Is ChatGPT Safe for Insurance Client Data?

By the end of this you will know which AI plans keep your clients' data out of model training, which ones will sign a BAA, and what never goes into a chat box at all.

A CSR has a loss run open. Six pages, four claims, the insured's name on every line. She copies the whole thing, pastes it into ChatGPT and types summarize this for the renewal meeting.

Ping. A tidy summary comes back.

So where did those six pages go?

That depends almost entirely on which plan she was signed into. And at most agencies, nobody knows. The 2026 Big "I" Agents Council for Technology report found 45% of respondents already use ChatGPT and other public models, while data privacy and compliance risk topped the list of concerns. Liberty Mutual's 2026 study found only 22% of agency principals and staff trust AI with client and business data.

They are right to be careful. Here is how to be careful and still get the hours back.

Short answer: the plan decides

Business plans from OpenAI, Anthropic and Microsoft do not train on your data by default. Personal plans can, unless someone flips a switch. Neither kind makes it fine to paste a Social Security number.

So skip "is ChatGPT safe?" and ask a sharper question: which ChatGPT is my team using, and who set it up?

Plan by plan: training and BAAs

This is what each vendor says today. Terms move, so check the linked page before you rely on it. Last checked November 2026.

Product and plan	Trains on your chats by default?	Will sign a HIPAA BAA?
ChatGPT Free, Plus, Pro (personal)	Yes, until you turn off Improve the model for everyone	No
ChatGPT Business (formerly Team)	No	No
ChatGPT Enterprise, sales-managed	No	Yes, through OpenAI sales
OpenAI API	No	Yes, on request
Claude Free, Pro, Max (personal)	Your choice, set in privacy settings	No
Claude Team	No	No
Claude Enterprise	No	Yes, once HIPAA settings are turned on
Anthropic API	No	Yes, after a signed BAA and enablement
Microsoft 365 Copilot and Copilot Chat, work account	No	Covered when properly configured, web searches excluded

Where those rows come from:

• OpenAI says it does not use ChatGPT Business, Enterprise or API data for training by default, while personal Free, Plus and Pro accounts have sharing on until the user opts out under Data Controls. OpenAI's BAA article explains how to request one, and Medcurity's 2026 breakdown spells out that only the API and sales-managed Enterprise or Edu accounts are eligible, not Free, Plus, Pro or self-serve Business.
• Anthropic's 2025 consumer terms update gave Free, Pro and Max users the training choice and states it does not apply to Claude for Work or the API. Its BAA page says Enterprise is covered only after the Primary Owner activates HIPAA compliance and accepts the BAA.
• Microsoft says prompts and responses in Copilot with enterprise data protection are not used to train foundation models, and that web search queries fall outside its DPA and BAA.

Read that last column twice if you place group health or benefits. A business plan is not automatically a HIPAA plan.

Training is the smaller risk

Model training gets the headlines. The everyday leaks are duller:

• Personal accounts. A producer uses his own Plus account at 10 p.m. because the agency account is on his work laptop.
• Shared links. Someone shares a chat with a client's dec page in it, and the link travels.
• Retention. Chats you never delete sit in a history that anyone with the password can scroll.
• Extensions and connectors. A browser add-on that "reads the page for you" reads everything on the page.

Turning off training fixes none of these. Your setup and your rules do.

What never goes in a chat box

Unless your plan, your contract and your state rules all say otherwise, keep these out of every AI tool:

1. Social Security numbers and tax IDs tied to a person
2. Driver's license numbers and MVR detail
3. Bank account, routing and card numbers
4. Dates of birth next to names
5. Health details: medical history, injury descriptions, workers' comp claim notes
6. Passwords and portal credentials of any kind
7. Anything a client told you in confidence that is not needed for the task

Notice what is not on the list: coverage forms, blank applications, your own email drafts, a carrier's underwriting guide. Most of the hours AI saves come from documents that never needed a name on them.

Which rules reach your agency

Three sets of rules usually apply. This is not legal advice, so read the sources and ask counsel about your states.

GLBA, through your state. The Gramm-Leach-Bliley Act covers the nonpublic personal information of insurance customers, but for insurance its rulemaking sits with state insurance regulators. Your state's privacy and safeguarding rules are the ones you answer to.

The NAIC Insurance Data Security Model Law (#668). States including Alabama, Connecticut, Delaware, Indiana, Iowa, Louisiana, Maine, Maryland and Michigan have adopted it, per a 2025 state roundup, and the NAIC keeps the current list on its adoption map. It asks licensees for a written information security program, oversight of third-party service providers, and notice after a cybersecurity event. The model exempts licensees with fewer than ten employees, including independent contractors, from the program requirement, though some states changed that threshold. An AI vendor that holds client data is a third-party service provider. Treat it like one.

HIPAA, if you touch health plans. Most property and casualty agencies are not HIPAA covered entities. Benefits brokers who handle group health plan data often sign business associate agreements of their own. If you did, protected health information only goes into a tool whose vendor signed a BAA with you. See AI for health and employee benefits brokers.

Redact before you paste

The habit that solves most of this takes thirty seconds. Strip the identifiers before the text leaves your machine, then tell the model to keep your placeholders intact.

1. Paste the text into a blank document first, not the chat.
2. Find and replace names, addresses, policy numbers, dates of birth and license numbers with placeholders.
3. Read it once more. Signatures and email footers hide names.
4. Paste the cleaned text with a prompt like this one.

You are helping an insurance agency with a redacted document.
Placeholders: [INSURED], [ADDRESS], [POLICY NUMBER], [DOB], [DL NUMBER], [CLAIMANT].
Keep every placeholder exactly as written. Do not guess what they stand for.
If you need a fact that was redacted, say which placeholder you need. Do not invent it.

Task: [WHAT YOU NEED, FOR EXAMPLE: list each claim with date, type, paid and reserved]

Text:
[PASTE THE REDACTED TEXT]

Then put the real names back in your own system, not in the chat.

Set up an agency account in five steps

1. Pick one business plan and pay for it: ChatGPT Business or Enterprise, Claude Team or Enterprise, or Copilot on your Microsoft 365 work accounts. One is enough. ChatGPT vs Claude for insurance helps you choose.
2. Ban personal accounts for agency work. Say it out loud and in writing.
3. Check the admin settings for retention, sharing links and connectors. Turn off what you will not use.
4. Write it into your AI policy, with the list above. Here is how to write an AI policy for your agency, template included.
5. Ask every AI vendor the same questions before anyone signs.

Questions to ask any AI vendor

• Do you train on our data, and where does that promise live in the contract?
• How long do you keep prompts, files and outputs, and can we shorten it?
• Will you sign a BAA, and for which features exactly?
• Who at your company can read our data, and when?
• How will you notify us after a security incident, and how fast?
• Can we export and delete everything when we leave?

A vendor that answers these on one page is a vendor that has been asked before.

There is another way to keep client data out of chat windows: stop moving it by hand. With Salience, you forward the request, Salience prepares the certificate, application or renewal, and nothing goes out until someone at your agency approves it.