How an AI collections call hands over to a person

What does a good escalation to a human agent sound like?

When a voice AI agent escalates a call, for example because the customer shows signs of vulnerability or financial difficulty, it should sound calm and specific. It says it's going to pass them to a colleague, says why in a way that doesn't embarrass them ("so someone can look at the options with you properly"), and says what happens next: they'll be connected now, or someone will call at a time they choose. Then the colleague picks up already knowing the story.

The opposite is easy to picture because most people have lived it: a transfer into a hold queue, a person who asks for the account number again, and a customer explaining a redundancy or an illness for the second time in ten minutes. A customer in collections has every reason to give up rather than explain it again.

Should it be a warm transfer, a cold transfer or a callback?

There are three ways a voice AI agent can escalate to human agents, and they differ in what the customer hears. A warm transfer is one where the person taking the call is briefed before or as they connect, so they start informed. A cold transfer passes the call with no briefing at all. With an AI agent, the briefing is usually a written summary on the colleague's screen as the call arrives; some agents also stay on the line briefly to introduce them.

Cold transfers should have almost no place in collections. A callback often does. Many handovers aren't urgent, and a customer who's just said they're struggling may prefer a call at 6pm, after work, to waiting on hold now. What makes a callback work is a named time the customer chose and a person who calls at that time, with the summary in front of them.

What should the person taking over receive?

A summary short enough to read in the seconds before they say hello. It should say who the customer is and whether the agent confirmed it was speaking to the customer themselves, a right party contact, the balance, days past due and anything already agreed, why the agent handed over, and what the customer said in their own words on the parts that matter.

The last item is the easiest to lose. A label such as "hardship" or "vulnerable" tells the colleague a category, and categories invite a scripted response. "I was made redundant in September and I'm waiting for my first benefit payment" tells them what to ask next. Keep quotations short, mark them as the customer's words, and never let the summary guess at a diagnosis the customer didn't give.

Include what the agent already offered and how the customer answered. A colleague who knows a six-month instalment plan was turned down won't open by offering it again, and can move straight to asking what the customer can afford, the affordability assessment, and to the relief options, such as a reduced payment or a pause, that only a person may offer.

How many people do you need for handovers?

Here's an illustration; replace each figure with your own. Say the agent holds 20,000 conversations a month and 6 percent of them hand over: 1,200 handovers. If each takes a person 12 minutes, including reading the summary and writing the outcome, that's 240 hours of work a month.

Spread across 21 working days of 8 hours, that's about 1.4 people busy at any moment, on average. Busy time isn't headcount, though: allow for breaks, leave and training, and for someone free to take a warm transfer at once, which usually means two people at minimum. Handovers also follow the agent's calls, so if those cluster in the early evening the handovers will too. Plan for two or three people at the busiest hours and book callbacks into the quieter ones.

Outside working hours, the agent should never pretend a transfer is possible. It should say when a person will be available and offer a callback then. Telling the customer when someone will be free is better than leaving them on hold.

How do you know the handovers work?

Watch a few numbers every week: the AI agent handover rate broken down by each reason in your escalation matrix, the list of situations that trigger a human handoff, so a sudden jump in one reason stands out; the time from handover to a person speaking; the share of callbacks made at the promised time; and the share of handed-over customers who end with a plan, a resolved dispute or a referral rather than nothing.

Then listen. Regular call reviews should look for two failures that no dashboard counts on its own: conversations that should have been handed over and weren't, and customers who had to repeat themselves after the transfer. The first is the more serious, because the customer who needed a person never got one. The second means the summary isn't doing its job.

Too many handovers is a problem too, if a smaller one. If most handed-over calls end with the colleague agreeing something the agent was allowed to agree, the triggers are set too wide, and the team is doing work the agent could have done.