When an AI agent should hand off

Handoff is the part of an AI agent rollout that gets written last and matters most. Everything a sales team is afraid of happens on the wrong side of a handoff: a customer who asked for a person three times, a price stated by a machine, a complaint answered cheerfully by something that did not understand the complaint. Write the triggers before the script, and treat them as hard rules rather than preferences.

The five triggers

1. The customer asks

Someone says “let me talk to a person,” “is there someone there?” or “I want to speak to a manager.” The right behavior is immediate. Not one more attempt, not a clarifying question, not an offer to help first. A system that asks “what is this regarding?” before transferring is arguing with the customer, and the customer notices.

This also covers the quieter version. Asking whether they are talking to a real person is usually a request underneath a question. Answer honestly that it is an automated assistant, then offer the person without waiting to be asked again.

2. The conversation is confused

Two signs are enough: the agent has asked the same question twice, or the customer has restated the same thing twice. Both mean the conversation is not converging. A third attempt rarely fixes it and reliably makes the customer feel unheard.

Background noise, a poor connection, an accent the system handles badly, a caller switching between two questions at once, a third person on the line: all of these produce the same failure and the same remedy.

3. The customer is frustrated or complaining

Anything about a bill that was wrong, a start date that slipped, a figure that changed, a person who did not call back. Even when the complaint is not about your company, the emotional register is wrong for an automated conversation, and staying in it makes the outcome worse.

This trigger should be generous. A false positive costs you a transfer. A false negative costs you a customer who tells people your company had a robot handle their complaint.

4. The conversation reaches price

“What would it cost for my situation?” “Can you do better than that?” “What would the monthly premium be with a higher deductible?” An agent can state a published price, because that is a fact. It cannot negotiate, because the next sentence always moves a number.

The refusal has to sound like service rather than evasion. Naming the person who will call, and a time, turns a dead end into a next step.

5. The business would be committing to something

Guaranteeing a price, confirming a start date or what a plan covers, promising an exception, approving a discount, and anything touching credit, contracts, account numbers or a deposit. Some businesses that arrange financing for customers count as financial institutions under the FTC's Safeguards Rule, which calls for an information security program to protect customer information. That alone is a reason to decide deliberately what any automated channel may collect at all.

How the handoff should actually work

Getting the triggers right and the mechanics wrong still produces a bad call. Four rules.

1. Warm beats cold. If a person is available, connect them and pass the context. A blind transfer that forces the customer to start over has undone the reason they were transferred.
2. Carry what was said. Name, number, the question asked and the reason for the handoff. The person should open with “I understand you are asking about adding your daughter to the policy,” not with “how can I help you?”
3. Never drop into a dead end. A transfer into a voicemail box nobody monitors is worse than the missed call you started with, because now the customer thinks they left a message with a person.
4. Say what happens next, then do it. If no one is available, the agent commits to a callback with a named person and a time window, that commitment lands in a queue somebody works, and the callback happens.

What it should say at the moment of handoff

The wording is short, and it is the same every time: acknowledge, name the person, state the timing, confirm the number. Synthetic example: “Of course, let me get Priya for you. She handles renewals. One moment.” When no one is free: “Priya is with a customer right now. She can call you back within the hour at this number. Is that the best one to reach you on?”

That wording avoids three things. It does not apologize at length, which draws attention to the fact that the customer had to ask. It does not ask the customer to explain the reason first. And it does not offer to help one more time before transferring, which is one of the most common ways a handoff turns into a complaint.

The hard case: nobody is there

A voice agent covering missed calls will often run when the office is closed. Handoff still has to mean something at 11 p.m. on a Sunday.

• Say the office is closed. Plainly, with the hours, rather than implying a transfer that cannot happen.
• Capture the name, number and reason. That is the whole job of the call at that hour.
• Commit to a specific morning window, and name who will call if you can.
• Escalate a complaint differently. An angry customer at closing time should create something a manager sees first thing, not a row in a list.
• Offer the urgent path if you have one, for an existing customer with a problem that cannot wait until morning.

The measure of an after-hours handoff is whether the promised call happens. A team that makes the 8:30 call it promised turns a missed call into a relationship. A team that does not has taught the customer that the automated line is where messages go to die.

What the rules cover, and what they do not

Federal law does not lay out a handoff checklist. What it does establish is that these calls are regulated calls. The Telephone Consumer Protection Act at 47 U.S.C. 227 restricts calls made with an automatic telephone dialing system or an artificial or prerecorded voice. The FCC's rules at 47 CFR 64.1200 say an artificial or prerecorded voice message must identify the business responsible at the beginning and give a telephone number during or after the message, and they bar telephone solicitations to a residential subscriber before 8 a.m. or after 9 p.m. at the called party's location. The FTC's Telemarketing Sales Rule sets the same calling window at 16 CFR 310.4. In a declaratory ruling adopted on February 2, 2024, the FCC confirmed that AI technologies that generate human voices fall inside the artificial voice restrictions. States add their own rules on recording, disclosure and automated calling.

In practice, the handoff and the disclosure are connected. A system that identifies itself honestly and transfers on request is far easier to defend than one that meets the letter of the rules and behaves evasively on the phone. Have your own counsel review your setup. This is general information, not legal advice.

Audit the handoffs, not the call volume

The number that tells you whether your rollout is healthy is not how many calls the agent handled. It is what happened at the boundary.

• Handoffs that should have happened and did not, found by listening.
• Handoffs that reached a person, against those that went to voicemail.
• Promised callbacks that actually happened, and how long they took.
• Calls where the customer had to repeat information they had already given.
• Calls where a customer asked for a person more than once.

The last one is the best health check you have. If it is anything other than zero, the trigger is not working, and no other number about the rollout means much until it is.