Anonymise documents before using ChatGPT or Claude

Staff at many firms use ChatGPT and Claude, often without the firm's approval and often with client data. In our view, a ban does not solve the problem; the protection has to take effect on the device, before any text leaves it. This page describes the legal issue, why manual redaction falls short, how pseudonymisation works and how Veil applies it locally on the user's computer.

Shadow AI is the normal case, not the exception

Writing on finews.ch, Peter Alder of Marcuard Heritage describes shadow AI as «nicht das Fehlverhalten Einzelner, sondern der Normalfall, mit dem jede Organisation rechnen muss»: not the misconduct of individuals, but the normal case every organisation has to expect. A ban tends to move use to private devices rather than end it.

For portfolio managers, more than data protection is at stake. Pasting client data into a public tool discloses it to a third party that is neither engaged as an agent nor bound to secrecy by contract. That touches professional secrecy under Art. 69 FinIA, which we cover in our article on professional secrecy and AI vendors.

Manual redaction misses what identifies a person

Manual redaction works for a single page but is unreliable at scale. An IBAN in a footnote, an address in a signature block or a client name in a file path is easily missed. So are indirect identifiers, such as a combination of profession, town and transaction size that points to one person. Redacted text is also hard to work with, because the model loses track of who did what.

Pseudonymisation keeps the meaning, not the identity

Pseudonymisation replaces each identifier with a consistent placeholder before the text leaves the device. «Anna Muster» becomes «Person 1» throughout the document, her bank becomes «Bank 1» and her account becomes «Account 1». The model can still work with relationships, dates and amounts. When the answer comes back, the placeholders are replaced with the real values.

Pseudonymised data remains personal data for whoever holds the key. The protection depends on where the key is kept. If the mapping never leaves your device, the AI provider receives only placeholders.

Veil works in four steps on your device

1. Detect. A language model running on your computer finds names, companies, addresses, account numbers and other identifiers, and replaces them with placeholders.
2. Review. A review panel shows every replacement before anything is shared. You add or correct entries.
3. Work. You use the AI tool of your choice with the pseudonymised text.
4. Restore. When you paste the answer back, Veil restores the original values. The mapping stays in a local store on your device and is sent neither to Unplex nor to the AI provider.

Pseudonymisation does not replace a policy

Pseudonymisation reduces what leaves the device. It does not determine which tools staff may use, which documents are suitable or how results are checked. Context can still identify a person if a document describes a unique situation in sufficient detail, which is why the review step matters.

In our view, an internal policy on AI tools should define at least which tools are approved, which categories of documents may be processed with them, whether pseudonymisation is mandatory before use, and who checks results before they are used with clients. Veil supports the third point. The others remain decisions of the firm, and the AI tools in use belong in the AI inventory.

Request a demo