Understand. Decide. Apply.

PII: protecting personal data before AI

AI should understand a task. It does not necessarily need a real name, bank account or private phone number to do so. Protection starts with selecting the information actually needed.

AI LabsUpdated 2 sources

PII masking in practice

After masking

Please send the draft to [NAME_1]: [EMAIL_1]. Follow-up questions also go to [EMAIL_1].

This local demo processes fictional examples only in the browser. Simple rules recognise email addresses, German IBANs and names with a salutation. It illustrates masking rather than production detection or guaranteed anonymisation.

What is personally identifiable information?

PII stands for Personally Identifiable Information: information that identifies a person directly or in combination with other details. An email address is an obvious example; a rare role in a small team can also identify someone.

Not all confidential information is PII. Product plans, margins or source code can contain business secrets without naming a person. A data policy therefore complements automated detection.

Detect, replace and restore under control

AI Link's protection path uses patterns and language-based detection, including Presidio, to identify personal information. Detected spans are replaced with placeholders for external processing. A protected mapping allows results to be restored to the right context afterwards.

Where re-identification remains possible, this is generally pseudonymisation and GDPR remains relevant. True anonymisation requires that a person can no longer be identified, including through reasonably available additional information.

Plan for the limits

No detector reliably finds every sensitive detail. Typos, unusual names, images and indirect clues are challenging. AI Link also considers approved providers and the trust boundary of each route. Protected paths should block when detection is unavailable; deliberately configured exceptions need a separate assessment.

The browser demo illustrates the principle with simple rules. It is not an evaluation of production detection quality. A pilot needs representative, approved test cases and a review of missed information.

Sources & further reading

Original research, provider documentation and further reading on the concepts and figures discussed here.

Talk to AI LABS

Bring AI into the business. Securely and step by step.

A personal conversation clarifies suitable use cases, data protection requirements and the next steps, from a first demo to a pilot.

Get in touch