How to use ChatGPT or Claude with client data without exposing it
Generative AI saves hours on contracts, memos and spreadsheets. The problem is what travels with the file: your clients' names, identifiers and figures. Here is how to keep the productivity and lose the exposure.
- Anything you paste or upload into a chatbot leaves your organisation and is processed on the provider's servers, under that provider's terms and your plan's settings.
- Banning AI rarely works: people switch to personal accounts on their phones, with no trace and no protection.
- Business plans improve the contract, not the content: the provider still receives your clients' names.
- The robust option is to remove identities before the file leaves the computer, with consistent codes the AI can still reason on, and to put the real values back locally afterwards.
What actually happens when you paste a client file into an AI
When someone drops a contract into ChatGPT, Claude, Gemini or Copilot, the full text is sent to the provider's servers to be processed. That includes every name, company number, IBAN, salary and address in it. What happens next depends on the provider, the product and the plan: how long conversations are kept, whether they can be reviewed for abuse monitoring, whether they can be used to improve models, and in which country they are processed.
None of that is necessarily malicious. But for an organisation that owes confidentiality to its clients or employees, it raises three concrete problems:
- Loss of control. Once the file has left, you rely on someone else's settings, contracts and security for data you are accountable for.
- Legal exposure. Under the GDPR, sending personal data to an AI provider is processing that needs a lawful basis, a processing agreement and, outside the EEA, a transfer safeguard. Lawyers, accountants, doctors and many others are also bound by professional secrecy that a contract with a vendor may not satisfy.
- No proof. If a client or a regulator asks what the AI received, most teams cannot answer.
The four ways organisations respond
1. Ban AI
It is the reflex, and it rarely holds. The productivity gain is too large: people keep using AI, on personal accounts, on their phones, outside any policy. A ban moves the risk out of sight. Many organisations now call this "shadow AI".
2. Buy a business or enterprise plan
Business plans from AI providers come with better terms: no training on your data by default, administration, shorter retention, sometimes data residency. That is a real improvement, and often a prerequisite. But it improves the contract, not the content: the provider still receives your clients' names and figures, and you are still trusting what happens on servers you cannot inspect.
3. Redact by hand
Some teams black out names before uploading. It works for one short document. It does not scale to a case file of forty documents, it is easy to miss one name, and it breaks the AI's reasoning: if every person becomes "[REDACTED]", the AI can no longer tell who did what. We cover the pitfalls in how to redact a Word or Excel file before an AI reads it.
4. Pseudonymise locally, then decode locally
The fourth approach removes identities before the file leaves the computer, but replaces them with consistent codes instead of black boxes. "Sarah Whitfield" becomes NOM-001 in every document of the case; "Atlas Logistics Ltd" becomes SOC-001. The AI can still reason, compare and draft, because the structure is intact. When the result comes back, the codes are replaced by the real values on your machine. The key linking codes to identities never leaves.
This is what lawyers and data-protection specialists call pseudonymisation, and it is the approach Hidden Tokens automates.
A checklist before your team uses AI on client files
- Map the flows. Which teams use which AI, for which documents? Ask; the answer is usually "more than you think".
- Choose the plan deliberately. Use a business plan with a processing agreement rather than personal accounts, and read its data settings.
- Decide what may leave the office. For most client work, the rule that is easiest to defend is: identities never leave, only codes do.
- Code whole case files, not single prompts. The same client must get the same code across every document, or the AI's output will be inconsistent.
- Cover the browser too. Files are only half the flow; messages typed into ChatGPT or claude.ai need the same protection.
- Watch for indirect identifiers. "The CEO of a Lyon biotech, born in 1971" identifies someone without a name.
- Keep proof. Be able to show a client, a DPO or an auditor what the AI actually received.
- Write it down. A one-page AI policy that people can follow beats a long one they can't.
What good looks like
In a well-run setup, a lawyer drops the case folder into a watched folder; every document is coded on the laptop within seconds; they work in Claude or ChatGPT on the coded copies; the memo the AI drafts comes back with the real names restored, in Word, on the same laptop. Nothing identifying left the office, and a local log shows exactly what the AI saw.
That is the workflow Hidden Tokens was built for. Whatever tool you choose, judge it on the same three questions: does the key stay with you, do the codes stay consistent across the whole case, and can you prove what the AI received?
Frequently asked questions
Is it legal to put client data into ChatGPT or Claude?
It depends on your legal basis, your contracts, your sector's secrecy rules and the provider's terms. Under the GDPR you need a lawful basis, a data processing agreement with the provider and, for transfers outside the EEA, an appropriate safeguard. Regulated professions such as lawyers and doctors are bound by secrecy obligations that go further. Removing identities before sending is the simplest way to reduce what is at stake.
Does ChatGPT or Claude train on what I send?
It depends on the product and the plan, and providers update these policies. Consumer plans and business plans generally have different defaults. Check the current terms and the data settings of the exact plan your team uses, and remember that not training on your data is not the same as not receiving it.
Is an enterprise plan enough?
An enterprise plan gives you better contractual commitments, administration and retention controls. It does not change the fact that your clients' names reach a third party. For many professional-secrecy situations, the safer standard is that identities never leave your office at all.
What about text typed directly into the chat?
File-level tools only protect files. To cover messages typed or pasted into ChatGPT, claude.ai, Gemini or Copilot, the protection has to sit in the browser, coding each message before it is sent. Hidden Tokens does this with its browser extension.