Triage rarely makes the AI strategy deck, which is a shame, because it is frequently the first Claude project to pay for itself. Every shared inbox and ticket queue has a hidden pre-processing tax: a person reads each message, decides category, urgency, and owner, and forwards it along — minutes per message, hours per day, before any actual work begins. The task is pure classification of messy human language, the stakes per decision are low (a misroute costs a bounce, not a lawsuit), and the volume is high enough that small time savings compound into real money.
What the triage step actually does
For each incoming message, a single Claude call produces a handful of labels: category (billing, technical, cancellation, sales inquiry, spam…), urgency (is a system down? is a deadline mentioned?), sentiment or escalation risk (is this customer about to leave?), routing target (which team or queue), and optionally a one-line summary so the assignee gets the gist before opening. Unlike the keyword rules most ticket systems offer, a language model reads meaning: "I was charged twice and nobody has answered in a week" is billing and urgent and escalation-risk, no matter which keywords appear.
The output must be machine-readable, since its consumer is your ticketing system, not a person. Define the labels as a schema — with each field limited to an explicit list of allowed values — using the technique in structured outputs. An enum of twelve categories you chose beats free-text categories the model improvises every time.
Why the economics work so well
Triage is short-input, short-output work at high volume — exactly the shape Claude Haiku 4.5 is built for. At list price ($1 per million input tokens, $5 per million output), a typical email of a few hundred tokens plus a compact label output costs a small fraction of a cent to classify. Set against even one minute of human pre-reading per message, the arithmetic is not close, which is why triage tends to clear ROI review faster than flashier projects. The workload also uses only the core Messages API, available on all four platforms — Bedrock, Vertex AI, Foundry, and Claude Platform on AWS — so your cloud choice doesn't constrain it. (If you were counting on the Batch API to classify a historical backlog cheaply, note it is not available on Bedrock or Vertex AI; a simple loop over ordinary requests does the job there.)
Design for the misses, not the hits
No classifier is perfect; good triage systems are built around that fact. Give the schema a confidence field and an explicit unclear category, and instruct Claude to use them instead of forcing a fit. Then route by confidence: high-confidence messages flow straight to their queue; low-confidence ones land in a small human-review queue — which is just the old triage job, shrunk to the genuinely ambiguous residue. Auto-routing is also the right ambition level to start with. Auto-responding is a different project with different risks; triage earns trust first by putting messages in the right hands faster.
A rollout that avoids the common traps
Run shadow mode first. For two or three weeks, classify every message but let humans keep routing. Comparing the model's labels to human decisions gives you a measured accuracy baseline before anything changes — and usually reveals that humans disagree with each other too, which resets expectations honestly.
Fix the taxonomy before the prompt. If your categories overlap ("billing" vs. "account issues") humans and models will both guess. Most "model accuracy" problems in triage are really taxonomy problems.
Mind the personal data. Emails carry names, account numbers, and grievances. Log the labels, not full message bodies; apply your existing data-handling rules to the pipeline. Running through your existing cloud platform keeps traffic inside a boundary your security team already governs — confirm specifics with your provider.
Watch for drift. New products, new failure modes, and new customer language appear constantly. Review a sample of the unclear queue monthly; it is a free early-warning system for categories your taxonomy is missing.
Where to go next
The schema-and-enum technique that makes triage reliable is covered in structured outputs. When you're ready to go beyond routing toward drafting replies, read building a customer support assistant. Model tiers and prices are summarized in the model lineup.