A good collections conversation is short because it is specific, not because it cuts corners. In banking collections, the cases that ended with a promise to pay averaged 116.5 seconds of handling time. In under two minutes, the customer learns who is calling, proves who they are, hears what is overdue, chooses a date and knows how to pay. This piece takes that conversation apart and explains the design choices behind it.

116.5 sAverage handling time (AHT)Average duration of conversations in cases that ended with a promise to pay (PTP).Cases ending with a promise to pay, banking, historical. AHT is not the outcome.

What fits into 116 seconds#

The structure of an early-collections call is fixed in its order and flexible in its wording. The order is not a matter of style. In banking it is a matter of rules: disclosure first, identity before any detail of the debt, and a business status at the end that the bank can act on.

  1. Disclosure and purpose. The customer hears that they are speaking with an AI calling on behalf of their bank. We have always said so; since 2 August 2026 Article 50 of the AI Act makes it an explicit transparency obligation.
  2. Identity verification. Configured per market and product. Until it succeeds, no detail of the debt is spoken, and that rule is enforced by the platform rather than written into a script.
  3. Payment status. Has the customer already paid? Many have, and a short, courteous confirmation ends the call there.
  4. The fact. Which product, how much is overdue, stated once and plainly.
  5. The commitment. A date within the horizon the bank allows for a promise to pay, chosen by the customer.
  6. How to pay. The method and reference, repeated afterwards in a text where the bank has agreed it.
  7. Status. The case returns to the bank as committed, with the date, so that tomorrow’s attempts can be suppressed.

Seven steps in under two minutes sounds tight. It is not, provided each step does one thing and the conversation does not wander. Most of the time a collections call loses is spent on things the customer did not need: long preambles, repeated legal wording, consequences recited to someone who has already said they will pay.

Why short respects the customer#

A customer who is a few days late on a loan instalment rarely wants a long conversation about it. Some are embarrassed, many are busy, most simply forgot. What they want is to understand quickly what this is, settle it and get on with their day. A call that respects that is more likely to be answered next time and more likely to end in a commitment the customer means.

Short is also clearer. When the call contains only what matters, the customer hears the amount and the date without noise around them. A long call tends to bury the one thing the customer needs to remember. And the longer a call runs, the more room there is for the conversation to drift from ‘when will you pay’ to a general negotiation that neither side planned.

The opening seconds decide whether there is a call at all#

Before a conversation can be short and successful, it has to survive its opening. In our outbound collections processes, a meaningful share of customers end the call within seconds of hearing who is calling. That is the second rung of the way we measure contact: after reach comes continuation, the share of calls that go on past the introduction. A design that saves ten seconds in the middle but loses customers in the first five has made the wrong trade.

So the opening gets more attention than any other part of the call. The order in which the bank’s name, the purpose and the AI disclosure are spoken; the pace; the voice. We learned early that the voice matters: our first outbound campaign only began to convert when synthesis was replaced by a recorded studio voice. Today we choose voices per process and language, and test them, because the cost of a weak opening is paid on every call that follows.

Specific beats soft#

The single biggest difference between a short call that works and a short call that fails is the specificity of the question. A vague question about whether the customer will be able to pay invites a vague answer, which is neither a commitment nor a reason to stop calling. A question that asks for a decision makes the next step concrete. Offering a choice within the bank’s allowed window gives the customer control without opening an unbounded negotiation.

The language around the question changes with the situation. Days past due shape the base conversation, and behaviour history adjusts it again, including how a broken previous promise is handled, calmly and directly. Different customer groups have their own dialogue paths, running in parallel in the same deployment.

None of this is improvised by a generative model in the middle of a call. In regulated processes, the replies come from a closed, versioned library approved by the bank; models recognise what the customer means, they do not write what the bank says. That is how a short call stays both natural and auditable.

When a call should take longer#

A short average is not a reason to rush everyone. Some calls should be longer, and the design gives them the room. The exceptions are where a collections conversation earns or loses trust, and each has its own path agreed with the bank.

Situations that deliberately leave the short path
SituationWhat the conversation does
Customer describes financial hardshipRecognises it, stops pushing for a date, routes to the bank’s specialist team or agreed referral
Customer disputes the debtRecords the dispute, pauses collection contact, routes to the right team
Customer asks for a personTransfers or arranges a callback under the agreed rules
Verification failsEnds without disclosing any detail of the debt
Customer withdraws consent to recordingFollows the market-specific stop rule
Death or legal proceedings mentionedHard stop and referral, configured per market

Exception paths from Vee’s early-collections process design; configuration differs by market and client.

These calls pull the average up, and they should. A supplier that drove AHT down by shortening them would be optimising the wrong thing.

AHT is a diagnostic, not the outcome#

We publish the 116.5-second figure to show what a well-designed conversation looks like, not to suggest that shorter is always better. Lower AHT is not inherently good. It has to be read alongside reach, the share of conversations that continue past the introduction, the share that reach a commitment, and, above all, whether the customers paid.

Consider two calls. One lasts forty seconds: the customer is verified, says the instalment went out this morning, and is thanked. The other lasts four minutes, ends in a promise, and the promise is broken. On a minute report the second looks like the harder-working call. On the bank’s ledger the first is the better outcome. AHT cannot tell them apart; only the payment data can.

This is also why per-minute billing sits so badly with collections. If the supplier earns more when calls are longer, the commercial model pulls against the design described here. We have moved new agreements away from minutes for exactly that reason. Our own operating data shows the direction: since 2026, our conversation minutes have fallen faster than the number of cases handled, because of shorter calls, a different campaign mix and our foreign processes.

Building a voicebot is not the same as serving a customer. Between the two lies a very long road. – Dawid Wójcicki, Founder & CEO of Vee

How a conversation gets this short#

Nobody writes a 116-second call on the first attempt. It comes from listening and iterating. Every day, in every project, people at Vee listen to selected calls, especially the unusual ones, to hear where customers hesitate, where a phrase confused them, where speech recognition struggled on a noisy line. Analysts look for correlations between wording, timing, voice and outcome.

Changes are tested with the client’s consent, on a large enough base to mean something: whether the change concerns the opening, the wording, the voice or the verification step. Optimisation cycles run in days, not weeks, and once a change is approved it applies to every live conversation at once. There is no team of agents to retrain and no drift between the best and the worst performer on a shift.

What happens after the 116 seconds#

The call ends; the work does not. Where the bank has agreed it, a text follows with the amount, the date and the payment reference, so the customer does not have to remember them. The case returns to the bank with its status the same day, and further attempts follow the rules agreed with the bank. If the payment arrives, the case is closed. If it does not, the bank’s allocation strategy decides whether the contract comes back, and when it does, the next conversation knows about the broken promise.

Questions worth asking about any collections conversation#

  • What is the average duration of calls that end in a promise to pay, and of those that do not?
  • Is identity verification enforced before any debt detail, by the system rather than by the script?
  • Does the conversation ask for a specific date within our allowed window?
  • How does the dialogue change with days past due and with a broken previous promise?
  • Which situations leave the short path, and where do they go?
  • How are wording changes tested, approved and rolled out?
  • Are we paying for minutes, or for the work and the result?

A short call is the visible end of a lot of invisible design. If your current collections calls are long and your promises are soft, the fix is rarely a better voice. It is a clearer conversation, a stricter structure and a commercial model that rewards both. That is the work we do, every day, for the banks we serve.

Follow The First Second

Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.

Follow via RSS