ChatGPT for Insurance Agents: Ten Safe Uses and Three to Avoid

An insurance agent can safely use ChatGPT for client emails, plain-English explanations of policy wording, renewal letters, meeting notes, checklists and marketing drafts, as long as no personal data goes in and a licensed person reviews every output. Three uses to avoid are giving advice, producing quotes and pasting personal client data.

Rules come first and uses second, because in an agency the risk is not that ChatGPT is clever; it is that ten people use it in ten different ways with client files open on the next screen. A short list of approved uses and banned ones, set by the owner and applied consistently, does more than any tool choice.

Follow me on Instagram@sagnikteaches

The rules that sit behind every use

Four rules apply to all ten uses and all three bans. Put them on one page and have every staff member sign it.

Connect on LinkedInSagnik Bhattacharya

1. No personal data goes in. No names, addresses, dates of birth, policy numbers, claim details, health or financial information. Use "[client]", "[policy number]" and a generic description of the situation.
2. Policy wording goes in only stripped of identifiers. The insurer's generic wording is fine to paste; the schedule with the client's details is not.
3. A licensed person checks every output before it reaches a client, and the checker is the named sender.
4. One account, set up by the owner. A ChatGPT Business, Enterprise or Edu workspace keeps inputs and outputs out of OpenAI's model training by default, according to OpenAI, whereas personal Free, Go, Plus and Pro accounts are opted in to training until the user switches it off. A business workspace lets you control access and remove leavers. If you stay on individual plans, each user must find the training switch under Settings, Data controls and turn it off.

Also check your carrier agreements, your data-protection obligations (data-protection law such as the GDPR if you have clients in the EU) and your errors and omissions policy. This tutorial on what to ask your carrier lists the questions. Nothing here is legal or regulatory advice; your compliance adviser decides what applies to your agency.

Subscribe on YouTube@sagnikteaches

Ten uses an agency owner can safely approve

1. Client emails from your own bullet points

Write three bullets about what you want to say and ask ChatGPT for a friendly email. Say a landscaping company client asks whether a storm-damaged fence is likely to be covered. You write: "ask for photos, explain that cover depends on the policy's storm wording, I will check it and reply by Friday." The prompt: "Write a warm email under 100 words from these bullets. Do not say whether anything is covered." The last sentence stops the model drifting into advice. You add the client's name in your own mail program.

2. Plain-English explainers from generic policy wording

Paste a clause from the insurer's wording (no client details) and ask: "Explain this clause in plain English for a small-business owner in under 120 words. Then list three questions the client should ask me about it." A sample clause on "waiting periods" becomes "cover for the event starts only after a set number of days; during that time, claims are not paid". Check the output against the original wording, because models sometimes smooth an exclusion into a softer statement. Label the final text "A summary, not the policy; the policy wording applies."

3. Renewal letters from a template

Create a letter structure once with placeholders: reminder of renewal date, what to review, what you need from the client, how to reach you. Ask ChatGPT to improve the wording and shorten it. Fill the placeholders from your agency system, not in the chat. Then use the template for the whole book. A realistic improvement: the model turns "Your policy is due to renew" into a clear first line with the date, followed by a bulleted list of changes to tell you about (new vehicles, staff, premises, work activities).

4. Meeting and call notes, anonymised

After a client meeting, type your own notes using initials or "the client", paste them and ask for action items, owners and dates. "Client wants to add a second van; check with the insurer; revert in a week." The output is a tidy list for the file. If your meeting-recording tool already produces summaries, stay with it and apply the same rules.

5. Checklists and onboarding steps

Ask for a checklist of information a new commercial client should provide: business activities, turnover bands, staff numbers, premises, previous claims history (as a request, not a record). The model drafts it; you check it against what your insurers actually require. This is a document with no client data in it at all, which makes it ideal.

An illustrative draft for a landscaping company joining the agency might list: main activities (design, hard landscaping, tree work above a set height, or none); annual turnover band; number of employees and subcontractors; vehicles and their use; tools and plant kept on site or in vans; premises and storage; contracts that demand specific liability limits; and incidents or claims in the last five years. The owner's edit is usually to add the questions your insurers ask that the model could not know, such as whether any work is done at height, and to delete anything that strays into an assessment of the risk.

