Deflection is not resolution
For a decade, “automation” in customer service meant deflection: a bot that pointed at an article and hoped the customer would go away. The metric was tickets avoided. Nobody measured whether the customer’s problem was actually solved, because the tools couldn’t solve it — they could only describe how a human might.
Agent changes the unit of work. It can look up the order, check the policy, issue the refund and confirm. So the metric changes too: resolution rate — the share of eligible conversations that end with the customer’s problem fixed and no human involved. Everything in this guide is about moving that number honestly.
A conversation is resolved when the customer confirms it, doesn’t come back, or a procedure completes with an action taken. Aftersales bills on this definition, so the number you optimise is the number you pay for — the incentives line up.
Week 1 — Find the fat head
Open Insights → Resolution rate by topic. Sort by conversation volume. In every support operation we’ve seen, the top five topics are 55–70% of volume, and they are boring: order status, returns, address changes, billing questions, password resets. That’s the good news. Boring is automatable.
- Pick the top topic where Agent is currently handing off rather than resolving.
- Read twenty handoffs. Write down the single action the human took in each.
- You will find two or three actions cover most of them. Those are your first procedures.
Week 2 — Write procedures a new hire could follow
A procedure is not a prompt. It is five to eight steps in plain language, each mapped to a system: look up the order in Shopify → check the return window → if in stock, create a replacement, else refund via Stripe → confirm with the customer. If a person on their first day couldn’t follow it, Agent shouldn’t either.
The best procedures are the macros your team already trusts, with the “ask a manager” step replaced by a rule.
Guardrails come first
Before a procedure goes live, set the spend limit, the escalation triggers and the tone. Agent enforces these at the action layer — a refund over the limit cannot be issued — so you can be generous with scope and strict with limits.
Week 3–6 — Ship one, measure, repeat
Turn on one procedure. Watch three numbers for a week: its resolution rate, its CSAT, and its Monitor score. If CSAT or quality drops, the procedure is wrong, not the customers. Fix the step, not the guardrail. Then ship the next one.
Teams that follow this cadence typically add 6–9 points of resolution rate per procedure for the top five topics. That’s how 40% becomes 70% in a quarter without a single “hallucination” debate — because Agent is doing things, not saying things.
Week 7+ — Close the loop on handoffs
Every handoff your team completes is a procedure waiting to be written. Inbox shows you which handoffs recur; Copilot drafts the procedure from the resolution. Review, approve, and the next customer never needs the handoff.
Median across customers after 90 days: 71% resolution rate, 41-second median time to resolve, 4.8 CSAT on Agent resolutions. Your mileage depends on how boring your fat head is — and most are wonderfully boring.