The Handoff: The Most Important Thing an AI Agent Does

The feature that decides whether an AI agent is an asset or a liability is not how well it answers. It is how reliably it stops.

An agent with no handoff condition does not fail by going silent — it fails by continuing. It answers a refund question with a plausible policy you do not have, reassures an angry customer in a tone that makes it worse, or confirms a delivery date it cannot see. Each of those is more expensive than not answering at all.

Six conditions that should always hand over

1. It cannot find the answer in your material. The most common and the most important. “I do not have that — let me get someone” is a correct, professional answer.
2. Money outside the published price. Refunds, disputes, discounts, custom quotes. An agent improvising here creates commitments you did not make.
3. A complaint, or any sign of anger. Not because an agent cannot write an apology, but because the customer wants to be heard by a person, and a smooth automated response reads as being handled.
4. Anything with a safety or legal dimension. Allergies, medical questions, anything involving a minor, anything a lawyer would want to see first.
5. Something it cannot verify. Stock levels, appointment availability, the status of a specific order. If the agent cannot look it up, it must not assert it.
6. The customer asks. “Can I talk to a person” is a complete instruction. An agent that tries one more time before complying is the single most irritating pattern in the category.

What does a good handover look like from the customer’s side?

Three properties, and they are all about honesty:

• It is explicit. “I will get a colleague to answer this” — not a silent change of tone that the customer is supposed to notice.
• It gives a time. “Within the hour” if you are open, “first thing tomorrow” if you are not. Vague is indistinguishable from being ignored.
• It does not make them repeat themselves. The worst handover is the one where a human opens with “hi, how can I help?” after the customer explained everything twice.

What does a good handover look like from your side?

The technical part matters as much as the wording, and this is where most implementations quietly fail.

A handover is only real if a person actually receives it. Three things have to be true:

1. Somebody is notified in a place they look — not a dashboard counter that nobody opens.
2. The conversation is marked as waiting for a human, so it does not sit in a list of two hundred looking like everything else.
3. The agent stops. If the agent keeps answering after handing over, two voices reply to the same person and the customer concludes nobody is in charge.

That third point is a real implementation detail and worth testing deliberately: hand over, then send another message as the customer, and see whether the agent speaks again. The inbox mechanics that make this work are in how a shared Instagram inbox works.

Should the agent hand over the moment it is unsure?

No, and this is the tuning question that decides whether the system is useful.

An agent that hands over at the first ambiguity produces no saving at all — you have added a step between the customer and the person. An agent that never hands over produces invented answers. The right setting is in between, and it is found by reading conversations rather than by picking a number.

A useful middle behaviour: ask one clarifying question first. Many conversations that look unanswerable are just underspecified, and one question resolves them. Two clarifying questions is usually one too many — at that point a person is faster.

What should happen out of hours?

This is where handoff design gets genuinely interesting, because “get a human” is not available at 2am.

The wrong answer is for the agent to stretch and answer anyway, because the pressure to be helpful is exactly when invention happens. The right answer has two parts: answer everything it legitimately can from your material, and for the rest, say plainly that a person will reply and when.

Do not promise “shortly”. Say the hour you open. A customer who knows they will hear at 9:00 waits; a customer told “soon” at 2am has been told nothing.

How do you know the handoff threshold is right?

Two numbers, read together, and neither is meaningful alone:

• Share of conversations handed over. If it is very high, the knowledge base has gaps — the fix is material, not tuning. The list of handed-over conversations tells you exactly which gaps.
• Share of handled conversations where the customer came back unhappy or asked again. If this is high, the agent is answering things it should not.

The goal is not to minimise handovers. It is to make sure every handover was necessary and every non-handover was correct. Getting there is mostly about the material: see training an agent on your own knowledge base.

The instruction to write down

If you write one behavioural rule for your agent, write this one:

If the answer is not in the material you were given, say that you do not have it and offer to get a colleague. Do not guess, do not approximate, and do not answer from general knowledge.

It is the difference between an agent that is occasionally unhelpful and an agent that is occasionally wrong, and those two failure modes cost very different amounts.

The distinction between an agent and a rule-based bot is in AI agent versus chatbot, and the Instagram-specific setup is in an AI chatbot for Instagram DMs. Chatonaut’s AI agent hands over on conditions you set, and the conversation is marked for a person rather than left in a list. Plans are on the pricing page.