6. Rewriting your own website and marketing copy

Paste a paragraph from your site and ask for a clearer version. Check compliance wording yourself: no guarantees, no implied cover, nothing that suggests more than the policy offers. Any claim like "cheapest" or "best cover" needs evidence or deletion. Keep the final copy within your regulator's rules for financial promotions, which the model does not know.

Before: "We find every client the best cover at the lowest price, so you are always fully protected." After a rewrite you would approve (illustrative): "We compare policies from several insurers, explain the differences in plain English and help you choose the cover that fits your business." The model's first rewrite often keeps "fully protected" in some form; take it out, because no policy covers everything and the phrase invites a complaint after a declined claim.

7. Training scenarios for staff

Ask for five fictional customer situations for a new team member to practise: a client who thinks they are covered for something they are not, an upset caller after a claim decision, a customer who wants to cancel mid-term. No real data is involved. Staff write their reply, then compare with a model answer reviewed by a senior person. See this tutorial on AI role-play for objections.

8. Turning a long document into a briefing

For a 60-page market update or an insurer's product guide (public documents, not client files), ask for a one-page briefing for the team with the changes that affect the book. Verify the figures and dates against the source; the summary will be quick, and the checking should be proportionate to what you do with it.

A sharper prompt adds structure: "List every change to cover, exclusions, limits or eligibility in this guide, with the page number, then a one-paragraph summary for the team." Say the briefing reports that a product's limit for tools kept in vans has changed. Open the guide at the quoted page and read the clause before anyone repeats it to a client, because a summary that drops "per item" or "overnight" turns a correct fact into a wrong one.

9. Complaint and apology wording

When a client is unhappy, you often have a clear view of the facts but not the words. Describe the situation in general terms ("a renewal letter went out a week late"), and ask for an apology and next-step email that takes responsibility without admitting more than the facts. Your compliance process and any complaints rules take priority; a model draft is a starting point for the person handling it.

10. Internal procedures and handover notes

Ask ChatGPT to turn your rambling description of how you process a mid-term adjustment into a numbered procedure, then correct it. Procedures written this way get done because the barrier is low, and they make holidays and sickness less painful. See also the Outlook and Word version if your agency works in Microsoft 365.

An illustrative result for a mid-term vehicle change: 1. Get the new registration, the date of change and the main driver's details from the client, and record them in the agency system, not the chat. 2. Request the change from the insurer. 3. Check the adjustment premium and new documents against the request. 4. Send the client the documents with the standard covering note. 5. Diarise a check that any extra premium has been paid. The owner's correction is usually to step 1, where the model forgets that the details must come from the client in writing.

Three uses to ban outright

A. Giving advice

Do not ask ChatGPT what a client should buy, whether a limit is adequate or whether a loss is covered, and never pass its answer on. Advice depends on the client's full circumstances, the exact policy and the rules you work under, and you carry the responsibility. The model can sound authoritative while being wrong, and clients rely on a confident tone. Use it to explain wording and to prepare questions for you, and keep the recommendation human.

B. Producing quotes or premiums

A model has no access to insurers' rating, and any figure it gives is an invention. Even "ballpark" premiums are dangerous, because clients remember the number. Quotes come from the insurer or comparison system. If a client asks for a rough idea, say it depends on the insurer, and give a range from your own experience only if your compliance rules allow it.

C. Pasting personal data

The failure most likely on a busy morning. Staff are busy, the file is open, and pasting a whole email chain is faster than removing names. Health details, claims history, financial position and identity documents are the most sensitive. A mistake here may breach your data obligations and your insurers' terms. Prevent it with the one-page rule, an example of a good and bad paste on the wall, and a no-blame reporting route.

