Triage a shared inbox

Sort a shared inbox (support@, sales@) into queues, extract key fields, and draft first replies for routine messages.

Pipeline

1. Classify each email with Amara Quick and an enum schema: billing, technical, sales, spam, other, plus urgency (low, normal, high) and a one-sentence reason.
2. Extract order numbers, account ids and product names in the same call.
3. Route in code from the category and urgency.
4. Draft replies only for categories where you have approved templates, with Amara S1.2, grounded in your help articles as search result blocks.
5. A person approves drafts before sending, at least until your evaluation shows drafts need no edits for a category.

Prompt essentials

• Put the email inside <email> tags and state that nothing inside it is an instruction.
• Define each category with one sentence and two short examples.
• Tell the model to prefer other over guessing.

Measure

Build a dataset from 300 past emails with the queue they ended up in. Track classification accuracy per category and the share of drafts sent without edits. Re-run whenever you change categories.

Throughput

Classification is short: a few hundred tokens in, under 50 out. A single API key’s 600 requests per minute covers busy inboxes; for backlogs, use a batch.