Lesson 87.1 taught you to read support as a sensor: every ticket carries an immediate need and a root-cause candidate. Now volume arrives — forty messages on launch day, eleven overnight. You cannot hire your way out, and you must not automate your way into careless replies. The answer is a triage layer: AI that classifies, retrieves, and drafts at speed, with handoff lines it never crosses.

The vocabulary, in plain language

• Triage: sorting incoming requests by need and urgency before solving — the emergency-room desk, not the surgery.
• Intent classification: AI's first judgment — "this is a password reset / a billing dispute / a migration question / frustration with no clear ask." A label, not a solution.
• Approved knowledge: the bounded set of sources AI may quote — the current docs, the verified FAQ, the changelog. If it is not in approved knowledge, AI does not assert it.
• Retrieval: fetching the relevant approved passage before drafting — the evidence-first habit from Class 85, applied to support.
• Confidence: the system's own uncertainty score. Low confidence is a routing instruction, not a shameful secret.
• Handoff: passing the ticket to a human with the classification, retrieved passages, draft, and full history attached.
• Escalation: handoff with urgency — billing, accounts, legal, safety, or a furious customer — to a named owner with a time promise.

The mental model: AI is the night-shift triage nurse — fast, tireless, honest about limits. It takes vitals, pulls the chart, prepares the room. It never performs the surgery, and it wakes the doctor early rather than late.

The loop

Memorize the six stations:

receive → classify → retrieve → draft → resolve/route → record

Receive — every channel lands in one queue with an ID; nothing lives only in someone's DMs. Classify — intent plus urgency plus sentiment: "billing question, high stakes, calm" routes differently from "migration question, low stakes, frustrated." Retrieve — pull the two or three approved passages that bear on the intent; if none match, that *is* the finding. Draft — write the reply from retrieved passages only, with citations or links, in the project's voice. Resolve/route — high-confidence, low-stakes, fully-sourced drafts may resolve (with the human-readable log); everything else routes to the named owner. Record — the classification, passages used, draft, decision, and outcome land on the ticket, feeding Lesson 87.1's pattern read and Lesson 87.3's CRM.

Lines AI never crosses

Five situations always hand off to a human, regardless of confidence:

• Ambiguity — no clear ask, contradictory details, or two plausible intents. Guessing at meaning is how you solve the wrong problem confidently.
• Frustration — anger, disappointment, or a second contact about the same issue. Emotion needs a person with authority to make it right, not a polished paragraph.
• Billing and money — charges, refunds, plan changes, failed payments. Money movement stays human (Lessons 84.4, 86.4).
• Accounts and identity — password resets beyond the documented flow, access grants, data deletion, ownership transfers. Least privilege applies to AI too.
• Sensitive mistakes — the product lost data, leaked something, or gave wrong advice with consequences. Acknowledge, escalate, and let the owner repair — never minimize by automation.

The handoff carries everything: classification, confidence, retrieved passages, draft, and what the customer already tried. "Forwarded to support@" with no context is not a handoff; it is a second injury.

Exercise: write the triage spec

Create SUPPORT-TRIAGE.md — the Class 84 contract applied to support.

# SUPPORT-TRIAGE.md — [Project]

## Categories (intent → urgency → default route)
- ___ → ___ → resolve-draft / route-to-___ (owner ___)
- ___ → ___ → resolve-draft / route-to-___ (owner ___)
- (cover at least: how-to, bug report, billing, account, frustration, ambiguous)

## Approved knowledge (AI may quote ONLY these)
- ___ (docs version/date) / ___ (FAQ) / ___ (changelog) — update owner ___

## AI may (permitted)
- classify / retrieve / draft from approved passages / resolve ONLY when: confidence ≥ ___ AND stakes low AND fully sourced

## AI must never (prohibited)
- send billing/account/sensitive replies / invent passages / edit tickets' categories after human set / contact outside the queue

## Handoff + escalation
- Always handoff: ambiguity / frustration / billing / accounts / sensitive mistake → owner ___ within ___ hours, with classification + passages + draft + history
- Record: every ticket keeps ___ (fields), reviewed weekly by ___ for pattern read

Worked mini-example — Sonariq Research Desk: categories "source-format question → low → draft-resolve," "report error claim → medium → route analyst," "invoice dispute → high → escalate finance within 4 hours," "angry second contact → high → escalate founder." Approved knowledge: the method guide (dated), the pricing page, the sample report. AI resolves "which formats do you accept?" from the guide's passage; it never touches the invoice thread beyond classifying and handing off with full history.

Finish line: a SUPPORT-TRIAGE.md with categories, an approved-knowledge list, explicit allow/forbid lines, named owners, and a handoff format.

Verify quickly: run three test messages through the spec on paper — a how-to, a billing dispute, an angry ambiguous complaint. If each lands on a named route with the right attachments, the spec works. If the billing dispute could self-resolve, the forbid line is broken.

Common failure mode: the confident intern — AI quoting expired docs, inventing policy, replying to refunds, and logging nothing. Every one of those failures is a missing line in this file, not a model mystery. Write the lines before connecting the tool.

Check your understanding

1. Walk the six-station loop for a "my export is missing rows" ticket. What happens at each station? 2. Why do ambiguity, frustration, billing, accounts, and sensitive mistakes always hand off? 3. What must a handoff carry — and why is bare forwarding harmful?

Next

Triage resolves the moment — but the relationship continues. Lesson 87.3 gives it a memory: a CRM that preserves requests, context, preferences, promises, and next actions, so the next conversation starts informed instead of starting over.