Agency security
by cirlet

Can insurance agencies put client information into ChatGPT?

By · Cirlet. Published . Page updated .

OpenAI API and ChatGPT Work documentation checked October 27, 2026. The examples are authored for this guide; they are not customer cases or AI test results.

Sometimes. The permission needs to cover the task, the information, the account and every service receiving it.

A paid plan or a no-training setting doesn't settle the whole decision. If your agency hasn't approved that combination, keep the client material out. Some jobs can be done with a generic brief instead.

Your agency's contracts and legal obligations matter too. This guide helps with the operating decision; it doesn't establish legal permission for a particular agency.

Four decisions staff can understand

These are teaching examples, not permissions granted to a real agency. An approval should tell staff what they can do as well as when to stop.

Does the task need the client document?

Suppose a staff member needs to remind a commercial client to send an updated vehicle list and payroll information. The source thread also contains names, addresses, a policy reference, financial figures and a claim discussion.

The reminder needs the names of the missing items and a tone. It doesn't need the rest of the thread.

The brief could say:

Prepare a short, courteous renewal-information reminder requesting an updated vehicle list and payroll information. Leave placeholders for the recipient, any confirmed due date and the agency's approved submission instructions. Make no statement about coverage or whether renewal is guaranteed.

An appropriate draft would be:

Hi [name],

We're preparing for your renewal review and still need your updated vehicle list and payroll information. Please send them using [approved submission instructions] by [confirmed date].

If you have questions about what to provide, please get in touch.

Check the wording and add the details in the agency's approved system. If no due date has been agreed, remove that part. This example shows that the task can be described without client-specific information. It doesn't prove a model will follow every instruction.

For repetitive reminders, a reviewed template may be enough. And keep the conversation itself clean: a generic new message doesn't make a conversation containing client records generic. Check the context, attachments and connected services used for the task.

What does a privacy promise actually cover?

“We don't train on your data” is a meaningful commitment about one use of information. Ask the other questions too:

A provider's public description and your workspace configuration answer different questions. Keep both pieces of evidence with the decision.

Check the account and the actual workflow

A personal account, an organization-managed workspace, a temporary conversation and a vendor application can have different terms and controls. Check which one staff will actually use. A paid personal subscription doesn't by itself establish agency approval.

Before relying on a temporary conversation or a setting that limits model training, ask the provider what happens to the content during processing and afterward. Check the exact account, feature, retention rules, exceptions and any connected services. We haven't verified the current personal-account and Temporary Chat settings in this guide, so don't treat it as an account-by-account settings reference.

When a vendor uses the OpenAI API

OpenAI says data sent to its API isn't used to train or improve its models unless the customer opts in. Its API data-controls documentation separately describes abuse-monitoring logs that may contain prompts and responses and are retained by default for up to 30 days, subject to stated exceptions. Some features also persist application state. Modified Abuse Monitoring and Zero Data Retention require approval and have feature-specific limits.

That doesn't tell you what the vendor's own application stores. Ask which components receive the document and which keep a copy: the vendor, a document-processing service, the model provider, logs or another integration. An OpenAI policy can't answer for another company's logs.

A file on your computer can still be sent elsewhere

For ChatGPT Work's local mode, OpenAI says a source file can stay on the device while relevant excerpts, prompts, screenshots, browser content or tool results are sent to OpenAI. Its documentation says covered business data in Business, Enterprise and Education workspaces isn't used to train models by default. Those statements apply to the documented services and conditions, not every vendor application or workspace configuration. OpenAI: ChatGPT Work local security.

Ask what actually leaves the computer in your workflow. The desktop icon and the location of the original PDF don't answer that.

Follow the information

Write down the route in plain words:

Agency source → AI workspace or vendor application → model service → any connected service → final output.

Adjust the route to match the real setup. Include what goes out, what comes back and any retained copies. An approval to work from one selected passage doesn't automatically cover connecting an entire mailbox. Permission to read a document doesn't automatically cover sending its summary to another app.

If the workflow will read, send or change agency records, use the agency AI permissions guide alongside this one. Data handling and action authority need to agree.

When the job needs client information

Some work can't be reduced to a generic brief. Extracting a policy field or reconciling a document against a client record may require the actual information.

The agency principal or designated operations lead should own the business-use decision. The administrator or IT provider supplies evidence about settings and data handling. Record the job and permitted fields, exact account and workspace, recipients, retention arrangements and the person reviewing the output. Give unresolved contractual or legal questions an owner.

Removing a name isn't enough on its own. In the reminder example, the claim narrative and financial details are still client-specific and unnecessary. If you need a redacted copy, prepare and verify it through an approved process. Sending an unapproved original to a cloud tool so it can redact it sends the original first.

