ChatGPT for Insurance Brokers: Client Emails, Summaries and Limits

A broker can use ChatGPT to draft client emails and to summarise policy wordings, provided identifying details are removed first, a business plan is used, and every statement about cover is checked against the policy document by a qualified person before it is sent. It should never decide what is covered.

The failure is rarely a leaked name. It is a fluent summary that sounds certain, drops an exclusion, and gets forwarded to a client as if it were the insurer's position. Stopping that takes two things: prompts that forbid statements about cover, and a clause-by-clause check of about five minutes per document.

Follow me on Instagram@sagnikteaches

The three jobs worth giving ChatGPT, and the one that is not

Split your week into four kinds of writing. Three suit a language model; the fourth does not, and the dividing line is whether the output creates a statement about cover.

Connect on LinkedInSagnik Bhattacharya

• Safe to draft: chasing documents, renewal reminders, meeting follow-ups, polite declines, explaining a process in plain language ("what happens after you report a claim").
• Safe to summarise, with a check: a policy wording or schedule turned into a one-page plain-English summary, with every point tied to a clause number.
• Safe to restructure: your own rough notes turned into a clean file note or a tidy email thread summary.
• Not for ChatGPT: "Is this loss covered?", "Which policy should this client buy?", "Is this client under-insured?". Those are advice or claims decisions. They belong to a named person with the policy in front of them.

If you want the bigger picture of where a small brokerage saves time, the tutorial on where AI saves time in a small brokerage maps it by task; this page stays on emails and summaries.

Subscribe on YouTube@sagnikteaches

Set the account up before anyone pastes anything

Three decisions come first, because they are far harder to fix after staff have formed habits.

1. Use a business plan, not personal accounts. OpenAI states that content on its business plans is not used to train its models by default. On an individual plan each person has to find the training switch in the data controls and keep it off. One brokerage-wide workspace removes that dependence on memory. A business workspace needs at least two seats, so a one-person practice either pays for two or uses an individual plan with the switch turned off. OpenAI's pricing page lists the Business plan for two or more users, so confirm the current seat terms there before you budget.
2. Write the paste rule in one paragraph. For example: "We paste wordings and clause text. We do not paste names, addresses, policy numbers, dates of birth, bank details, claim narratives or medical information." Pin it where staff open ChatGPT.
3. Keep a Project per job type. A ChatGPT Project holds instructions and reference files for a repeating job, such as "Renewal emails". Put your tone rules and your sign-off in the Project instructions so each draft starts consistent. Check how many files your plan lets a Project hold before you upload a whole library of wordings.

OpenAI is retiring custom GPTs and moving them to plugins, so do not start a new library on a GPT; a Project is the simpler home for a small team. Staff who use personal accounts anyway are the usual leak; stopping staff pasting client data into free tools covers that problem.

Anonymise by replacement, not deletion

Blanking out details leaves the model guessing and your own reader confused. Replace each detail with a consistent placeholder so the draft can be finished locally.

A quick test: could a colleague outside your firm work out who this is from what you pasted? If a risk is unusual enough (a single-site specialist manufacturer, say), generic wording matters more than placeholders. For a longer method, see how to anonymise client data before you paste it into AI.

Summarising a food manufacturer's combined policy, clause by clause

Illustrative case. A broker holds a commercial combined policy for a small food manufacturer. The client has asked, "What does this actually cover, in a page?" The broker pastes only the section text of the wording, with the schedule's names and numbers swapped for placeholders.

You are helping an insurance broker prepare a plain-English summary for a client.
Use ONLY the wording pasted below. Do not use general knowledge about insurance.

Task: summarise each section in no more than four bullet points.
For every bullet give the clause number in brackets.
After the summary, list separately:
1. Every exclusion, limitation or condition you found, with clause numbers.
2. Every defined term that changes the meaning of a section.
3. Anything in the text you could not interpret, or that seems to refer to a document I have not pasted.
Do not say whether any specific event would be covered.

