If you paste client documents into a chatbot, you send your client's data to a company you have no agreement with. It feels like using a tool. In fact it is a disclosure to a third party, and it can breach your contract, your professional duty and data-protection law at once.
Why it happens
Say it's Monday. The partner on the file wants a summary of a 90-page client contract by Thursday. You paste it into ChatGPT, and a day later you send a polished analysis. Your partner asks how you did it so fast, and you don't answer.
The gain is real. AI turns hours of synthesis into minutes, it needs no technical skill, and nobody sees you do it.
The chat window also feels private, and the AI seems to "just process text." So the habit spreads quietly through law firms, hospitals, accounting practices and government agencies, mostly where no policy covers it. None of this makes it acceptable.
What you actually uploaded
The contract held the client's legal name, registered address and bank details. It named the counterparty and set out the deal terms and penalty clauses. It may also have shown the strategy behind a deal the client's board has not yet seen.
All of it went to cloud servers run by a US company, under terms most people never read. What happens next depends on the account:
- Free and consumer accounts usually let the provider train its models on conversations, including your contract.
- Enterprise accounts usually promise no training use. Most knowledge workers are not on them, and those terms do not meet every GDPR data-residency rule.
- Human reviewers at most major providers may read flagged or sampled conversations.
- Subprocessors may route the data through several countries, each with its own laws.
The four risks
Your confidentiality agreement. Most engagements bar sharing client data with third parties. Under the plain text of most agreements, an AI provider is a third party. A client who finds out may have a claim even if no harm followed.
Your professional duty. Lawyers, accountants, doctors and HR staff owe confidentiality independent of any contract: state bar rules for lawyers, HIPAA for much health data. Sharing client data with an AI provider without informed consent can lead to a misconduct complaint. What happened counts, not what you meant.
Data-protection law. When the people in the document are in Europe, the GDPR applies. You are the controller and the provider becomes your processor. You need a processing agreement and a legal basis for the transfer abroad, or you risk infringing Article 46. Article 83(5) sets the highest fines for that: fines of up to 20 million euros or 4% of worldwide annual turnover, whichever is higher.
Your client's trust. Clients rarely mind that you use AI. They assume their details were removed before any outside service saw them. They may learn through an access request or a breach notice that the names, figures and clauses were sent intact. Then the relationship may end. In professional services, lose the relationship and you lose the business.
What a privacy policy does not do
A provider may promise not to train on your data. That can be true for one account type, in one jurisdiction, under the terms in force that day. It still gives you no lawful basis under the GDPR, does not release you from your agreement and is no defence to a misconduct complaint. It also fails you if the policy changes or you misread your tier.
Remove the identities first
The AI does not need to know the counterparty is "Acme GmbH, Friedrichstraße 123, Berlin" to analyse the contract. It needs the structure, the wording and the logic of each clause. A penalty clause means the same with the names removed.
So anonymise the document before you paste it. Replace names with [P0ABCD], addresses with [A1WXYZ] and Social Security numbers with [S2KLMN]. The analysis is just as good, and you map the codes back afterwards on your own machine. promptShield does this on your own device.
Common questions
What happens to a client document pasted into a chatbot?
It leaves your control. The provider sets retention, subprocessors and human review, and deleting the chat does not undo the processing. The details are in what happens to the text you paste.
Does a paid or enterprise plan solve it?
Only in part. Those plans usually promise not to train on your content. They give you no lawful basis for the transfer and do not override your own confidentiality clause.
What should I do instead?
Remove the identities before the text leaves your machine, keep the mapping locally and restore it afterwards. The model rarely needs the names to do the task. As we argue in why undisclosed AI use cannot be protected, you can only protect what you can see.