Sentinel is a five-stage autonomous intelligence pipeline that continuously monitors every client relationship in a consultancy's active portfolio. It ingests real communication data, classifies it into structured signals, scores each account's health, diagnoses the ones at risk, and drafts human-reviewed recovery or expansion outreach.
Nothing sends automatically. Every drafted action lands in a review queue. The system's job is to do the cognitive work of noticing and diagnosing — not to act unilaterally on a client relationship.
Agent 0 — Data Ingestion & Identity Resolution
(real email/calendar/manual events → sentinel_raw_events)
↓
Agent 1 — Signal Collector
(classifies raw events into structured signals)
↓
Agent 2 — Health Scorer
(0–100 score + state: healthy / drifting / at_risk / expansion_ready)
↓
[conditional routing]
↓
Agent 3 — Diagnostician (only for drifting / at_risk)
(root cause, relationship timeline, recommended approach)
↓
Agent 4 — Action Drafter
(drafts state-appropriate outreach, awaiting human review)
An account manager BCCs client emails to a dedicated ingestion address, or logs a quick manual note after a call. Sentinel resolves which client the communication belongs to, summarizes it while preserving friction and tension (never smoothing over conflict), and classifies it into a structured signal with sentiment and concern flags.
Every account gets a health score, recalculated whenever new signals arrive. Scores that swing more than 20 points without a genuinely concerning event are automatically dampened — the system doesn't cry wolf.
For any account showing drift or risk, a diagnostic agent produces a grounded root-cause report — citing the actual signals, never inventing history it can't see. An action-drafting agent then writes the recovery email, call-prep notes, or upsell message, flagged for human review before anything is sent.
A live web dashboard gives the consultancy: a portfolio overview with health states across every account, full signal timelines per client, diagnostic reports, a review queue for AI-drafted actions, a live cross-portfolio signal feed, and an unmatched-events queue for teaching the system which email addresses belong to which client.
If the system flags one at-risk $80K–$200K client relationship early enough to save it, it has paid for itself several times over in a single engagement. That is the entire ROI conversation, and it takes under a minute to make.