ChatGPT in Business: How to Protect Sensitive Data

When employees upload documents to ChatGPT or enter text containing personal data, the following happens:

Data Transfer to OpenAI

All inputs are transmitted to OpenAI's servers in the United States. With the free version and ChatGPT Plus, this data may be used for model training — unless the user explicitly disables this option. Even with ChatGPT Enterprise, where no model training occurs, data is processed on US servers.

Typical Workplace Scenarios

The range of data protection risks is enormous. Employees regularly upload the following documents to ChatGPT:

• Contracts — with full names, addresses and account numbers of the contracting parties
• Emails — with sender and recipient data, often including signatures with phone numbers
• Personnel files — salary data, social security numbers, performance reviews
• Client lists — names, addresses, order histories, payment information
• Expert reports and assessments — with patient data, client information or trade secrets

Legal Consequences

Transmitting personal data to OpenAI without a legal basis constitutes a GDPR violation. Companies risk:

• Fines — up to EUR 20 million or 4% of annual global turnover (Art. 83 GDPR)
• Compensation claims — affected individuals may claim damages (Art. 82 GDPR)
• Reputational damage — data leaks via AI tools are increasingly reported publicly
• Loss of trade secrets — once fed into training data, the information is irretrievably lost

Pseudonymization is a procedure explicitly referenced in the GDPR (Art. 4(5)) in which personal data is replaced with consistent placeholders. The key difference from anonymization: the replacement is reversible — the original data can be restored via a protected replacement table.

How the Workflow Works

The workflow from the original file to the finished AI result consists of four steps:

Step 1: Import Document

You load the document containing sensitive data into the pseudonymization software. This can be contracts, emails, expert reports, personnel files or any other documents. The software supports over 70 file formats including PDF, Word, Excel and scans.

Step 2: Automatic Pseudonymization

The software automatically detects personal data and replaces it with consistent pseudonyms:

• John Smith becomes Person_A
• 123 Main Street, London EC1A 1BB becomes Address_A
• GB29 NWBK 6016 1331 9268 19 becomes IBAN_A
• john.smith@company.co.uk becomes Email_A

Crucially, the pseudonymization is consistent. If John Smith appears in 50 documents, he becomes Person_A everywhere. This preserves context without revealing the real identity.

Step 3: AI Processing

The pseudonymized document is submitted to ChatGPT (or any other AI tool). The AI sees only pseudonyms — no real names, addresses or account numbers. The analysis, summary or translation is performed on the pseudonymized data.

Step 4: Re-Identification

After AI processing, the pseudonyms in the result are automatically replaced with the original data. The replacement table ensures that Person_A becomes John Smith again, Address_A becomes 123 Main Street, and so on. The final result is a complete document with original data — as if the AI had worked directly with the real data.

Not all data is equally sensitive. The following overview shows which categories of personal data are particularly worthy of protection and how they are pseudonymized:

Docuflair Mask automatically detects all of these categories and replaces them consistently across the entire document set. The categories are fully configurable — you can determine which data types should be pseudonymized and which should remain in plain text.