Handover: the moment your AI agent should stop talking

Almost every disappointing AI deployment fails in the same place, and it is not the part people test. The agent handles the straightforward majority perfectly well. Then something falls outside its scope, and a confused customer arrives in front of a colleague who has no idea what was already said.

The answers were fine. The seam was the problem.

Handover is a feature, not a failure

Teams treat escalation as the thing that happens when the AI loses, which leads to prompts that try to avoid it. That is exactly backwards. An agent that hands over cleanly at the right moment is doing its job; one that grinds on for four more turns to avoid handing over is not.

Design for it deliberately and it stops being a defeat.

Name the conditions, do not infer them

"Escalate when appropriate" pushes the entire decision into a word that means nothing. Write the list instead. Five or six named situations, each saying where the conversation goes:

They asked for a person

Immediately. No "I can help with that first". This is the one people most often get wrong, and it is the one customers resent most.

They are upset

To a human, with the transcript. Do not defend, do not explain policy.

Money is being negotiated

Refunds, discounts, contract terms. An agent that can be talked into a price is a liability.

Two turns with no progress

A simple, brutal counter that catches most of what your named conditions miss.

It is outside scope

Your refusal list, from the prompt.

Confidence is low

Useful as a backstop, useless as your only rule.

The three failure modes

1. Handover into a void. The agent says a person will take over, and nobody does, because it happened at 11pm or because "a person" was not a specific person. If there is nobody to hand to, the honest handover is a booked callback — not a promise.
2. Handover without context. The colleague gets a chat with no history and asks the customer to start again. This is the one customers actually complain about.
3. Handover that isn't one. A "mode" gets toggled somewhere, the bot is technically off, but the chat is still nobody's. It sits until someone notices.

What makes it clean

Three properties, and they are structural rather than clever:

• Handing over means assigning the conversation to a named person. Not a mode, not a flag. If it is assignment, the chat is always somebody's and "unassigned" is a state you can filter for and act on.
• The whole thread survives. Same conversation, same record — the person picking up scrolls up rather than starting over.
• The context is already there. The contact record beside the thread: the last call and how it went, the note a colleague left, the open task, the lead stage.

In Convarza AI a bot-handled chat is simply an unassigned one, marked with an AI badge. Assigning it takes the bot out and puts a person in — there is no third state where both or neither is responsible. How that works in the inbox.

If your handover is a mode rather than an assignment, you will eventually have a conversation that belongs to nobody. It is only a question of volume.

What the agent should say

Short, specific, and no apology theatre:

That one needs a person. I've passed this to the team with everything we've discussed — someone will pick it up shortly.

Two things matter there: it names what happens next, and it says the context went with it. "I'll connect you to an agent" says neither, and the customer braces for repeating themselves.

Measure the seam

• Time from handover to first human reply. The number that decides whether the handover felt good.
• Handovers with no reply for an hour. Your void detector.
• Customers who repeat themselves after handover. Read ten of these a week; each one is a context failure with a specific cause.

Fix the seam and the same agent, with the same prompt, becomes noticeably better — because most of what people disliked was never the answers.