Anonymization
How Copilot via API detects and anonymizes personally identifiable information (PII).
Copilot via API can anonymize personally identifiable information (PII) in messages to protect user privacy.
Overview
Messages that you attach to a thread or send via the Chat-Completions API can contain PII. When anonymization is configured, the service detects PII before it processes a message. Detection uses a combination of static pattern matching and named-entity recognition (NER). Then the service replaces each detected entity with a placeholder tag.
The following entity types can be anonymized:
| Entity Type | Placeholder |
|---|---|
| Email addresses | <EMAIL_ADDRESS> |
| Phone numbers | <PHONE_NUMBER> |
| IBANs | <IBAN> |
| Credit card numbers | <CREDIT_CARD_NUMBER> |
| German tax IDs | <TAX_ID> |
| German social security numbers | <SSN_DE> |
For example, the service replaces an email address with <EMAIL_ADDRESS> and a tax ID with <TAX_ID>. Downstream services only see the anonymized form.
The set of entity types is part of the configuration of each Copilot. The configurations range from no anonymization to all entity types in the table above.
We conclude a data processing agreement (German: Auftragsverarbeitungsvertrag, AVV) with each customer to protect personal data that messages can contain.
Even though PII is anonymized, we strongly recommend that your users share only the minimum personal information necessary in their messages.
Names and birthdates
The service can also detect names (<PERSON>) and birthdates (<BIRTH_DATE>). We do not anonymize these entity types by default. The false positive rate within Haufe content is high. For example, European laws are referred to by names.
Data deletion
When you delete a thread, the service does not anonymize the messages. It redacts the content of every message. See Data Deletion for details.