WORDING:
[paste clause text]

A realistic output looks like this (illustrative): a tidy set of bullets for "Material damage" and "Business interruption", each with clause numbers, and a list of exclusions. This is where you look hard. Typical problems found on a read-through:

• The summary says "business interruption is covered for the loss of gross profit" but omits that the indemnity period is a stated number of months, which sits in the schedule that was not pasted. The prompt's third list is there to surface this, and here it did.
• An exclusion for gradual deterioration is merged into a bullet about accidental damage, softening it. Fix by reading the clause itself and rewriting the bullet.
• A defined term, such as "Premises", has a narrower meaning than the everyday word, so a bullet reads wider than the policy.

The broker corrects those three points, adds the schedule figures from the real documents, and sends the result under their own name with a line such as "This summary is a guide; the policy wording governs." In this illustration the summary took about fifteen minutes against perhaps forty from scratch, and that difference only exists if the check below is done every time.

The five-minute verification pass

Every summary goes through the same pass before it leaves the building. It is boring on purpose.

1. Clause match. Open each cited clause and confirm the bullet says what the clause says. If a bullet has no clause number, delete it or find the clause.
2. Numbers and periods. Limits, excesses, waiting periods and indemnity periods come from the schedule, typed by you. The model should never supply a figure.
3. Exclusion sweep. Compare the model's exclusion list against the exclusions section yourself. Missing one is the most expensive error.
4. Verbs. Search the draft for "will", "always", "covers" and "guaranteed". Soften to "the wording says" unless you are certain.
5. Sign-off. A named person approves it. Record who and when in the file.

Client email prompts with a line that blocks cover talk

Emails are the lower-risk use, but they still need rules. Each prompt below pairs a task with a line that blocks the common failure.

Chasing missing documents

Write a short, courteous email from a broker to a client asking for three missing items: [ITEM 1], [ITEM 2], [ITEM 3].
Mention that the insurer needs them by [DATE] to keep the renewal terms.
Tone: friendly and plain. Under 120 words. Do not mention any cover, price or claim outcome.
Use [CLIENT FIRST NAME] and [BROKER NAME] as placeholders.

The final sentence is the guardrail: no mention of cover, price or outcomes keeps the draft administrative.

Explaining a renewal that has gone up

Clients ask why the premium moved, and the honest answer is often "the market", which is thin. Give ChatGPT the reasons you know to be true, not the facts it might invent.

Draft an email explaining that a client's renewal premium has increased.
Use only these reasons I give you, in this order: [REASON A], [REASON B].
Do not add other reasons, market statistics or industry trends.
Offer a call to discuss options. Do not suggest a specific alternative insurer or product.
Under 150 words.

Before: "Premiums are rising across the market due to increased claims costs." After the rewrite: "The insurer has raised your rate because of the revised rebuild value we agreed in March and because your stock limit went up." The second is specific and verifiable because you supplied it. For the renewal routine at scale, see how to automate insurance renewals with AI.

Setting expectations after a loss

This is the riskiest email, so draft only the process, never the outcome.

Write an empathetic email to a client who has just reported a loss.
Cover only: we have noted the claim, the insurer will contact them, what documents to keep (photos, receipts, repair quotes), and that we will update them when we hear back.
Do not say or imply the claim will be accepted, paid, or how long it will take.

Read the output for phrases such as "don't worry" or "this should be straightforward". Strike them. A client who hears reassurance and then gets a declined claim will remember the reassurance. The tutorial on handling claims enquiries with AI gives more on tone.

A polite decline to quote

Write a brief, kind email declining to quote for a prospect whose risk is outside our appetite.
Reason to give: [REASON]. Suggest they ask a specialist broker. Do not name any firm.
Under 90 words.

Meeting notes into a file note

Paste your own bullet points from a client call, with names replaced, and ask for a file note with sections for "Client's request", "Information given", "Actions" and "Open questions". Add the instruction: "Do not add anything that is not in my notes." A reasonable output is a half-page record in the order an auditor expects. The usual mistake is the model tidying an uncertain remark into a firm one: your note said "client thinks the roof may have been replaced in 2018" and the draft says "roof replaced 2018". Search your draft for every date and amount and tick it against your notes.

