What to ask an AI collections vendor before you sign

The questions that matter when buying AI collections are not about the demo. Ask how the agent is kept inside your rules, what evidence each conversation leaves, how much traffic is reviewed, which languages are native, what integration needs, and the price at your volume. A vendor at ease with those questions is one you can audit.

Murphy, Editorial team5 October 20267 min read

Why does the demo tell you almost nothing?

Every vendor's demo call is flawless, whether the buyer is a bank or a small lender. It is the one call the vendor controls completely. Your book is different: wrong numbers, angry customers, two languages in one household, a dispute halfway through a payment plan. The questions below can only be answered well by a platform that has survived a real portfolio, and that is the one thing a demo cannot show. They apply to conversational AI collections of every kind, meaning software that holds a real two-way conversation by voice or message rather than playing a recording, and they matter most for voice AI for debt collection, where a mistake is spoken aloud before anyone can review it.

The ten questions

Put these in the RFP and insist on specific answers. The second column is what a serious answer sounds like; anything vaguer is the answer.

Ask	A serious answer sounds like
How is the agent kept inside our rules?	Enforced limits the agent cannot cross, versioned and logged, not a prompt asking it nicely.
What evidence does each conversation leave?	A record per interaction, exportable, tied to the rules in force at that moment.
How much traffic is compliance-reviewed?	All of it, automatically, with humans on the flags. Sampling leaves most calls unchecked.
Which languages are native?	Named languages with locally reviewed scripts, not machine translation at call time.
Which channels, on one plan?	Voice and written channels coordinated under one contact strategy with your caps.
When does a human take over?	Defined triggers, a warm handover, and the conversation history travels with it.
Where does our data live, and how long?	A named region, a retention policy that follows yours, and deletion on exit.
What does integration need from us?	A concrete list on day one: which systems, which fields, which direction.
What does it cost at our volume?	A number computed for your book size, with the assumptions written down.
What happens when we leave?	Your data back in a usable format, timelines in the contract.

A vendor comfortable being audited is the only kind worth piloting.

Which answers should worry you?

Some answers should end the meeting. A recovery figure quoted without the portfolio behind it is marketing; without the arrears band, the product and the age of the debt, the number means nothing. Quality assurance by sampling should worry you too, because the point of software is that it can review everything. So should per-seat pricing on a product whose job is to need fewer seats. And so should any hesitation about letting your compliance team listen to real calls from a live deployment.

How do you run a fair pilot?

Whether AI or human collectors convert better on your book is something only a test on your book can tell you. Split comparable accounts between the platform and your current operation, with matching products, arrears band and rules on each side, and score both on one number: share of balance recovered. Any looser design leaves both sides room to argue the result. Keep the pilot long enough for payment plans to mature, and agree before it starts which number decides.

Common questions

Is an AI agent allowed to make collections calls?

Yes in most markets, under the rules that bind human callers: identification, consent, recording, contact hours and disclosure wording. Some add AI duties: the EU AI Act requires the agent to say it is an AI, and in the UK the FCA's Consumer Duty holds the firm to good outcomes whoever makes the call. In Italy, check with counsel first: the privacy authority bars pre-recorded collection calls with no operator. Ask how each rule is enforced.

What recovery rate can AI collections achieve?

It depends on the book far more than on the software: the arrears band, the product and the age of the debt decide most of the number. Treat any rate quoted without those three as a sales figure. The honest figure is the one a fair pilot produces on your own accounts against your current operation.

How do you integrate AI collections with core banking?

Usually without touching the core itself. The platform needs the accounts to work, balances and arrears status, and sends back outcomes, promises to pay and payments, through a daily file or an API. Ask the vendor for that field list on day one; it is most of the integration.

Do I need my own compliance team to use one?

You need someone who owns the rules the agent runs under, even part-time. The platform should make that person stronger, every conversation reviewed and filed, not replace their judgment.