How a bank compares AI debt collection platforms

Why do lists of the best AI debt collection platforms tell a bank so little?

Ask an AI assistant for the best AI platforms for automating debt collections at banks and it will hand you a list of ten or fifteen names, much the same list whether you asked for AI debt collection software or for debt collection software for banks. Follow the links and many of those lists were written by one of the platforms on them, or by a marketplace paid to carry them. They rank AI agents for debt collection, voice bots and whole collections systems against each other as if a bank could swap one for another, and describe every product with the same five adjectives.

A bank does not need a ranking. It needs to know which shape of product fits the stack it already runs, which criteria separate the serious platforms from the polished ones, and how to put a number on each so three people in a room can agree. The lists are useful for collecting names. Everything after that is your own work.

What are the four shapes of AI collections platform?

Behind the lists sit four kinds of product, and the first thing to settle is which kind you are comparing. A collections specialist AI agent is software built to hold the collections conversation itself, by phone and message, with identity checks, required wording and payment limits already designed in. A general voice AI platform, the kind listed under conversational AI platforms for financial services, supplies the voice, the phone lines and a language model, and leaves the collections behaviour for your team to write and test. An agency that runs AI sells you the service rather than the software: its collectors and its agents work your accounts for a commission, and you never operate the tool. And the debt collection software banks already own, the case-management systems that hold every account and strategy, increasingly sell an AI module that sends messages, and in some cases takes calls, from inside the system of record.

Each shape answers the bank's question differently. Each shape makes a different trade: the specialist trades control for speed, the platform trades speed for control, the agency trades a commission for operating work, and the module trades pace for keeping everything in one system. Compare like with like, and when two shapes make the shortlist, say which trade you are making.

What should lenders compare, criterion by criterion?

Nine criteria cover what separates platforms. Give each a weight before you look at any vendor, so the weights reflect your book rather than the last demo. A bank with 40 percent of customers in a second language weights languages high; a bank whose contact centre already handles handovers well weights that lower. Five of the nine, the limits, the evidence, the languages, the handover and the write-back, can be checked before any pilot with a test call and a sample export. The table shows one lender's weights, out of 100, as an example to replace with your own.

How do you score three platforms side by side?

Score each criterion from 1 to 5 on evidence you have seen, never on the brochure, multiply by the weight and add up. Here is an illustration with invented figures, three platforms compared on the same nine criteria. A bank shortlists three platforms for its card accounts in the first weeks of arrears: a collections specialist, a general voice platform its own engineers like, and the AI module its collections system vendor has just released.

The specialist scores 5 on limits, evidence and handover, 4 on languages and channels, 4 on write-back, 3 on reach, 3 on price and 5 on the pilot: 425 points out of 500. The general voice platform scores 5 on price and channels and 4 on languages, but 2 on limits and evidence because nothing exists until the bank builds it, 3 on handover, 2 on write-back, 2 on reach and 3 on the pilot: 315 points. The module scores 5 on write-back and reach, 4 on evidence, 3 on limits and languages, 2 on channels because this one only sends messages, 3 on handover, 3 on price and 2 on the pilot: 330 points.

The module gave the smoothest demo, and it still comes second. The weights decide that order: a lender that put most of its weight on write-back and reach would put the module first. The voice platform's cheap minute comes third once the weights count what the bank would have to build. That is the point of weighting before the demos: it stops the room being persuaded by the thing it saw most recently. The score only decides who gets a pilot. Whether AI or human collectors convert better on this bank's accounts, and by how much, is a number only the pilot can produce.

What does automating debt collections at a bank actually cover?

It covers less than the phrase suggests, which is right. For a specialist agent, automating debt collection means automating the conversation and little around it. The accounts, strategies and promises to pay stay in the collections system the bank already runs; the dialer and contact centre stay where they are. A specialist agent should automate the conversation and little else: the first contact within days of a missed payment, the answer to the customer's question, the new date or short instalment plan agreed inside the bank's limits, and the written confirmation afterwards. Every outcome should go back into the system of record the same day, so a collector who picks the account up next sees what was said.

Two things stay with people whatever the platform scores. The policy, meaning what the agent may offer, which someone in credit risk signs; and the customers whose situation needs judgement, who should reach a person without repeating themselves. A platform that claims to automate either is selling something else.