Mixed policies, cut-off wordings and other summary slips

Translating an insurer's non-renewal letter for a landscaper

Insurers send letters that are dense and defensive, and clients ask the broker to translate them. Take an illustrative letter to a small landscaping business: it says the insurer will not renew the contractors' plant cover on current terms, lists three conditions (a tracker on machines over a set value, a named security standard for the yard, and a higher excess), and gives a deadline. The broker pastes the text with the client name replaced and asks for a three-part note.

Below is a letter from an insurer to a broker's client. Rewrite it as a note to the client in plain English.
Structure: (1) what the insurer has decided, (2) what they are asking for, as a numbered list, (3) the deadline and what happens if it is missed.
Keep every figure, condition and date exactly as written. If anything is ambiguous, list it under "Questions to ask the insurer" instead of guessing.
Do not recommend whether the client should accept.

LETTER:
[paste text]

A good output keeps the three conditions as a numbered list and flags one ambiguity: the letter says "approved tracker" without saying who approves it. That flag is the useful part. The broker rings the underwriter, adds the answer, and the client gets a note they can act on the same afternoon. If instead the draft had quietly written "any GPS tracker will do", the client would buy the wrong device. This is why the instruction "list ambiguities, do not guess" sits in nearly every prompt on this page.

Standing rules for the client-emails Project

Retyping rules in every prompt is how rules get dropped. Put the standing rules in the Project instructions once. A workable set for a "Client emails" Project (adapt the wording to your firm):

You help staff at a small insurance brokerage draft emails.
Always: use British spelling, short paragraphs, plain words, no jargon without a one-line explanation.
Never: state or imply that a claim is covered, accepted or will be paid; quote premiums, limits or dates unless the user supplied them in this chat; recommend a product or insurer.
If a request needs a cover opinion, reply only: "This needs a qualified colleague to answer from the policy."
Use placeholders [CLIENT], [BROKER NAME] and [DATE]. Never ask the user for personal data.

The refusal line matters. It gives staff a visible signal when they drift into advice, and it is easy to spot in review.

A team exercise: find the exclusion the model missed

Written rules fade in a fortnight. A short session with the whole team works better, and you can run it yourself.

1. Five minutes: show the paste rule and the placeholder table above.
2. Ten minutes: give everyone the same anonymised, invented wording extract and the summary prompt. Each person lists the points the model got wrong.
3. Ten minutes: compare lists. Someone usually finds the missed exclusion, and the room sees why the check exists.
4. Five minutes: agree who signs off summaries and where drafts are saved.

Use an extract you have invented, not a real client's, so nobody is tempted to paste live data in the training itself.

Smaller jobs that fit the same rules

• Subject lines: paste a finished body and ask for five subject lines under 50 characters. Pick one; discard any that sound like alarms ("Urgent: cover at risk").
• Reading-age check: ask for the same email rewritten at a simpler level, then compare. Keep whichever says exactly what you mean.
• Checklists for new clients: ask for a list of documents a business of a given type usually needs to gather for a first quote. Treat it as a starting list and strike anything your insurers do not ask for.
• Renewal call scripts: turn your own bullet points into a five-question call outline. Questions only; no answers.

Comparing two quotes without letting ChatGPT pick one

Clients often ask which of two quotes is better. The answer is advice, so ChatGPT must not give it, but it can lay the two documents side by side so that you spot differences faster. Paste the relevant sections of each quote under the labels "Quote A" and "Quote B", with names removed, and ask for a table with one row per feature: limit, excess, main exclusions, conditions, and anything present in one and absent from the other.

Illustrative result for a small print shop: the table shows Quote A with a higher stock limit and Quote B with a lower excess, and flags that only Quote B mentions a condition about alarm maintenance. You still read both conditions yourself, because the model compared phrases, not meaning. A "no equivalent clause found" cell sometimes means the clause sits under a different heading, so search the original before you tell the client a feature is missing.

Then you write the recommendation, in your own words, with the client's circumstances in mind. If you do record a suitability reasoning for the file, it is yours and so is the responsibility. That split is the one to keep: the model finds differences, you decide what they mean for this client.

