate agent read, "Good news, [CLIENT FIRST NAME]'s income checks out comfortably and the lender should approve this week." Two problems in one sentence. It discloses the client's financial position to someone who has no right to it, and it predicts a lender decision. The adviser caught it because the guardrail line says "no financial details" and the draft plainly broke it. The fix was procedural rather than editorial: for partner updates, paste only the stage and the next date, never the case notes. If the model has nothing else to work with, it cannot leak it.

One adviser's day on a purchase case, timed

Illustrative. A solo adviser has six live cases and spends roughly 90 minutes a day on routine messages. On one purchase case the day looks like this.

1. 8:45 Prompt 1 chases two missing bank statements. Draft in 40 seconds; the adviser deletes the line about "securing the best rate" and sends. Time: 3 minutes against 8.
2. 11:30 After a call, the adviser pastes shorthand notes (client replaced with [CLIENT]) into prompt 8, gets a file note, and fixes one sentence that turned "might" into "will". Time: 6 minutes against 15.
3. 14:00 The estate agent asks where the case stands. Prompt 9 gives two sentences. Time: 1 minute.
4. 16:30 The adviser runs prompt 6 on three cases to prepare Friday's review. Time: 10 minutes against 25.

Total saved on that day: roughly 29 minutes, which over a five-day week is about two and a half hours. These are illustrative figures, not a benchmark; your own timings will differ, so time a week before and after.

A two-minute word search before anything is sent

It takes about two minutes and catches the slips models make most often. Make it a habit that the adviser, not the admin who drafted, signs off anything containing a product, a rate or a decision.

Edge cases that need an adviser's own voice

• Joint applicants who disagree. Drafts that summarise two people's positions risk presenting one as the agreed view. Keep each person's statements labelled in notes and in the summary.
• Clients who may be vulnerable. Bereavement, illness or financial difficulty call for a human tone and your firm's vulnerable-client process. Do not use a model for the first message, and never paste the circumstances.
• Clients who read English as a second language. A "simpler wording" pass is useful, but have someone who knows the client check that simplified text has not changed meaning. Tone and idiom shift easily.
• Complaints. A complaint reply has formal requirements and potential consequences. Draft it yourself or with a compliance colleague.
• Anything involving a decision already taken. A decline, a withdrawal, a changed offer: give the model the facts in one line and let it draft only the courteous frame around them.

A fortnight's rollout for four to eight advisers

A solo adviser can keep the rules in their head. A team of four to eight cannot. A workable rollout takes about a fortnight.

1. Week one, one Project, two prompts. Start with the document chaser and the handover summary. They are the lowest risk and the most repeated.
2. Week one, a review buddy. Each person has a colleague who reads one sent AI-drafted email a day for the first week.
3. Week two, add recaps and partner updates. By now people have noticed the typical slips, so the guardrails make sense to them.
4. Monthly audit. Pull five sent messages at random per adviser. Tick each against the table above. If more than one in five contains an outcome promise or a rate, stop adding prompts and retrain.

Keep a shared document of prompts that have been tried and which edits they needed. Over a couple of months the list shows which prompts are worth keeping, and which only look useful. If you also use a CRM, choosing a mortgage CRM with AI built in explains which built-in drafting tools deserve the same guardrails.

A down-valuation message written three ways

The same client situation can be written three ways, and only one stays on the right side of the line. Illustrative situation: a client's valuation has come back and the adviser needs to tell them it is lower than the purchase price.

Notice what the model is good for in this example: the courteous frame and the plain structure. The options are for a person.

Plain-English explainers you can reuse

A second source of safe value is generic explanation, written once and reused. Ask the model for a one-page explainer of a process step, such as "what a valuation is and who it is for", then have an adviser edit it against your own process and save it in the Project as reference text. Because it describes the process, not any client's case, it carries no personal data and little advice risk. Check each explainer for statements about how lenders behave and delete them. Over a quarter you can build a small library of eight or ten explainers (valuation, offer, exchange, completion, what documents are for) that cut down repetitive questions.

