Redaction
Our models find and replace sensitive information before anything leaves the platform, and restore it in the reply you read.
How Redaction works
OurAI combines fixed patterns (regex) with a language model (NER) that catches what patterns cannot, in both Danish and English. Together they find personally identifiable information (PII), business details and the terms your organisation has defined itself. Each detected piece of information is swapped for a token before the message leaves the platform (tokenisation), and the same information always becomes the same token, so the conversation still reads coherently for the AI model.
Personally identifiable information (PII)
Information that points to a person, including names, CPR numbers, telephone numbers, e-mail addresses, home addresses and dates of birth.
Business details
Information that identifies your organisation or its finances, including CVR numbers, organisation names and account numbers.
Your own terms
Words and phrases your organisation defines itself, such as customer names, project names or internal terms.
Review and approval
Everything is held for review before it is sent. Every detected piece of information is swapped out and highlighted, so you see exactly what leaves the platform. If you judge a piece of information safe, you can keep it, and you can mark anything that was not caught. Nothing is sent until you approve.
What is stored
What you send is stored without the sensitive information. The link between a token and the real information is never stored. It exists only while the reply is being generated.
| DATA | STORED? |
|---|---|
| What you send, without the sensitive information | Yes |
| The reply, as you read it | Yes |
| The link between token and information | Never |
Retention periods for all data types are described under Data retention.
Developed further at DTU
Redaction is built by OurAI and is continuously improved through a special course at DTU. The design follows a single principle: when there is doubt about a piece of information, it is swapped out. Missing something is a serious error. Removing too much at most costs the AI model a little context. Your data is never used for development or testing.