How semantic redaction works

Semantic redaction detects sensitive entities in a prompt, replaces them with context-aware tags, forwards the safe version to the model, and records what happened for review. The model still sees the structure of the task, even though the raw identifiers have been replaced before it goes out.

Step 1: Inspect the prompt in context

The first step is classification. The supervision layer evaluates the prompt in the context of the firm's industry and risk model to decide whether it contains sensitive information that should not reach an external AI provider.

That includes obvious identifiers such as names, account numbers, and medical records, and it includes contextual sensitivity such as matter details, compensation, portfolio information, and confidential strategy.

Step 2: Replace raw values with semantic tags

Once the layer identifies a sensitive element, it swaps the raw value for a tag that tells the model what kind of thing it is. This preserves the logic of the request without exposing the underlying data.

A prompt about a client portfolio keeps its structure even when the client name, the value, and the account reference are all rewritten into tags.

Step 3: Send the sanitized prompt to the model

The external model receives only the safe prompt. It can still generate a useful answer, because the prompt retains the task intent, the entity roles, and the overall structure.

This is the difference from brittle masking that strips so much context the model can only produce vague output.

Step 4: Preserve evidence for review

A serious implementation does not stop at rewriting the prompt. It records what was classified, what was transformed, who initiated the interaction, and whether the event should be reviewed later by compliance, legal, or security.

• The timestamp and the model provider.
• The risk category and the data types detected.
• Any reviewer status changes and exported evidence.

That is what turns redaction from a narrow privacy feature into a defensible governance control.

Common questions

Is semantic redaction only for personal data?

No. It also protects confidential commercial information such as deal terms, portfolio allocations, underwriting logic, or client campaign plans.

Does it slow users down?

It should not meaningfully disrupt normal usage. The goal is to preserve the employee workflow while improving control and visibility.

Why is the audit trail important?

Because regulated firms need to demonstrate supervision, not just claim that redaction happens. The evidence is what lets leadership and compliance defend the control later.