The rule. Build the escalation path to a human before launch: explicit triggers, a visible exit, and a conversation that travels with the user.
Why. Ashktorab, Jain, Liao, and Weisz (2019) compared eight chatbot repair strategies and found that deferring to a human ranked below every other strategy when a repair attempt succeeded, but jumped to second place when the repair failed: people want to try with the bot first, then want out fast. Liu et al. (2021) trained handoff prediction on annotated customer service dialogues where a turn counts as transferable not only on an explicit request for a human but also on unsatisfactory answers, negative emotion, and repeated utterances, and they note that a delayed handoff risks user impatience and abandonment. The trigger set is bigger than "agent please," and the timing window is narrow.
Seen in the wild. Intercom's Fin agent escalates by default on a clear request for a human, strong frustration, or a repetitive loop, and lets teams tune, offer, or prevent escalation per scenario in plain language (verified as of September 8, 2026).