How to Anonymize Company Data Before Using ChatGPT or Claude

Anonymisation starts with purpose

Write the AI task in one sentence before transforming the data. For example: summarise the reasons customers churned, classify support themes, or identify anomalies in monthly costs. The purpose determines which relationships and values must survive.

Then remove whole fields, records, and attachments that are not necessary. Data minimisation is more reliable than trying to mask every item in an oversized source. It also produces a smaller, clearer prompt.

Do not assume that company information is low risk because it is not personal data. Prices, forecasts, product plans, customer lists, contract terms, system architecture, and investigation findings can create serious commercial or contractual exposure.

Identify direct and indirect identifiers

Direct identifiers include names, email addresses, telephone numbers, customer numbers, legal entity names, account IDs, and public profile links. Indirect identifiers—sometimes called quasi-identifiers—include combinations such as location, exact date, role, amount, rare event, or project description.

A single indirect field may be harmless. Several together can make a subject obvious. A row that says Head of Security, Sapporo office, joined 4 April, and handled a specific incident may identify one employee even after the name is removed.

Review free-text fields separately. Meeting notes, ticket descriptions, contract clauses, filenames, worksheet names, comments, and code examples often repeat identities that a column-based transformation misses.

Choose a transformation that preserves only what you need

Use consistent placeholders where analysis depends on repeated actors. Randomly replacing the same customer with a different label on every row destroys the relationship and can make the output misleading.

If a lookup table can restore the original identities, protect that table separately and do not upload it with the transformed dataset. Treat the result as pseudonymised data with remaining risk rather than claiming it is anonymous.

Test the transformed copy

1. Search for every original name, domain, ID prefix, project term, and distinctive phrase.
2. Review the smallest groups and rarest records; these are often easiest to re-identify.
3. Ask whether public knowledge or another internal dataset could reconnect the remaining details.
4. Inspect metadata, comments, filenames, hidden content, and attachments.
5. Check that totals, ordering, and relationships required for the AI task still make sense.
6. Have a reviewer unfamiliar with the transformation try to infer the subjects.

NIST describes de-identification as a risk-management process rather than a one-time masking operation. The acceptable residual risk depends on the data, likely recipients, context, and consequence of re-identification.

ChatGPT and Claude still require destination review

A prepared copy does not remove the need to check service terms and settings. Consumer and commercial versions of an AI service can handle content differently, and policies can change. Confirm the exact workspace, training controls, retention, administrator access, and any connectors before upload.

When risk remains high, use an approved enterprise environment, an internally controlled model, a synthetic sample, or no upload. The tool selection follows the data decision—not the other way around.