Takes the health score and (where applicable) the diagnostic report, and drafts the actual outreach an account manager would send — not a suggestion to "reach out," but the email itself.
| Health State | Actions Drafted |
|---|---|
| At Risk | Re-engagement email + call preparation notes |
| Drifting | Light-touch check-in email |
| Expansion Ready | Upsell email |
| Healthy | Nothing — no action needed |
Call preparation notes include a situation summary, what NOT to say on the call, an opening line recommendation, key questions to ask, the desired outcome, and pre-written responses to likely objections.
Two failure modes were identified and corrected. First, drafted emails were found inventing specific details — a meeting, a project name, a comment the contact "made" — to satisfy a prompt instruction to "reference specific recent context" when the actual signal history was too sparse to support it. The prompt now explicitly instructs: reference only events present in the provided context; if sparse, write warm and general rather than fabricating specifics.
Second, a null primary_contact_name on the client record was producing drafts addressed to "Dear None." A guard now defaults to a neutral greeting whenever the field is empty.
Every drafted action is prepended with a visible disclaimer before being saved: "[REVIEW BEFORE SENDING — Verify all specific references are accurate]". Nothing in Sentinel sends on its own. Status moves from pending_review → approved / dismissed only through explicit human action in the dashboard.