What a collections voice agent has to get right on the phone

What does a voice AI agent face that a chat agent never does?

A voice agent, sometimes sold as a voicebot or a conversational AI call agent, is software that holds a phone conversation by itself: it listens, works out what the customer meant, decides what to say and says it in a synthetic voice. For banks, lenders and other financial services firms that want collections calls made this way, the phone is far less forgiving than a chat window, and almost every way it is less forgiving can be heard.

In writing, a reply that takes four seconds is fine and a typo can be read past. On the phone, four seconds of silence makes the customer say "hello?" and a misheard number becomes a wrong amount spoken aloud, with no way to take it back. And before any of that, someone has to pick up, which a message never needs.

So the questions that decide whether a voice agent works are mostly about sound and timing, not about the model behind it. You can test most of them with a phone, a list and a few days.

How fast does a voice agent need to reply?

The customer should never have to wonder whether the line dropped. Studies of conversation across many languages find that people typically leave around 200 milliseconds, a fifth of a second, between one person finishing and the other starting. Nobody expects software to match that exactly, but the further a reply drifts past a second, the more the call sounds like a machine thinking, and customers fill the silence by talking over it.

The delay between the customer finishing and the agent starting is called latency. It adds up from several steps: turning speech into text, deciding the reply, turning that reply back into speech, and the phone network itself. Ask a vendor, whether a collections specialist or a general voice AI platform, for the typical figure and the slow end, measured on real phone calls rather than in a browser demo, because the network adds time a demo never shows.

Then listen for how the agent covers the slow moments. A person checking an account says "let me just look at that". An agent that does the same, briefly and only when it is actually waiting on something, sounds natural. One that pads every reply with filler sounds like it's stalling.

Can the customer interrupt it?

They will, constantly. People interrupt to correct a wrong detail, to say they've already paid, or because they've understood and want to move on. The ability to stop speaking when the customer starts is often called barge-in, and an agent without it keeps reading its sentence while the customer says "no, that's not right" underneath.

Test three things. Does it stop within a fraction of a second when you speak over it? Does it then respond to what you said rather than picking its old sentence back up? And does it tell a real interruption from background noise: a cough, a television, a child in the room. It should not stop for every background noise, or the customer can never finish a sentence either.

How should it say amounts, dates and reference numbers?

Say them the way a careful person would. "1,250.50" should come out as "one thousand two hundred and fifty euros and fifty cents", not as a string of digits and not as "one point two five thousand". Dates should be days the customer can place: "this Friday, the 20th" rather than "20/11". Long reference numbers should be read in short groups, with a pause between them, and offered again by message if the customer wants to note them down.

Every figure the customer gives should be read back before a promise to pay is agreed. If they say "I can do fifteen on Friday", the agent should confirm whether that means 15 or 50 euros, and which Friday. Spoken numbers are an easy place for a call to go quietly wrong, because "fifteen" and "fifty" sound alike and both are plausible payments.

What happens when a machine, or the wrong person, picks up?

Many outbound calls reach voicemail, so the agent, or the dialer placing its calls, has to tell a recorded greeting from a live person, which the trade calls answering machine detection. Both mistakes cost something. If the agent takes a person for a machine, it either hangs up or plays a message at someone who said hello. If it takes a machine for a person, it starts a conversation with a recording. Say 1,000 calls in a week are picked up by people and the agent misreads 4 percent of them: 40 customers had the call they answered thrown away.

Decide what a voicemail message says before launch. Anyone in the household may hear it, so a sensible voicemail message is short and neutral: the company's name, a person to ask for and a number to call back, nothing about the account.

The same care applies when a live person answers who isn't the customer. The agent should confirm it has a right party contact, the customer themselves, before mentioning any debt, and when it's a partner or a colleague on the line, it should say only enough to arrange another time.

Will customers answer its calls at all?

Only if the number looks like someone worth answering. A number flagged as possible spam, or one the customer does not recognise, often goes unanswered, and then the quality of the agent behind it does not matter.

Every unanswered call lowers your contact rate, so ask how the numbers the agent calls from are chosen, registered and monitored for spam labels, and whether the number shown is one the customer can call back and reach the same company. Keep the same few numbers for the same customers, and retire any number that starts to collect spam labels instead of calling from it harder.

How do you test AI voice agents for debt collection before launch?

Write the test calls before the vendor runs any. Group them by what can go wrong: the customer who interrupts, the one who gives an amount that could be misheard, the one who answers for someone else, the one on a bad line, the one who asks for a person. Ten categories with 30 calls each gives you 300 calls, enough to see a pattern rather than an anecdote, and a small team can make them in a few days.

Make them on real phones, on mobile networks, from rooms with noise, because that's where your customers are. Score every call against the table above, with a native speaker of each language listening for amounts, place names and surnames. Then repeat the same set after every significant change to the agent, so a fix in one place can't quietly break another.