Someone pasted client data into ChatGPT. What now?

Someone has just told you that a client's information went into ChatGPT: a contract to sum up, a spreadsheet to tidy, an email thread to reply to. Most likely it was on a personal account, and they probably did it to get work done rather than out of carelessness. What happens in the next few hours decides whether this becomes a short entry in an incident log or a conversation with a client you would rather not have.

First, thank the person who told you

They have done the most useful thing anyone in this situation can do, and how you respond decides whether the next person tells you or deletes the chat and says nothing. Most people started using AI before there were any rules. Treat it as an event to manage, not misconduct to punish, and say so to them in those words.

In the first hour: get the facts

Sit with the person and establish, in writing:

1. Exactly what went in. Ask them to show you the conversation rather than describe it.
2. Which tool and which account: a personal plan, a company workspace or something else, such as a browser extension.
3. Whether the setting that lets their conversations improve the model was on, if it was a personal plan.
4. When it happened, and whether it happened more than once.
5. Whether they created a shared link to the conversation, or whether the tool saved anything to its memory.
6. Whether any output from it has gone to the client or anywhere else.

Contain it

Then, with the person at their own account:

• Delete the conversation. OpenAI says deleted chats are scheduled for permanent removal from its systems within 30 days, unless it has to keep them longer for security or legal reasons.
• Delete any shared link to it. Anyone who opened the link and continued the conversation keeps their own copy, which deleting the link does not remove.
• Check the saved memories. They are stored separately from chat history, so deleting the chat does not delete anything the tool remembered from it.
• If the training setting was on, turn it off for future conversations.
• If it happened in a company workspace, tell whoever administers it, since they may have records and retention settings the individual cannot see.

Work out what it was

What went in decides what follows. Sort it into one or more of these:

• Personal data about identifiable people, and whether any of it is sensitive, such as health or financial information.
• Information covered by a confidentiality clause or a non-disclosure agreement with the client.
• Legally privileged material.
• Information the client has told you not to share with third parties.

For the checks that should come before personal data goes anywhere near a tool, see AI and data protection.

Decide with your advisers what you owe

These are decisions for leadership to make with your advisers within hours, and nothing here is legal advice.

• Data protection. Under UK and EU law, a personal data breach has to be reported to the regulator within 72 hours of the business becoming aware of it, unless it is unlikely to result in a risk to the people concerned. The clock runs through weekends, a breach that is not reported still has to be recorded with the reasoning, and people at high risk have to be told. Rules differ elsewhere, including from one US state to the next, so take advice on where your client's people are.
• The client contract. Many contracts require notice of any disclosure the client has not approved, sometimes within a set time.
• Any professional regulator you answer to. Law firms, for example, now have a specific warning notice to consider.
• Your insurer. Cyber and professional indemnity policies often require incidents to be notified, and a late notice can affect cover.

Whether and when to tell the client is a leadership decision, guided by the contract and your advice. Whatever is decided, the facts you gathered in the first hour are what make that conversation accurate.

Record it

Write down what happened, when you found out, what was contained and when, what you assessed, and what you decided and why. If you hold ISO 27001, this is the incident record its controls expect, as set out in what ISO 27001 expects once AI is in use. If you do not, it is still the document you will want if the client, the regulator or your insurer asks.

Fix the cause

Ask why a personal ChatGPT account was the easiest way to get that piece of work done. Usually it is because there was no approved tool, no written rule or both. To stop the next one, you need an approved tool on a company account, a short acceptable use page people have read, and a route for reporting that this incident has just shown works.

Where this fits

Because one incident is rarely the only one, the readiness snapshot shows which other accounts are in use and what they can reach. From there, Govern puts the rules, the approved tools and the company sign-in in place, so the next person has a better option than a personal account.