THE REVIEW WINDOW BEFORE ANYTHING IS SENT
1
Sensitive content is detected in the document and highlighted in place, so the user sees exactly what was found and where.
2
Protected items can be reviewed and corrected — an incorrect detection can be removed, and a missed value can be marked manually.
3
The user explicitly continues to the AI workflow — nothing is sent until the review is confirmed.

Why review exists#

Detection is layered and tunable, but no detector should be presented as perfect. Human review is the architectural answer to the question every security team asks — what if the model misses something? Before content leaves the environment, the person who actually understands the document gets to see what was detected, fix what was wrong, and add what was missed. The correction happens at the moment it is cheapest: before egress, not after an incident.

Where review appears#

Human review exists in both product experiences where configured:

  • Salus Desktop — the Gateway's endpoint layer shows the review window for governed traffic captured on managed devices.
  • Salus Workspace — the same review flow is built into the client, including manual marking of items the detector missed.
  • Policy-required approval — the policy engine supports a require-approval mode per data class or workflow, so selected categories cannot leave without an explicit human decision.

Review is a configurable control, not a mandatory gate: most workflows run straight through with automatic tokenization, and policy decides which classes, users, or destinations warrant the extra step. That keeps the protection proportionate — friction where the risk justifies it, invisibility everywhere else.

What review feeds back#

Every correction is signal. Removed detections and manually marked items are recorded (as token-level events, not raw values) and inform detection tuning — thresholds per class, custom entity definitions, and observe-mode measurement of the miss rate on your own traffic. The review window is both a control and the ground truth that makes the automated layer measurably better.