Phase 1 — Observe#
The deployment starts narrow: one team, one application or agent pointed at the gateway, and a small group of MDM-managed devices running the Salus Desktop endpoint agent. Salus runs in observe mode — nothing is blocked, nothing is rewritten.
What you get out of this phase is evidence, not opinion: a measured report of what detected sensitive data was leaving to external AI providers, from which surfaces, in which data classes. Detection quality and added latency are measured on your own traffic during this phase rather than quoted from a datasheet.
Phase 2 — Protect#
With the observed baseline agreed, tokenization is switched on for the pilot scope. Users keep the same tools and the same workflow; the difference is that external providers now receive typed tokens for detected governed values, and answers are restored inside your perimeter.
Before enforcement begins, failure behavior is agreed per data class — which classes fail closed, which surfaces may degrade visibly, what gets held for human review. The defaults are conservative, and the choices are recorded as policy, not tribal knowledge.
Expand — when ready#
Expansion is additive rather than disruptive. Each increment reuses what the pilot already proved:
- More teams and device groups — the MDM package rolls out to wider rings.
- More applications and agents — each new caller is a base-URL change plus a scoped virtual key.
- More policy depth — additional data classes, approval requirements, and per-destination rules.
- Higher-assurance controls — HSM/KMS key custody, network-lane capture, and air-gapped operation come in when and where the environment requires them, as described in Deployment Options.
There is no fixed pilot duration — the trigger for each step is that the previous scope is measured and stable, not a date on a slide.