Three months of an AI support bot: two roles saved, complaints doubled
One retailer let us see the numbers. The savings are real, and so is the price.
The retailer had six support staff. Three months ago they put an AI bot on the front line, and they now have four. Payroll is down about NT$ 90,000 a month.
Over the same period, tickets escalated to a human went from 180 a month to 370.
The problem was not wrong answers. It was fluent ones
Their read: the bot almost never says it does not know. Faced with something it cannot handle, it produces an answer that sounds reasonable and solves nothing. The customer follows it, fails, and comes back — and by then the mood has changed.
They changed one thing: the talk-to-a-human button moved from the bottom of the conversation to the top, and the bot now volunteers “I might get this wrong, shall I pass you to a person” when confidence is low. Complaints came back down to 240 a month.
{"q":"Our first KPI was bot resolution rate. That KPI was pushing it not to hand over.","by":"Their support manager"}
The conclusion is not don’t use one. It is don’t make “avoid escalation” the metric. Saving two roles assumes the remaining four handle harder cases, and their training cost belongs in the maths.