What staff say when a client asks the banned question

Bans hold better when staff have a sentence ready. When a client asks "Am I covered for this?", a scripted reply avoids both the model and a guess: "Good question. Cover depends on your policy's exact wording, so I'm going to check it and come back to you by [day]. Could you send me photos and a short note of what happened?" Pin it beside the phones. It buys time for a licensed person to read the wording, and it needs no AI at all.

A renewal letter template, built and checked end to end

This is illustrative. Say a four-person agency wants one renewal letter template for small commercial clients. The owner writes the bullets: renewal date, ask the client to tell us about changes, mention that we will review options with them, give a reply-by date, offer a call. The prompt:

Write a renewal letter template for small-business insurance clients. Placeholders in square brackets: [client first name], [renewal date], [reply-by date], [agent name]. Content: remind them cover renews on [renewal date]; ask them to tell us about any changes in the last year (staff numbers, premises, vehicles, equipment, new work activities, claims or incidents); explain we will review the renewal terms with them; ask for a reply by [reply-by date]; offer a call. Friendly, plain English, under 200 words. Do not say what is or is not covered. Do not mention prices.

A realistic output has a clear subject line ("Your cover renews on [renewal date]: a quick check first"), a short opening, a bulleted list of changes to tell us about and a closing offer of a call. What the owner fixes: the model adds "to make sure you are fully protected", which implies a promise, so it goes; it adds "we will find you the best price", also removed because no agent can promise that. The bulleted list is kept because it prompts clients to disclose changes, which is where renewals go wrong. The senior agent approves the template, saves it in the agency system, and every renewal letter starts from it, with details added outside ChatGPT.

Clause, call note and complaint: three before-and-afters

A coverage explainer, before and after

Before (generic wording): "The insurer shall not be liable for loss or damage arising from gradual deterioration, wear and tear, or failure to maintain." After, from the prompt in use 2 (illustrative): "The insurer will not pay for damage that builds up slowly over time, such as rust, rot or ageing parts, or for damage caused by not looking after the property. Sudden, accidental damage is treated differently." The question list that follows reads: "How do I show I maintain the equipment? What counts as sudden? Does the policy define wear and tear?" The agent reads the last sentence against the clause, notices that "sudden" is not defined in the wording supplied, and removes the sentence rather than let a client rely on it.

A call note, before and after

Before: "spoke to the client about van, thinks might get second one in spring, asked about cover for tools, wants to know about public liability limit, call back Tues." After the prompt in use 4: actions list, "1. Check with insurer whether a second vehicle can be added mid-term (agent, by Friday). 2. Review tools-in-transit cover wording (agent). 3. Confirm current public liability limit and what the client's contracts require (agent, client to send one contract). 4. Call back Tuesday." The structure is useful. The agent still checks that no action was invented, such as a promise the client never received.

A complaint reply, with the limits visible

Prompt: "A client's renewal letter reached them eight days after we planned. They were annoyed. Write a short, sincere apology email that states what happened in one sentence, what we have changed (we now send renewal letters 30 days in advance), and offers a call. Do not mention compensation or fault beyond that." The draft is useful for tone. The agent must confirm "we now send 30 days in advance" is true before keeping it; an invented process change in an apology is a promise.

One shared agency Project with the bans built in

Once the owner has approved uses, put the prompts in one place. A ChatGPT Project (ChatGPT's shared workspace for instructions and files) lets everyone open the same setup, and sharing a project is available on every ChatGPT plan, including Free, per OpenAI's help article on projects. Put four items in the instructions: the four rules, the three bans, the tone for client emails and the line "If the user pastes personal data, stop and ask them to remove it." The last line will not catch everything, but it reminds the user at the moment it matters. Store the template letter, the checklist and the explainer prompt as files. Review the Project each quarter, and remove access promptly when a person leaves.

If a client's file gets pasted by mistake

Write the response plan before you need it. An illustrative case: a staff member pastes an email chain including a client's name, address and claim description to get a summary. What happens next: they tell the nominated person that day, delete the conversation, and record what was pasted and when. The owner decides whether any client or regulator notification is needed with the help of the compliance adviser, because duties depend on your rules and the data involved. Then the team fixes the cause. If the cause was a rushed morning, add a rule: summaries come from your own notes, never from a pasted chain.

