Group chats changed where support happens. On WeCom — WeChat Work, the messaging platform a large share of Chinese brands run service on — one customer group can hold hundreds of buyers, and every unanswered "does this come in size L" is visible to all of them. The usual reaction is to drop an auto-reply bot into the group. The bot then answers everything, in arrival order, with the same voice. The result is not automation; it is noise at scale. The step that gets skipped is routing: deciding who or what handles each message before anyone writes a reply.
Arrival order is the wrong default
A naive bot consumes the queue from the top and produces one reply per message. Three failure modes follow. The bot answers a question the human agent already answered thirty seconds earlier, because they were working in parallel. It interrupts a consultation that was clearly headed somewhere the bot cannot go. And it keeps replying after the customer has stopped talking to it — the human took over, and now two voices alternate in the same thread. Buyers notice. The fix is not a smarter bot; it is a routing layer in front of the bot.
A three-tier routing table
Most groups need only three tiers, and the boundaries are about risk, not topic.
Tier 1 — factual and repeatable. Stock, shipping windows, price, return policy. These have stable answers and low blast radius when phrased carefully. This tier goes to templates, and matching should require a confident keyword or intent hit — a half-match on "when will my order arrive" should fall through, not guess.
Tier 2 — consultation. "Which model fits my kitchen?" has no stable template, but silence is worse than an acknowledgment. The routing rule here is a split response: the automation confirms receipt and promises a person, then stands down while a human answers. What it must never do is keep answering alongside the person.
Tier 3 — emotion and disputes. Complaints, refund anger, anything with an exclamation mark and a history. The route is a human from the start. Automation's only job is speed of triage: tag it, surface it, and get out of the way.
Silence is a routing rule too
The most underused setting in group automation is suppression. After a staff member speaks in a thread, the bot should stand down for that thread for a while — buyers can tell when a person and a bot take turns over them, and it reads as being passed around. A workable default: once a human has replied in a conversation, the bot only re-engages on a high-confidence Tier 1 match, and never on follow-up questions in the same thread. This one rule eliminates most of the "two voices" complaint in group support.
Routing is configuration, not code
None of this requires custom engineering. What it requires is writing the table down before buying anything: message type, tier, who responds, what suppresses the bot, what re-arms it. Most team conflicts about automation ("it interrupted me again") are really routing bugs, and routing bugs are fixed in a table, not in prompts or code. Teams that skip the table end up turning the bot off entirely — usually in week three, usually right after the double-voice incident.
References: a worked example of these rules applied to WeCom customer groups lives in this group message routing guide, and the broader setup context is in this WeChat auto-reply guide.
Top comments (0)