Human handoff: the most important moment in an AI assistant

We have all been there: talking to a support bot, the problem isn't solved, and the bot keeps offering the same three options. At that moment the problem isn't the AI. Nobody designed for "I can't".

In building assistants, the handoff to a person matters as much as the quality of the answers.

Why this moment decides everything

An AI assistant is excellent at common cases. Trust, however, is made or broken in the hard ones: a complaint, an exception, a sensitive decision. Handle those well and the customer trusts the whole system. Handle them badly and every good answer before it is forgotten.

When to hand over

These triggers belong in the design:

• Low confidence. The answer isn't in the available knowledge, or the sources conflict.
• Dissatisfaction. The tone has sharpened, or the customer has repeated a question several times.
• An explicit request. The customer asks for a person. This must always be honoured, without resistance.
• Sensitive topics. Financial, legal, health — anywhere an error is costly.
• Out of authority. Something the assistant isn't allowed to do, such as a price exception.
• A conversation going nowhere. Several turns without progress.

How to hand over

1. Be transparent

The customer should know what happened: "I'm passing this to a colleague." Hiding that they were talking to an AI assistant destroys trust.

2. Send a summary with it

The worst experience is explaining everything again to the operator. The assistant should hand over a summary: what the customer wanted, what was said, and why it was escalated.

3. Manage the wait

If no operator is available right now, say so honestly: when they will get a reply and through which channel. A promise of "right away" that isn't kept is worse than silence.

Common mistakes

• Hiding the human option. Making operators hard to reach to reduce support load. The result is an angry customer, not a lighter load.
• Handoff without context. The conversation moves, but the operator starts from zero.
• Too much handoff. An assistant that escalates every simple question isn't doing anything.
• No out-of-hours path. It must be clear what happens at night and on holidays.

Tuning the threshold

How early the assistant escalates is a product decision, not a technical one. In a store with expensive items the threshold should be lower: you want a person involved sooner. For simple, frequent questions the assistant can be more independent.

Tune it with data: which conversations were handed over, and which should have been but weren't. That is part of evaluating assistant quality.

In practice

In Pasokhinoo, handoff was part of the design from the start, not a later patch: wherever confidence is insufficient or the customer is unhappy, the conversation reaches an operator. Rahro, which deals with relationship wellbeing, applies the same thinking differently: the AI coach doesn't replace a professional, and the user is guided to a human counsellor when that is needed.

The takeaway

An AI assistant isn't judged only by its right answers. It is judged by what it does when it has none. A transparent handoff, with a summary and a realistic expectation, turns the assistant from a tool into a colleague people trust.

Related: AI agent vs chatbot.