Deal health

When to run

• "Which deals are at risk this quarter" / "score my pipeline" / "is the Northwind deal healthy"
• Before forecast calls, at stage-gate reviews, when a close date slips

Computer scopes

• CRM: opportunities, stages, amounts, close dates, contact roles
• Email + calendar: thread recency and meeting cadence per opportunity
• Support: open tickets tied to the account
• Notes: call summaries mentioning the opportunity

Workflow

1. Select scope: one opportunity, one owner's book, or the full open pipeline as asked.
2. For each opportunity compute the signal set: days since last inbound reply, meeting cadence trend, stakeholder breadth (single-threaded?), stage age vs. this pipeline's historical median, open support friction, and champion activity.
3. Score 0–100 with a band (healthy / watch / at-risk). Weightings live in this skill's references file; do not improvise new signals silently.
4. For every flag, attach the evidence: the specific thread, the missed meeting, the aging stage.
5. Rank output by risk-weighted amount, not raw score — a shaky $400K deal outranks a shaky $4K one.
6. Recommend one concrete next action per at-risk deal (reengage champion, multi-thread, requalify).

Guardrails

• Autonomy: L0. Scores and evidence only; never edits the opportunity or emails anyone.
• Label the score as model-derived guidance, not ground truth; show the signal breakdown on request.
• Single-account users see only their permitted book.

Output

Eval fixtures

• Set: 120 closed opportunities (60 won / 60 lost) scored as-of 30 days pre-close, blind.
• Rubric: AUC of risk band vs. actual outcome; every flag citation resolves to a real record.
• Ship gate: AUC ≥ 0.75; citation resolution 100%; no flag without evidence.