Tell an AI agent to escalate “refund requests” and it will still answer some of them, because a category isn’t something the model can match against — a sentence is. The fix is writing the trigger as the literal phrases customers actually type, then routing each case by how urgent it is, not by one rule for every refund.
Why “escalate refunds” is not a rule
A refund request in Thai chat rarely arrives as the word “refund.” It shows up as “ขอคืนเงิน,” “คืนเงินได้ไหมคะ,” “อยากได้เงินคืน,” “ยกเลิกออเดอร์ ขอเงินคืนด้วย,” or a customer typing the English word on its own. An escalation rule built around a single label or a single keyword catches whichever phrasing the person who wrote the rule happened to think of, and misses the rest. The AI answers a policy question instead of stopping, and the team’s first sign of trouble is a chargeback or a public review, not a queued conversation.
Write the trigger with your customers’ own words
The fix is mechanical, not a smarter model. Pull five to ten real messages from your own chat history that were actually refund requests, and list them — not the single word “refund” — as a “contains any of” condition: ขอคืนเงิน, คืนเงิน, เงินคืน, ยกเลิกออเดอร์, refund. Covering the phrasing your customers actually use gets you further than tuning a confidence threshold, because the model isn’t unsure about these messages — it just was never told they counted. Treat the list as maintained, not finished: add a phrase every time a refund conversation slips past the rule and gets answered instead of handed off.
Route the case by what happened, not by the word “refund”
Not every refund message needs the same handling. Whether the order has already shipped, and whether the customer is asking calmly or has already said they’re unhappy, changes who should pick it up. A cancellation before an order ships is close to an ordinary order change, so routing it to whichever agent is free next works fine. A request that arrives after the parcel has gone out usually involves a return or a dispute over the item’s condition, and holds together better with one specific agent who can see it through, rather than being redistributed mid-conversation. A refund request that shows up next to a complaint — repeated messages, a mention of posting about it, a message in capitals — should escalate the same way your team already handles other urgent approvals, whether that’s a named person or the same group chat used for other sign-offs.
Test it against real conversations before switching it on
A rule that has never been tried against your own customers’ phrasing is a guess. Before it goes live, run every phrase you collected through a simulated conversation and confirm the agent stops and hands off each time — not only for the wording the rule was written around. If any of them still gets an AI-composed answer instead of a handoff, the rule is too narrow, and the fix is adding that exact phrase, not lowering a confidence setting elsewhere. Retest after every change to the rule, the same way you’d retest any handoff trigger before trusting it with a live customer.
What the agent handling it needs to see immediately
A refund case that lands with a human still costs the goodwill a fast handoff bought if that person has to ask the customer to explain everything again. Carry three things with every handoff: the phrase that triggered it, the order’s shipping status, and whatever payment or slip status is already on file. They’re the same three things that make a routed cart finishable in one reply instead of a fresh conversation. A customer who has already explained their order once should never have to explain it again just because the case changed hands.
Start with the refund messages sitting in your chat history right now. Five real phrases, tested once, beat a rule that only recognizes the word you’d have used yourself.
