AI for Tax Advisors: Protecting Client Data Under GDPR

Yes, tax advisors may use AI tools. But never with real client data in plain text. Typing names, tax IDs, or amounts straight into ChatGPT, Copilot, or Gemini risks two violations at once: a breach of the statutory duty of confidentiality (Section 203 of the German Criminal Code) and a breach of the GDPR. The fix is simple. You replace sensitive details with placeholders (pseudonyms) before you send anything, then work with the AI.

Why AI is especially sensitive for tax advisors

Tax advisors are bound by a statutory duty of secrecy. They may not disclose secrets entrusted to them. When you enter client data into an AI tool, that data leaves your firm and lands on third-party servers. That alone can break confidentiality and breach the GDPR at the same time. Two sets of rules apply here:

• Criminal law: Section 203 of the German Criminal Code (StGB) penalizes the disclosure of private secrets. Tax advisors are named explicitly. A breach can be punished by up to one year in prison or a fine.
• Professional law: The German Tax Advisory Act (StBerG) requires confidentiality. A breach can lead to professional sanctions from the chamber.
• Data protection law: The GDPR requires a legal basis, data minimisation (Art. 5), and a contract with every processor acting on your behalf.
• Special categories: Health or social data is specially protected under GDPR Art. 9. It never belongs in an AI tool unprotected.

Which client data is critical

Not every field carries the same risk. But the typical data in a tax firm is almost always identifying. The table below shows which item carries which risk, and how to protect it before you work with an AI.

How to anonymize before the AI

The safest rule is easy to remember: no plain text goes to the AI. You replace every sensitive detail with a placeholder and restore the real values only in the answer. Here is the step-by-step approach:

1. 1Scan the text: which names, numbers, amounts, and addresses appear?
2. 2Replace each sensitive item with a neutral token, for example [CLIENT_1] or [TAX_ID_1].
3. 3Send only the anonymized text to the AI. The real values stay with you.
4. 4Restore the original values in the answer once the result comes back.
5. 5Do not store client data in chat histories you cannot control.

Doing this by hand is possible but error-prone. A single missed tax number is enough for a violation. That is why a tool that handles the replacement automatically and locally helps.

What ONYRI Sanitize handles for you

ONYRI Sanitize is built for exactly this step. It detects sensitive data in your text and replaces it with reversible tokens before the text reaches an AI. What matters is how this works technically:

• Detection runs 100 percent in your browser. The mapping from token to real value never leaves your browser.
• When you use the built-in Chat, only the already-anonymized text is sent to the AI model, never the original data.
• The real values are restored only in the answer, right inside your browser.
• You can add your own rules, for example for client numbers or your firm's internal codes.

To be clear: ONYRI pseudonymizes, it does not make your data anonymous in the legal sense. That lowers your risk clearly and supports data minimisation under GDPR Art. 5. It does not replace a legal review or any of the duties below.

Organizational duties in the firm

Technology alone is not enough. Alongside anonymization, you need clear rules and contracts. These points belong in every firm that uses AI:

• Data processing agreement (DPA): Every external service processing data on your behalf needs a contract under GDPR Art. 28. Check whether the AI vendor offers one.
• Check vendor statements: Whether a service uses inputs for training differs by plan. According to OpenAI, content from its business and API tiers is not used for training by default. Such statements change often, so check the vendor's current documentation.
• Internal policy: Put in writing which AI tools are allowed and that only anonymized text may be entered.
• Training: Every staff member must know which data never goes to an AI in plain text.
• No special categories: Health and social data (GDPR Art. 9) never belongs in an AI tool unprotected.