There is a moment that decides whether an AI agent feels like a colleague or like a badly built phone tree: the handoff.
If the human receiving the call has to ask "sorry, what was this about?", the customer has now told their story twice — once to a machine, once to a person. That repetition is the thing automation was supposed to eliminate, and it is the most common reason people say "I'd rather just talk to someone". A handoff is a data problem disguised as a courtesy. Three things go wrong:
The transcript is passed. A five-minute transcript is not a brief. A person on a phone cannot read it, and skimming one while the customer waits is worse than asking.
Nothing structured is passed. Name and number only. The human rebuilds the case from scratch.
The customer's own words are lost. The summary says "billing issue" when the customer said "you charged me twice and I've been waiting a week". The specificity is the value.
A handoff that works carries five things, in this order:
1. What the customer wants. One sentence, in their terms. "Wants a refund on order 4471 because the item arrived damaged" — not "refund enquiry".
2. Identity, verified or not. Phone number captured, and explicitly whether anything was verified. This matters enormously: an agent that says "verified" when nothing was verified creates a fraud path.
3. What was already promised. If the agent said "we can probably do that", the human needs to know, because it is now a commitment. An agent that promises freely and a human who refuses three minutes later is a broken experience and a justified complaint.
4. What the customer should not be asked again. Their date, their order number, their preferred slot, whatever the agent already collected.
5. The preferred outcome and the constraint. "Wanted Thursday evening, but said any day after 5pm works." The human needs the goal and the real limitation, not just a time.
There is a version of this that goes too far: passing everything, including things the customer did not intend to disclose.
The brief should contain what is needed to continue the conversation, and nothing more. In practice that means: no emotional analysis, no inferred attributes, no speculation about the customer's situation. If the agent guessed, it stays out of the brief.
This is also a legal question in most jurisdictions — but the practical rule is simpler than the legal one: the customer should be comfortable seeing the brief.
The ideal handoff is short and specific:
"Hi, this is Sarah. I can see you're calling about a damaged item on order 4471, and that my colleague said we'd sort a replacement. Can I confirm it's the blue one, and would Thursday after 5 work?"
Everything the customer said is present. They are not repeating themselves. The human's first sentence already demonstrates the system worked.
Compare:
"Hi, how can I help?"
That single sentence undoes every minute the agent spent.
Three delivery points, and you need at least the first two:
If the brief only exists as a spoken sentence at the start of the call, it is gone the moment the call ends. The most damaging handoff failure is not a missing summary. It is a summary that records a promise the business will not keep.
"An agent told me I'd get a refund" is a sentence that has won arguments against real companies. The fix is a rule, not a technology: the agent may state what is true, and may say it will pass a request on. It may not promise an outcome it cannot guarantee.
If your agent can promise refunds, your escalation policy is wrong, not your summary format. Do this once a month:
- Pick a real handoff from the log. 2. Read only the brief, not the transcript. 3. Ask: could I continue this conversation in one sentence, without asking anything already answered?
If the answer is no, the brief is missing items — and the customer will pay for that gap with their time. Every handoff is a moment where automation is either justified or exposed.
A good handoff makes the customer feel that the system is coordinated: someone already knows. A bad one makes them feel processed, and it is the moment they decide to mention it in a review.
The cost of fixing it is a data structure with five fields. The cost of not fixing it is every conversation you automated, undone in the first sentence.
Disclosure: I work on RingBot, an Agent-as-a-Service platform where a handoff to a person carries a structured summary rather than a transcript, and where agents can be restricted from promising outcomes. The five-item brief above is deliberately platform-independent.
Related: when to hand off, five things a transfer needs, missing default route, identity before disclosure.