Draft by default, send by exception
The tempting version of email automation sends replies on its own, and it is the version that produces the incident. Complaints, legal matters, anything with a number in it and anything from a customer already unhappy are all cases where a confident wrong reply is far more expensive than a slow right one. The default is therefore a draft prepared in the agent's own reply, with the classification, the retrieved context and the reasoning attached, so the person approving it spends seconds rather than minutes. Auto-send is enabled per category, deliberately, for categories where being wrong is cheap.
A taxonomy your team already uses
Triage quality depends far more on the categories than on the classifier. Categories that mirror how the team actually divides work produce useful routing; categories invented for the project produce arguments about where things belong. We derive the taxonomy from a sample of real messages and the existing ownership map, keep an explicit unclear bucket rather than forcing every message into a category, and track what lands there — a growing unclear bucket is the earliest signal that the taxonomy or the business has moved and the rules need revisiting.
Test inbox triage from receipt to resolved owner
For a shared inbox pilot, capture the mailbox, message and thread IDs, sender, received time, subject and attachment references before classifying intent and urgency against the team's existing categories and SLA rules. Keep attachments untrusted input and follow the mailbox's malware and access controls; send ambiguous, sensitive or high-impact messages to a person, with auto-send disabled unless a category is separately approved. Create one owned task or draft per message ID so webhook retries do not duplicate work, and preserve a repair queue if CRM or ticket creation fails. Compare baseline and pilot routing corrections, time to assigned owner, draft edit rate and unresolved backlog on the same representative sample; the inbox owner reviews exceptions and verifies recovery before closing the pilot.
How to measure the pilot
A pilot is accepted against the current process, not an industry promise. Agree the sample, baseline and owner before launch, then review the same measures after the workflow has handled real cases.