How do you bring AI onto a collection floor without hurting results?

The short answer

Start where the risk is lowest, and bring AI in one change at a time. In our view, work that helps collectors behind the scenes, such as call notes and compliance checks, is a safer first step than anything that talks to consumers. Decide how you will score each change before it starts, and test it against a control group run in the same weeks. Keep people on the calls that need persuading. And keep watching after launch, because AI tools change and so do their results.

Start behind the scenes

A sponsored recap of an industry webinar on insideARM (2026), written by Dan Ward, a vice president at Finvi, quoted one panelist describing the approach as wrapping AI around the agent rather than putting it between the agent and the consumer. It listed lower-risk uses behind the scenes: call note summaries, real-time compliance monitoring, post-call documentation checks and help deciding which accounts to work first. It urged caution with early outreach, sensitive conversations and anything that needs empathy.

In our view that is the right order: a tool that writes call notes can be checked against the notes collectors already write. A tool that talks to consumers changes the conversation itself, and in our experience the conversation is where the money is won or lost. Will AI voice bots replace human collectors?

Change one thing at a time

If a voice agent, an agent-assist tool and a new dialer setting all go live in the same month and collections move, nobody can say which one did it. Bring in one change, measure it, then start the next. Or hold everything else equal inside a test: split accounts at random, give both groups the same scripts, offers and dialer rules, and let the only difference be the change being tested. How agent matching fits with agent assist and speech analytics.

Decide the score before you start

Agree the measure before the test begins, not after the results come in. In our view the main score is dollars collected per account contacted, with complaints and compliance findings alongside it. In our view promises are a weak score for AI, because a bot and a person can count a promise on different rules, and a promise made to a machine may not be kept the same way. How to prove a routing change worked.

Keep watching after the first weeks

In our view some effects take months to show. In "How Good is AI at Twisting Arms?" (NBER working paper w33669, 2025), Choi, Huang, Yang and Zhang studied calls to late-paying borrowers at an online consumer lender and found AI callers substantially less effective than people. Switching to human callers six days into the delinquency closed much of the gap. But borrowers first contacted by AI had repaid less of the initial late payment a year later, and were more likely to miss later payments, than borrowers first contacted by people. A test that stopped after the first weeks would have missed that.

The tools also change. The insideARM recap stressed that models change, that even small updates can change what a tool produces, and that results need checking on an ongoing basis. In our view that means keeping a small control group running, or rerunning the test after each significant update, rather than trusting a result from launch month for good.

What to ask an AI vendor before you start

• What does it do with consumers, if anything? Does it call, text or email them, or only help your collectors? If it contacts consumers, how does a consumer reach a person when they ask for one?
• Who is accountable when it gets something wrong? The insideARM recap was clear that an agency cannot hand its compliance risk to a vendor. Your compliance team and counsel decide how the rules apply; BestPair does not provide legal or compliance advice.
• What data does it use, and where does it go? Including anything it records or keeps from calls.
• How will we measure it? Ask for a test against a control group on your own accounts, and the data to rerun the result yourself. How to check a vendor’s lift claim on your own data.
• What changes when the model is updated? Will you be told, and can you retest?
• How do we switch it off? And what happens to your floor and your data when you do.

Bring the floor with you

In our experience a tool the collectors don’t use does nothing, however good the test looked. In our experience the people who have worked the phones longest are often the most doubtful, and they are usually the ones others follow. Show them what the tool does for them, keep their day as familiar as you can, and have managers keep it going after the launch week. How to get agents to actually use new tools.

Where agent matching fits

Agent matching is one of the changes a floor can bring in, and it works behind the scenes: it changes which collector works each account, and people make every call. BestPair sends each account to the eligible collector who has done best with accounts like it, from the floor’s own call and payment history, inside the rules and workload limits the agency sets, and measures the result against a control group. It does not dial, call, text or contact consumers, and it does not listen to calls. Like any change, it should go in on its own, so its result is clear.

Sources

• Safe, Confident AI Adoption in Collections: What the Experts Recommend: insideARM, 12 May 2026, a sponsored recap of an insideARM webinar, written by Dan Ward, Vice President of Payments at Finvi.
• How Good is AI at Twisting Arms? Experiments in Debt Collection: Choi, Huang, Yang and Zhang, NBER working paper w33669, 2025.