PII protection with audit
Privacy that stays out of your way.
Detected personal data is replaced before any text reaches the model, then restored after the answer. You work with plain text; the model sees placeholders.
Where does your data go when you ask an LLM?
Every request to a cloud model leaves the building. If a customer name, a personnel number or a diagnosis sits in the prompt, it lands with the provider — processed, logged, possibly stored.
For personal data that is a GDPR problem. The usual escape — just banning cloud AI — costs you the benefit. There is another way.
Pseudonymise, ask, re-identify
Before the hand-off the text passes a PII detection step via an external Presidio installation — as a Docker container in your environment or hosted by evanto. Detected entities such as names, addresses, emails or IBANs are replaced with stable placeholders.
The model only ever processes the pseudonymised version. In the answer, the placeholders are swapped back to the original values. Every step lands in the audit log.
The path of one request
-
Detect
Presidio scans the text for personal entities.
-
Replace
Each entity is swapped for a stable placeholder: "Ms Müller" becomes [PERSON_1].
-
Process
The model only sees the pseudonymised version.
-
Re-identify
In the answer the placeholders are swapped back — plain text again, for you.
-
Log
Every step sits in the audit log, traceable and exportable.
What this means for you
- Detected PII never reaches the model in plain text — not even with cloud providers.
- Stable placeholders keep context intact: same person, same placeholder.
- A tamper-evident audit log as evidence for your GDPR documentation.
- Presidio runs externally — as a Docker container or hosted by evanto.
Sounds right for your organisation?
In a 30-minute demo we show the concept on your concrete use case — no obligation, with concrete suggestions at the end.