Deleting the chat is not instant erasure. OpenAI's help centre says deleted chats are scheduled for permanent deletion within 30 days, unless OpenAI must keep them longer for security or legal reasons, and a business workspace's own retention and compliance settings can also apply. So do not tell a client the data is gone until that period has passed and your adviser agrees.

An illustration: one week in a four-person agency

Say a four-person agency selling commercial and household cover approves uses 1, 3, 4 and 5 for the first month. Monday: the owner builds a renewal template and a new-client checklist (uses 3 and 5), reviewed by the licensed senior agent. Tuesday and Wednesday: staff use the email prompt (use 1) with bullet points for twelve client messages. Thursday: the team meets; one person pastes anonymised notes for action items (use 4). Friday: the owner samples five sent emails and checks each for personal data and for any sentence that sounds like advice.

In the sample the owner finds two issues. One email says "you should be fine" about a claim: advice, removed. Another includes a full policy number the sender had pasted into the prompt. The owner reminds the team of the rule, with no blame, and adds a "scan for numbers" step to the checklist. Four weeks later, the owner adds uses 2 and 7.

The allowed-and-banned table for the office wall

A monthly sample that shows whether staff follow the bans

Do three things monthly. Sample five sent messages and ask the sender to show the bullets they started from, which reveals what went in. Look for forbidden phrases in the sample, such as "you are covered", "this will be covered", "should be fine", or any premium figure. And ask staff one question: has anyone been unsure whether to paste something? Their uncertainty tells you which rule needs an example. The goal is a team that asks, rather than one that guesses.

A quick sum shows why the sample should rotate. An agency sending about 60 client messages a week sends roughly 260 a month, so a five-message sample covers about 2 per cent of them. Rotate whose messages you pick so every sender is sampled each quarter, and look hardest at the uses added most recently.

If you use a wider set of tools, this comparison of seven AI tools for agents shows where specialist software does the job better than a chat assistant, particularly for calls and quotes, and the life-cover version covers suitability rules for follow-up. Start small, put the rules in writing, and let the licensed people stay responsible for every word that goes to a client.

Questions agency owners ask about ChatGPT

Can I paste a policy document into ChatGPT?

Paste the policy wording only after removing the policyholder's name, address, policy number and any claims details. The wording itself is generic text from the insurer; the identifying details are the risk. Check your carrier agreement and your own data policy first.

Does ChatGPT count as giving advice?

If it outputs a recommendation and you pass it to a client, the advice is yours. That is why the second avoid rule exists. Use it to explain wording and structure messages; the licensed agent makes and owns every recommendation.

Is my agency's errors and omissions cover affected?

It may be. Policies differ on AI use and on who is responsible for mistakes. Ask your carrier or broker in writing whether AI-assisted drafting changes anything, and keep the reply on file.

What should staff do if they pasted something they should not have?

Report it the same day to the person you named in your policy, delete the chat, and record what was pasted. A no-blame reporting rule gets problems reported early, which is what you want.

Further reads

• AI Answering Service for Insurance Agents: Claims Calls and Quote Leads — Phone answering for claims calls and quote leads.
• Software for Insurance Agents and Brokers: Where AI Fits the Stack — Where AI fits the wider agency software stack.
• HubSpot for Insurance Agents: Pipeline, Renewals and What Breeze Adds — Pipeline and renewals in HubSpot.
• Zapier for Insurance Agents: Five Intake and Certificate Automations — Five intake automations for an agency.
• Can AI Answer Calls and Qualify Leads for an Insurance Agency? — Can AI answer calls and qualify leads for an agency?
• ChatGPT Prompts for Financial Advisers: Follow-Ups and Review Prep — Admin and communication prompts for advisers, written around confidentiality: placeholders, no client data, and a compliance check before sending.
• ChatGPT for Commercial Real Estate: Source-Checked Reports and Comps — A commercial broker's method for ChatGPT market reports, comps summaries and owner emails, with a check on every figure before it reaches a client.
• 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, data controls and deleting chats (checked 4 November 2026).