Someone pasted client data into a public chatbot: what to do that day, and how to make it rare
A staff member put a client file into a consumer AI chatbot. How to assess what happened using the provider's own data rules, who to involve, and the changes that make the next one less likely.
Monday morning
A project coordinator at a small engineering firm tells her manager, a little embarrassed, that on Friday she pasted a client’s site report into a free AI chatbot to turn it into a summary. The report included the client contact’s name, phone number, the building’s access arrangements and a list of security deficiencies. She used her personal account because the firm has not set up an AI tool. She wants to know if she is in trouble. Her manager wants to know if the client is.
This is one of the most likely AI incidents a small organization will face, and it is a hard one. Nobody attacked anything. The data left on purpose, through a tool that worked as designed, by someone trying to do her job faster. Treating it as a disciplinary matter guarantees that the next person will not tell you. Treating it as nothing ignores a real disclosure.
It is not hypothetical at any scale. In 2023, Samsung restricted generative AI tools on company devices after sensitive internal data was accidentally leaked to ChatGPT, according to reporting in TechCrunch based on a Samsung memo first reported by Bloomberg. Samsung told staff not to submit company information or personal data to such tools.
Why it is hard
- You cannot pull it back. You can delete the conversation from the account, but whether copies remain, and for how long, depends on the provider’s rules, not yours.
- The rules differ by provider and by account type. Consumer and business versions of the same chatbot usually handle data differently, and the settings the user had turned on matter.
- The person who reported it is your best control. How you respond decides whether the next mistake is reported.
The same day: five steps
1. Thank the person, then get the facts
Ask, without blame: which tool, which account (personal, free, paid, or a company plan), what exactly was pasted or uploaded, when, and whether the chatbot’s settings for history or model training were changed. Get a copy of the text that was shared if she still has it.
2. Read the provider’s own data rules for that account type
This is the step that tells you how serious it is. The terms differ, and they change, so read the current page for the specific product. Two examples show how much they vary:
- Google Gemini Apps (consumer). Google’s Gemini Apps Privacy Hub says a subset of chats is reviewed by human reviewers, including trained service providers, after being disconnected from the user’s account. Chats with Keep Activity off, and temporary chats, are kept for 72 hours and are not used to train models unless the user sends feedback. Reviewed chats are retained for up to three years, and deleting activity does not remove chats that were already reviewed. Google’s own advice on that page is not to enter confidential information you would not want a reviewer to see. Google notes that work or school accounts may have different terms.
- Anthropic Claude (consumer). Anthropic announced in August 2025 that Claude Free, Pro and Max users choose whether their chats are used to train future models. If the user allows it, retention extends to five years; if not, the existing 30-day retention applies. Deleted chats are excluded from future training. Anthropic said its commercial services, including Claude for Work and API use, are not affected by that change.
Check the same points for whichever tool was used: is content reviewed by people, is it used for training under the user’s settings, how long is it kept, and does deleting the chat remove it.
3. Have the user delete the conversation, and record that it was done
Deletion may not remove every copy, as the Gemini example shows, but it limits further exposure and is usually the only lever available. Note the time it was deleted.
4. Classify the data and decide who else needs to know
What was shared decides the next steps:
| What was shared | Usually means |
|---|---|
| Internal notes with no personal or client-confidential data | Record it, use it as a training example, close it |
| Client-confidential business information | Review the client contract’s confidentiality and notice terms; involve the account owner |
| Personal information about individuals | Ask counsel whether a breach-notification or privacy-law obligation applies |
| Regulated data (health, financial account, government ID) | Treat as a potential reportable incident and involve counsel the same day |
Whether a disclosure to an AI provider triggers a legal notice is a legal question that depends on the data, the law and the provider’s terms. Do not guess, and do not wait until the end of the week to ask.
5. Write it down
A short incident record: what was shared, through what, when, what the provider’s terms say, what was done, who decided, and what the client was told. If a client asks later, this is your answer.
Making it rare
The coordinator used a personal chatbot because the firm had not given her a better option. Most of the fix is there.
Give people an approved tool
An approved AI assistant on a business plan, with data terms you have read, removes the main reason people reach for a personal account. Choose it the way you would any vendor that will hold client data: check the terms on training, retention, human review and admin controls.
Write a rule people can remember
Long AI policies are not remembered. A rule that fits on a card is:
- Use the firm’s approved AI tool for work.
- Do not put client names, personal information or anything marked confidential into any other AI tool.
- If you make a mistake, tell your manager the same day. Reporting a mistake is never the problem.
Add a technical control where you can
Endpoint data loss prevention, managed browser controls or a secure web gateway can warn or block when sensitive data is pasted into public chatbots on company devices. Microsoft, for example, documents endpoint DLP policies in Purview that warn or block users from pasting sensitive information such as credit card numbers into third-party generative AI sites on onboarded Windows computers. Start in warn mode so people learn the rule at the moment it applies, and remember that no control on company devices covers a personal phone.
Rehearse it
Use this exact scenario in your next tabletop exercise. Ask who would be told, who reads the provider terms, and who calls the client. The answers will show the gaps faster than any policy review.
Who does what
| Role | Responsibility |
|---|---|
| Staff member | Reports promptly, deletes the conversation, provides details |
| Manager | Responds without blame, starts the record |
| Security or IT lead | Reads the provider terms, classifies the data, recommends next steps |
| Counsel | Advises on notification and client contract terms |
| Owner | Decides on client communication, approves the AI tool and rule |
When to get help
If your firm has no approved AI tool and no rule yet, setting both up takes less time than handling the incident that comes from not having them.
Want to talk it through? Book a free 30-minute call with Security Medic Consulting.
Sources
- TechCrunch, Samsung bans use of generative AI tools like ChatGPT after April internal data leak, May 2, 2023 (accessed 2026-11-05).
- Google, Gemini Apps Privacy Hub, privacy questions section updated October 26, 2026 (accessed 2026-11-05).
- Anthropic, Updates to Consumer Terms and Privacy Policy, August 28, 2025 (accessed 2026-11-05).
- Microsoft Learn, Microsoft Purview data security and compliance protections for Microsoft 365 Copilot and other generative AI apps, May 27, 2026 (accessed 2026-11-05).