1. Build the peak baseline
Review a comparable sale period or the latest four weeks. Count questions by category, channel and outcome—not just total volume. Separate product and promotion questions, WISMO, address changes, cancellations, returns, payment problems and cases that reopened.
Baseline sheet: category → conversations → median first response → resolution → reopenings → owner. If no history exists, start with a seven-day sample and label it as provisional.
2. Prevent the predictable
Make offer dates, exclusions, delivery expectations, return conditions and product facts consistent across the storefront, campaign messages and support answers. Shopify Inbox supports instant answers and quick replies, but every published answer still needs a current source and a named owner.
Pre-purchase: promotion terms, sizing, stock and delivery cutoff. Post-purchase: tracking route, expected scan gaps and the information needed for an exception review.
3. Draw the human boundary
Automate facts; escalate decisions. Refund approval, suspected fraud, duplicate charges, privacy, a parcel marked delivered but missing, sensitive address changes and commitments outside policy require a person with authority and a recorded decision.
Rule: trigger → required evidence → approved draft → human owner → response target → prohibited promise. A fast wrong answer is still a failed support outcome.
4. Run a seven-day rehearsal
Test the twenty questions most likely to arrive, including five edge cases and five prompts that must hand off. Measure accuracy, unresolved conversations, reopenings and escalation quality alongside speed. Freeze unreviewed automation changes during the live peak.
Continue if reviewed answers stay accurate and unresolved or reopened cases do not worsen. Stop immediately on a privacy, payment or unauthorized-refund error.