Confidence score
One number from 0 to 100, shown on every request across web and iOS, answering the only question an approver actually has: how safe is it to say yes?
The five bands
| Band | Score | Read it as |
|---|---|---|
| Allow | 85 to 100 | Strong, consistent evidence this is a legitimate binary. A quick identity glance is enough. |
| Soft allow | 70 to 84 | Probably safe. Approve unless something in the context (parent process, path, arguments) looks off. |
| Review | 30 to 69 | Genuinely uncertain: mixed or missing signals. Read the AI summary and the launch context before deciding. |
| High risk | 15 to 29 | Multiple negative indicators. Deny unless you can independently verify the business need. |
| Hard block | 0 to 14 | The evidence says malware or tampering. Deny, and look at what else that device has been doing. |
What feeds the number
The score has one producer at a time, labeled next to the badge:
- ML: once the machine-learning scorer has trained on enough of your tenant's own approve/deny history, its confidence becomes the score. It reflects how your team actually decides.
- AI: before the ML scorer matures, the score falls back to the AI analysis legitimacy score for the binary.
Both produce 0 to 100 and both are clamped before banding, so the pill always reads the same way regardless of source. See Risk scoring for what each layer looks at.
What the score is not
It is advisory, never an enforcement mechanism. A high score does not bypass the signature gate, MFA requirements or deny rules; a low score does not block anything by itself. Enforcement belongs to rules and policy. The score exists to make the human decision faster and better calibrated.
Auto-approval (when enabled in policy) keys on repeated identical approvals of the same binary for the same user, not on the confidence score alone. The score never silently approves anything.