Three numbers that show whether drafts help clients

Track three numbers for a month: minutes spent on routine messages per week, the share of AI drafts that needed a guardrail fix, and the number of client queries that arrive because a message was unclear. The first should fall, the second should fall as prompts improve, and the third should not rise. If the third rises, your drafts have become shorter but vaguer, and the clients are paying for it with extra calls.

Requests that should hit the refusal line

The refusal line in the Project instructions is there for moments like these. A junior colleague types, "Write an email telling the client why a five-year fix is better for them." The Project replies that it needs an adviser's judgement. That reply is a feature: it is a nudge to hand the case to someone qualified. If a colleague starts rewording prompts to get round the refusal, that is a training conversation, not a software problem.

Prompts that ask about specific products or lenders are a related temptation. Models describe lender criteria from training data that is out of date and unsourced, and criteria change. Look them up in the lender's own materials and write the message yourself. For the tools side of the stage-by-stage picture, best AI tools for mortgage brokers at each stage of a case is a good map.

What goes in the case file, not the chat

Whatever is sent to a client is part of your file. Save the final, reviewed version in your case system; the chat is only where the wording was shaped. If a regulator or a complaint ever asks how a message was written, you want the sent text and the name of the adviser who approved it, not a chat history. Ask your compliance contact whether and how to record that drafts were prepared with software, and check what your professional indemnity insurer expects.

The aim is modest: take routine writing off advisers' desks, keep the advice where it belongs, and make the boundary visible in the tools themselves, so that it does not rely on anyone remembering it at four in the afternoon.

Mortgage broker questions about ChatGPT and advice

Can ChatGPT recommend a mortgage product to a client?

It should not be used for that. Recommending a product is regulated advice that rests on a full fact-find, current lender criteria and your own suitability judgement. A model has none of those and can sound sure while being wrong. Use it for wording, structure and summaries, and keep product choice with the adviser.

Is it safe to paste a client's income into ChatGPT?

Avoid it. Income, employer, address, date of birth and account details are exactly what a draft does not need. Use ranges or labels such as [INCOME BAND], and type the real figures into your own system. Check your data-protection obligations and your client privacy notice before pasting any real client information at all.

How do I stop staff drifting into advice?

Put the guardrail into the saved Project instructions so it applies every time, give staff a refusal line to use when a request needs advice, and review a sample of sent emails monthly. Make it normal for an adviser to approve any message that mentions affordability, rates or lender criteria.

Should I tell clients I use AI to draft messages?

Check your regulator's expectations and your own terms of business. A plain statement that drafts may be prepared with software and are always reviewed by a named adviser is a reasonable approach. Ask your compliance contact before changing client-facing wording.

Further reads

• Is It Safe to Let AI Listen to Mortgage Advice Calls? — Whether AI note-takers are safe on advice calls.
• Using AI to Spot Remortgage Opportunities in Your Client Bank — Finding clients whose deals are about to end.
• HubSpot for Mortgage Brokers: Leads, Referrers and Case Stages — Setting up leads, referrers and case stages in one CRM.
• AI Answering Service for Mortgage Brokers: Cost and What It May Say — What an AI phone answering service costs and may say.
• AI Loan Processing: Catch Missing Documents Before Underwriting — Catch missing documents before a file reaches underwriting.
• Does ChatGPT Train on Client Data? Business vs Free Plans — How business and free plans differ on training.
• How to Anonymise Client Data Before You Paste It Into AI — The placeholder method in full, before anything is pasted.
• AI Tools for Loan Officers: CRM, Follow-Up and Compliance Checks — A CRM-first method for loan officers: compliance fields, guard-railed follow-up prompts, a pre-check step and a log, with no rate quotes from AI.
• Mortgage Broker Social Media With AI: A Month of Compliant Posts — A four-week content calendar for mortgage brokers, with a per-post advertising-rule check, disclosure wording and a review step before anything goes live.
• Can Claude Read Payslips and Bank Statements for a Mortgage Case? — A broker's document-extraction workflow with source pages, salary-credit tie-outs and exception review.
• 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 custom GPT migration guide and its Data controls help article; your regulator's current rules and your lenders' criteria are the authority on advice and suitability.