Removed staff, old chats and the client file

Decide where the approved version of anything sent to a client is kept, and treat the chat as a scratchpad. Copy the final text into your broking or document system, and delete chats you no longer need. Staff changes are why this matters. OpenAI's workspace documentation says that on a Business workspace a removed member's chats are kept indefinitely, and that handing their projects to a colleague does not let the owner read their private conversations. So when a broker leaves, whatever they summarised in ChatGPT is neither deleted nor available to you. If the approved text already sits in the client file, nothing is lost; if it lives only in a departed colleague's chat history, it is effectively gone.

Timing chasers and summaries for a fortnight

For two weeks, note the minutes spent on each repeated email or summary before and after. Illustrative sum: a six-person brokerage sends about 60 chasing and renewal emails a week at six minutes each, 360 minutes. If drafts cut that to three and a half minutes with the check included, the team saves about 150 minutes a week, roughly 2.5 hours. That is worth having but not life-changing, which is a healthy way to see it. Time long-document summaries separately from emails: the food-manufacturer summary above saved about 25 minutes on its own, roughly what a whole day of chasing emails saved, so mixing the two hides which job is paying off.

Also count errors caught in review. If the check finds nothing for a month, people have stopped looking; seed a deliberate mistake into a test summary and see if it is spotted.

What to write down for the regulator and your insurer

You do not need a thick policy. A single page is enough: which tool and plan, what may be pasted, what may not, who reviews, and where reviewed drafts are saved. Check your regulator's current expectations and your professional indemnity insurer's questions on AI use before you start, and take advice from a qualified adviser where the answer is unclear. For client-facing wording, the tutorial on telling customers you use AI gives phrasing you can adapt.

The rule of thumb is simple: let ChatGPT shorten the writing, never the reading. Everything it produces about cover goes back to the document it came from.

Questions brokers ask before using ChatGPT at work

Can ChatGPT tell a client whether a claim is covered?

No. Whether a loss is covered depends on the exact wording, endorsements, exclusions and the facts of the loss, and a language model can state a wrong answer confidently. Use it to summarise what the wording says, quote the clause, and leave the cover opinion to a named, qualified person at your firm.

Is anonymising enough to make pasting safe?

It removes the most obvious risk, but not all of it. A rare risk description, a unique address fragment or an unusual claim history can identify a client even without a name. Check what your client agreements and regulator expect, and treat anonymising as one layer alongside a business plan and a written rule.

Which ChatGPT plan should a small brokerage use?

A business plan is the sensible starting point because business content is not used for model training by default, and you can set one workspace policy. Individual plans have a training switch in the data controls that each person must remember to check. Confirm current terms on OpenAI's pages before you roll it out.

Do I have to tell clients that I use AI?

Your regulator, your insurer partners and your client terms decide this, so check them. A sensible default is to say that drafts may be prepared with software and that a named person reviews everything sent. The site's disclosure tutorial covers wording you can adapt.

Further reads

• How Brokers Use AI to Compare Policy Wordings at Renewal — Renewal comparisons of wordings, a heavier job than summarising one policy.
• What Does AI Cost a Small Insurance Brokerage in Year One? — What AI costs a small brokerage across its first year.
• Can AI Answer Calls and Qualify Leads for an Insurance Agency? — Whether AI can answer calls and qualify leads.
• Does Your Business Insurance Cover AI Mistakes? — Whether your own insurance covers AI errors.
• AI Submission Intake: Stop Re-Keying Data Into Insurer Portals — Cut re-keying between client forms and insurer portals.
• Questions to Ask Before Buying AI That Touches Client Data — Questions to ask any AI vendor touching client data.
• AI Insurance Quote Comparison for Brokers: Carrier Quotes to Proposal — A broker's workflow from three carrier quotes to a client proposal: normalise, flag gaps, draft the narrative, and check every cell before advising.
• Copilot for Insurance Agents: Outlook and Word Tasks for the Agency — Email triage, client letters from policy wording and meeting recaps: which Copilot tasks an insurance agency can hand over, and at what cost per seat.
• AI in General Insurance: What a Small Broker Can Use Today — Renewal reminders, quote summaries, claims chasing and client emails: what a small non-life broker can hand to AI today, and what stays with a person.
• 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 documentation on removed members and custom GPT migration, and its Data controls help article; the insurer's own policy documents are the source for every cover statement in your work.