Lesson 87.2's SUPPORT-TRIAGE.md needs a concrete starting grid — this is it. Copy the matrix, assign owners, set your confidence bar, and run three test messages through it before connecting any tool. AI sorts; humans own what matters.

The matrix (intent → route)

| Intent | Example | Urgency | AI may | AI must never | Route → owner (SLA) |

|---|---|---|---|---|---|

| How-to | "which formats do you accept?" | Low | retrieve + draft-resolve if confidence ≥ ___ and fully sourced | invent passages | resolve-draft → ___ |

| Bug report | "export missing rows" | Med | classify + retrieve + draft for reviewer | promise a fix date | route → product ___ (24h) |

| Billing | "charged twice" | High | classify + handoff only | reply, refund, quote policy | escalate → finance ___ (4h) |

| Account/identity | "transfer ownership" | High | classify + handoff only | grant/revoke access | escalate → ops ___ (4h) |

| Frustration / 2nd contact | "still broken, unacceptable" | High | classify + handoff with history | send calming prose as resolution | escalate → ___ (same day) |

| Ambiguous | "just fix everything??" | Med | ask one clarifying question OR handoff | guess the ask | route → ___ with context |

| Sensitive mistake | "your advice deleted my data" | Highest | acknowledge receipt + escalate instantly | minimize, explain away | escalate → founder ___ (hours) |

Approved knowledge (AI quotes ONLY these): docs version ___ / FAQ ___ / changelog ___ — update owner ___. Everything else gets "I don't have a verified source — routing you to ___," never a plausible-sounding answer.

Handoff format (every escalation carries)

Classification + confidence + retrieved passages + draft + customer history + what was tried. Bare forwarding ("fyi") is a second injury — the customer repeats themselves and trust drops twice.

Weekly tune (15 min, feeds 87.5)

Review the week's misroutes: which intents were confused, which confidence threshold misfired, which approved passage went stale. Log one pattern (recurring question → docs/product candidate) and one threshold or passage fix. The matrix is a living file — date each change, keep the history. Start strict (more handoffs), loosen only when the logs earn it.

Filling it in: three worked rows (Sonariq Desk)

How-to → resolve. "Which formats do you accept?" classifies how-to/low; retrieval pulls the method guide's format passage; confidence high + fully sourced → AI resolves with citation and link. Log keeps classification, passage ID, and draft. Owner reviews weekly samples only.

Billing → escalate. "Charged twice for the report" classifies billing/high regardless of confidence; AI takes no further action beyond attaching history and routing to finance within 4 hours. The workflow's credential cannot issue refunds (scope denial per 88.3), so even a misconfiguration can't pay out.

Ambiguous → clarify once, then human. "Just fix everything??" classifies ambiguous/medium; AI asks one bounded clarifying question ("which report and what looks wrong?") and hands off with the reply attached if the answer stays unclear. Guessing the ask is forbidden — the cost of solving the wrong problem exceeds the cost of one human read.

Calibration notes

Set the confidence bar high at first (e.g. ≥ 0.85 plus fully-sourced plus low-stakes for any auto-resolve) and lower it only on evidence: two clean review cycles with zero misroutes. Track per-intent accuracy, not global accuracy — 95% overall with 60% on billing is a failure wearing a good average. Date every threshold change; when a passage goes stale (pricing, policy, feature), update the approved-knowledge version the same day and note it in the decision log (89.5).