Loading
Low confidence, where the assistant is unsure. Explicit request, where the customer asks. And topic, where you have decided in advance that a person must handle it — complaints, cancellations, anything clinical or financial.
The third is a configured boundary rather than a judgement, which is what makes it reliable.
The cost asymmetry is stark. Handing over one conversation that could have been resolved costs a few minutes of agent time; failing to hand over one that needed it costs the relationship.
So set thresholds conservatively. Deflection lost to caution is cheap, and the alternative is not.
On low confidence, on request, and on any topic you have listed. Early, in all cases.
The full conversation, the customer record, and what the assistant already tried.
Say so with a realistic expectation. Never loop them back to the bot.
Half an hour with your own data usually saves reading three of these. The guides will still be here afterwards.