從 FAQ 試算表遷移至規則表，並確保回覆不中斷

A FAQ sheet and a rule table look alike — question, answer, maybe a keyword column — but they are different artifacts. The sheet documents knowledge; the table executes it. Paste a sheet into a bot and you get keywords that cross-trigger, stale prices that survive a catalog update, and rules whose owner left long ago. The migration below treats the move as a project: classify first, cut into rules second, switch traffic over gradually, and write the rollback conditions before launch day.

Classify Before You Convert

Walk the sheet row by row and sort entries into clusters: pricing, shipping and logistics, sizing and specs, after-sales, process and booking. Expect duplication — the same shipping answer often sits under three differently worded questions, and the clusters exist to expose that. Merge duplicates, drop rows that have drawn no questions in months, and mark the rows that should stay human regardless: discount authority, compensation decisions, anything negotiable. A rule can still acknowledge those topics and page a person, but the decision itself stays with a person. Whatever debt you import into the rule table, you will service for as long as it runs — so leave the dead rows behind.

Cut Clusters Into Rules

Within each cluster, define one intent per rule and attach its variants: exact matches for precise strings such as order numbers, keyword matches for phrasing that varies. Then order the rules — a specific rule outranks a broad one, and a default fallback sits underneath to catch the remainder. Carry the original sheet row reference into the rule name so anyone auditing a fired reply can walk back to its source. For the fallback and priority ladder, the layered-script model maps cleanly: each customer tag points to a scenario and a difference point, specific tags outrank generic ones, and a default reply catches what the tags miss; the tag-to-script mapping guide describes that three-column format and how to review it after launch.

Switch Traffic Over in Gray Steps

Do not flip the whole account in one motion. Whitelist a small slice of traffic — a handful of regular customers, or a single channel — and let the new rule table answer while the old setup stands by. Track three counters on both sides: replies sent, escalations to a human, and messages that mapped to no rule. Expand the slice when the numbers hold steady, then expand again. Where the rules quote from source material, such as price lists uploaded so answers cite real figures, the setup guide for AI auto-replies on personal chat accounts covers how uploaded documents become the fact base the replies draw from.

Write the Rollback Conditions Before Launch

Decide in advance what a failed rollout looks like, in observable terms: the unmapped share climbing after an expansion, a wrong rule answering a pricing question, a complaint that quotes a confusing reply. Name the person who flips the switch, and make reverting a one-step action — the old sheet stays importable, the new table stays editable. After the switch, keep reading the logs on a weekly rhythm, and re-read the tag-to-script mapping monthly, because traffic mix shifts and rules that fired on day one can go quiet without anyone noticing. A migration that plans its own retreat reads, to buyers, as no migration at all.