Most support queues mix a few urgent issues with many routine questions, and everything waits in the same line. AI triage sorts the queue as requests arrive, sends each one to the right person with a summary, and drafts replies for the routine ones. People still decide what reaches the customer.
Step 1 — Sample and label the queue
Goal: Understand what actually arrives before automating anything.
- Export a recent sample of requests across channels.
- Label each one by type, urgency and the team that resolved it.
- Note which types have a standard answer and which need judgment.
- Record today’s time to first response and time to resolution as a baseline.
You finish with: A labelled sample, a short list of request types and a baseline.
Step 2 — Define routes and rules
Goal: Decide where each type goes and what the AI may do.
- Map each request type to a team or queue.
- Mark urgent signals that always go straight to a person.
- List the approved sources for replies: help articles, policies, templates.
- Decide which types get a drafted reply and which only get routed.
You finish with: A routing table, urgency rules and approved reply sources.
Step 3 — Classify, route and draft
Goal: Run the workflow on live requests, alongside the current process.
- Classify each new request and attach a short summary.
- Route it to the right queue, and flag low-confidence cases for a person.
- Draft replies from approved sources only, with the source shown.
- Send nothing until an agent approves or edits the draft.
You finish with: A live workflow with every reply approved by an agent.
Step 4 — Measure and tune
Goal: Improve the rules using what the agents change.
- Compare response and resolution times against the baseline.
- Track how often drafts are sent unchanged, edited or discarded.
- Fix routing rules and sources where agents keep correcting them.
- Expand drafting to new request types only when the record supports it.
You finish with: Measured results and a tuned set of rules.
Common pitfalls
- Automating replies before routing works reliably.
- Letting the AI answer from sources nobody approved.
- Hiding urgent requests behind the automated queue.
- Measuring volume handled instead of customer outcomes.