What AI collections is, and what it does on a call

What is AI collections, and what isn't it?

AI collections is software that holds the collections conversation with the customer, on the phone or in writing, inside limits the lender sets. When the customer asks why this month's amount is higher, it answers. When they say they can pay half on Friday, it checks your rules and either agrees, offers what the rules do allow, or passes the call to a colleague. The trade also calls this conversational AI collections, meaning software that understands free replies and responds to them, as opposed to anything that only dials or sends.

Two other things are sold under nearly the same words, and they are easy to confuse. Collections automation is the older layer: the dialer that places calls in bulk, the queues that decide which account a collector sees next, the templated SMS that goes out on day 3. It moves work around and starts contact, but every reply that isn't a payment lands on a person's desk. A collections platform, also called the system of record, is where the accounts live: balances, days past due, the strategy each account sits in, every promise to pay (a customer's commitment to pay a set amount on a set date) and whether it was kept.

Some AI debt collection software is sold as an AI collections platform, which can mean an agent with a light record attached or a record with a chatbot bolted on. The quickest way to sort a pitch is to ask what each piece leaves to your people.

What does an AI collections agent do on a real call?

In outline, a call runs the way a good collector would run it. The agent makes sure it is talking to the right person before it mentions any debt, and says nothing about the account to anyone else who picks up. It says it is an AI and who it is calling for. It explains what is owed, asks what happened and listens, because the answer decides what kind of conversation this is. Voice AI for debt collection follows the same course as a written chat; only the medium changes.

From there it can only agree what the lender has already approved, such as a new date or a short payment plan. Anything outside those limits, any sign that the customer is struggling or vulnerable, any dispute and any request to speak to a person goes to a colleague. Whatever happens, the outcome is written down where the team can see it, so nobody has to ask the customer to start again.

That is also how to tell real AI collections from scripted automation in a demo, where the vendor's call tends to go perfectly. Take it off the script: interrupt it, answer a simple question with a long story, propose something it was not expecting, or mention a hard personal situation and see whether it slows down and offers a person. A decision tree with a synthetic voice falls apart at the first answer it was not built for. Guardrails, evidence and price belong to the vendor meeting, which has its own list of questions.

What changes when every account can get a conversation?

Capacity stops being a headcount question. Today the size of the team decides how much of the book hears from you, and the collections strategy is written around that ceiling: call the first days of arrears as hard as you can, send letters after. When a conversation costs a fraction of a person's time, the ceiling moves. Which accounts get worked, and how often, becomes a decision someone has to make on purpose.

Here is an illustration, with invented numbers to replace with your own. A consumer lender sees 4,500 accounts a month fall into arrears and keeps them in its early collections team until 60 days past due. It has 10 collectors, each with about 5.5 hours a day on the phone, 21 days a month. Working an account properly, with the unanswered dials, the conversation and the notes, averages 15 minutes. That is 4 accounts an hour per collector, 220 a day across the team, about 4,620 a month: roughly one proper attempt per new account and very little after it. At 3,500 euros a month per collector, fully loaded, each worked account costs about 7.60 euros.

Now suppose the strategy you'd actually want is five touches per account across those 60 days, at around day 3, 10, 20, 35 and 50. That's 22,500 worked accounts a month. With people alone it takes nearly 49 collectors, about 171,500 euros a month, so few teams do it, and the accounts between day 20 and day 60, the ones closest to default, get a letter. Assume, purely for the sum, that an agent costs 2 euros per worked account all in: the same plan then costs 45,000 euros. Whether that pays depends on how many more accounts cure, meaning return to fully up to date, and the cure-rate sum for that is in the guide to choosing debt collection software. Run it with your own roll rates, the share of accounts that slide into the next arrears band each month.

The harder change is that someone now owns the choices the ceiling used to make. Do days 1 to 5 get a message only, because many of those customers pay unprompted? Does a broken promise go back to the agent or straight to a person? What do the ten collectors do now? The sensible answer to the last one is the work that was always being squeezed out: hardship cases, complaints, disputes, larger balances, and the customers who asked for a human.

What does AI collections not do?

It doesn't set policy. The agent can offer a three-month split because someone in your credit risk or collections team decided three months was right for that product and signed it off. It can't decide that this customer deserves four, however good the reason sounds. The limits are configuration, and an agent that can be talked past them is a defect to report, whatever the vendor calls it.

It doesn't replace judgement on hardship, vulnerability or disputes. A customer who has lost their job, who sounds confused or distressed, who mentions an illness or a bereavement, or who says they never took the loan needs a person who can weigh circumstances and step outside a menu. The revised Consumer Credit Directive, which member states must apply from 20 November 2026, requires lenders to exercise, where appropriate, reasonable forbearance before enforcement, taking individual circumstances into account. The agent can gather the facts in the customer's own words and pass them on. Weighing them is a person's job.

That's also where the argument over AI vs human collectors usually goes wrong. On many collections floors much of the volume is repetitive: the first reminder, the rescheduled date, the confirmation, the follow-up after a missed promise. The agent takes that. People keep the conversations where judgement is the work, so a team handles fewer conversations and each one is harder.

What are the rules for AI debt collection in Europe?

For debt collection, the EU AI Act's main demand today is transparency. Article 50, which has applied since 2 August 2026, requires AI systems that interact directly with people to be designed so that the person is told they're dealing with an AI, clearly and at the latest at the first interaction. There's an exception where it's obvious from the context, and a natural-sounding voice on a phone call is the opposite of obvious, so the disclosure belongs in the opening line. Formally the duty sits with the provider, the company that builds the system. It's still your customer and your complaint if it's missed, so listen for it in the first seconds of real recorded calls.

The Act lists AI used to evaluate people's creditworthiness or set their credit score as high-risk, with much heavier obligations. A collections conversation isn't normally that. But if the affordability answers an agent collects feed a model that scores customers for credit decisions, you're closer to that category, and that's a question for counsel before go-live rather than after.

The GDPR covers every recording and transcript. Tell the customer the call is recorded and why, have a lawful basis for it, and keep it no longer than your retention policy allows. Calls about arrears often carry more than a balance: customers mention illness, and health information is a special category of data with stricter conditions. Article 22 of the GDPR also gives people the right not to be subject to decisions based solely on automated processing that significantly affect them. An agent refusing a payment plan with no person involved could come close, which is one more reason refusals and anything outside the limits should route to a colleague.

Everything else is the rulebook you already follow. Contact hours, contact frequency, who you may speak to and what you may say about consequences apply to an AI exactly as they apply to a human collector. Some national rules go further: Italy's privacy authority bars pre-recorded collection calls made with no operator, and how that applies to an AI voice agent is a question for Italian counsel before go-live. Where a credit servicer works the loans, the EU Credit Servicers Directive requires dealing with borrowers in good faith, fairly and professionally, with information that isn't misleading and without harassment, coercion or undue influence. In practice, AI collections compliance in Europe is that rulebook enforced on every call, plus the duty to say what's calling. A rule-breaking line repeated across a whole book is a breach on every call it was said in.