Your customer support agent is mid-conversation with a frustrated customer about a delayed shipment. The agent has just finished looking up the order and identified that a refund-and-replace is straightforward and approved by policy. The customer's next message arrives: "I'm done with this. Get me a human."
The agent has the resolution ready. It would take one tool call to execute.
Should the agent surface the resolution first, or escalate immediately?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
When a customer explicitly requests a human, that request takes precedence over the agent's resolution capability. Overriding it — even to deliver a good outcome — communicates that the customer's stated preference is conditional on the agent's judgment about what's best for them. That damages trust in a way no quick fix recovers. The principle: explicit customer requests are not heuristics to be optimized against; they're constraints to be honored. The cost of an "unnecessary" escalation is much smaller than the cost of teaching customers their requests will be overridden when the agent thinks it knows better.
This is the genuinely tempting answer because it sounds reasonable — the resolution exists, the customer wants the underlying problem solved, why add latency? The misconception is that the customer's stated preference (a human) and their underlying preference (a solved problem) can be cleanly separated. They can't. By the time a customer says "get me a human," they've often lost trust in the automated path entirely; even a correct resolution at this point may register as the agent dismissing their request.
Offering a choice sounds neutral but functionally still asks the frustrated customer to do more work — read, decide, respond — when they've just told you they're done with that. "Give them the choice" is reasonable in many contexts but not after an explicit demand for escalation.
Applying the resolution silently while transferring sounds like the best of both worlds, but it ignores the customer's stated preference (they asked for a human, not a resolution + a human) and creates confusion when the human agent picks up a case that's already been acted on. Hidden actions during escalation are a reliability hazard.