A harder example: extracting fields

Suppose a vendor proposes extracting the named insured, policy number and expiration date from a declarations page into an internal review sheet. Removing those values would defeat the task.

For this authored exercise, the agency requires that submitted client content not remain with outside processors after the task. That's a deliberately strict sample condition, not a claim about insurance law or a universal recommendation.

The fictional vendor replies:

We use the OpenAI API, so your information is not used to train models. We delete your uploaded PDF after processing.

That's not enough to approve the upload. Deleting the PDF doesn't establish what happens to extracted text, prompts, responses or copies in logs. OpenAI's API documentation treats monitoring logs and application storage separately, and it doesn't establish this vendor's configuration.

Ask the vendor to identify the original file, extracted text, model request, generated result and logs containing client content. For each, request the service, retention period, deletion mechanism and applicable configuration.

If a required component retains content after completion, the workflow doesn't meet the sample condition as written. Change the workflow, find a suitable arrangement or have the responsible person reconsider the condition. Staff can use the established manual process while that's resolved.

Try three sample cases before client files

You can test whether a workflow catches missing or conflicting information without starting with a client's policy. Use the sample cases below, then make more that look like the documents your agency actually handles. A short test can reveal a problem; it can't establish how often the tool will be right on real files.

Before the trial:

Give the tool this instruction with the samples:

For each case, prepare a review row for the policy expiration date. Include the exact date shown, the source ID and the text supporting it. If a date is missing or sources disagree, say so and leave the approved value blank. Do not guess, send a message or change a record.

  1. Case A: A1: Policy expiration date: 2027-06-30. Expected: 2027-06-30, with source A1 and the supporting text.
  2. Case B: B1: Policy number: SAMPLE-42. Expiration date: not shown. Expected: Missing expiration date; no date filled in.
  3. Case C: C1: Policy expiration date: 2027-06-30. and C2: Policy expiration date: 2027-07-31. Expected: Both dates and their source IDs, marked as conflicting; no approved date chosen.

These are authored samples and expected review outcomes, not results from a vendor or a model. Compare the actual output with each source. A citation still needs checking: the tool may point to a passage that doesn't support its answer. Keep the established manual process available if the trial fails or the vendor cannot meet the agency's information-handling conditions.

Record what happened

For each run, note the date, product and account, available model or version, relevant settings, instruction, case ID, result, reviewer and decision. Record a wrong date, omitted case, unsupported source, unexpected action or missed conflict as a failure for that case. Keep sensitive records out of the test log.

For example, “2 of 3 cases matched the expected review outcome; case C picked a date despite the conflict” is a useful test record. It is an illustration, not an actual result or an accuracy rate for the product. If a case fails, stop using the workflow for that task, correct or change the setup, and rerun the cases before considering it again. A passed practice run doesn't authorize client uploads. The agency's designated owner must separately approve the real information, account, recipients, permissions and review process.

If an incorrect result has already entered a record or communication, follow the agency's correction and incident procedures. Record who owns the correction and check where the result went; do not assume changing the draft corrected every downstream copy.

NIST's AI risk-management playbook suggests defining tests and acceptable performance limits, tracking errors and documenting human oversight. This small exercise applies those ideas to one document task; it is not NIST certification or a complete evaluation.

Decide what happens to copies before uploading

“We'll delete it afterward” needs a defined process. Find out where the conversation, files, outputs, saved information and connected-service copies can live. Then establish who removes each copy, when and under which retention rules.

For the specific ChatGPT Work workflow, OpenAI's cloud-security documentation says deleting a conversation doesn't delete a separately saved Library file and doesn't necessarily delete saved memory, while connected-system copies follow the source system's rules. Check the documentation and settings for your own product and workspace before relying on a deletion process.

If someone has already uploaded material outside the approved process, report it promptly through the agency's incident process. Record what was shared, when, which account and services were involved, and any known sharing or onward action. Keep the record in the approved incident system without copying the sensitive document again unnecessarily.

The responsible person can determine containment, deletion requests, evidence preservation and any required advice or notification. Deleting the visible chat alone isn't proof that the issue is resolved.

Give staff a rule for the next request

Here's sample language to adapt after reviewing your own arrangements:

Use the agency-approved AI workspace only for the tasks and information listed in its approval record. Start with a generic brief when the job can be done without client details. Don't add client documents, connect a service or expand the task beyond that approval. If something doesn't fit, pause and ask the named owner. Report an unintended upload promptly. Review every draft before it enters a client record or communication.

Staff need to know where the approval record is, who owns it and which process they can use today. Use the provider action plan to assign the open questions without putting client information into it.