Can AI collections raise recovery rates without harming customers?

Where is recovery usually raised at the customer's expense?

For a long time, raising the debt recovery rate meant pressing harder. More dials per account, firmer scripts, the consequences mentioned earlier in the call, collectors paid on what they brought in that month. It can work for a while. The costs, complaints, customers who stop answering, customers who clear the balance and close every product they hold, tend to land later, where nobody counts them against the campaign.

Vendor pages promise to boost recovery rates, and buyers remember how recovery used to be raised. That is the fear behind the question of whether AI collections software can improve recovery rates without harming customer experience: that software will do the same thing, only faster and to everyone. It can. An agent told to press will press on every call, consistently, which is worse than a tired collector who lets some customers off. So the answer depends less on the software than on what it is set to do, and on whether anyone is measuring the customer side at all.

What does AI collections software change that moves both numbers?

Five things, none of them pressure. The first is timing. A conversation within days of a missed payment reaches a customer whose problem is still small and who still answers; a team that can only call a fraction of its accounts reaches the rest weeks later, when a second payment may have been missed and the number may already carry a spam label. Software that holds the conversation itself can reach every account early, which is where recovery is easiest and the customer is spared the escalation.

The second is the offer. An agent that asks about the customer's month before naming a figure, and then offers an instalment plan the customer can actually keep, has a better chance of being paid over six months than one that demands the full balance today, and leaves the customer with a plan rather than a threat. The third is tone: the agent uses the same courteous wording on every call, with no bad afternoon, no commission at stake and no escalation the rules didn't allow. The fourth is channel and hour: a WhatsApp message to the customer who never picks up unknown calls, a call in the evening to the one who does. The fifth is the handover. A customer who says they have lost their job reaches a person within a minute, with their words already written down, instead of a collector working through a script.

Each of these can lift the share of balance collected for the same reason it protects the customer: the customer gets a conversation they can act on, early, in private, on terms they named. That is why the two numbers can move together, and why they only do when the agent is set up to offer rather than to press.

What should you measure on both sides?

One scorecard with two columns, read together every month. Read on its own, recovery rewards pressing harder; read on its own, customer experience rewards never calling. Side by side, they show whether the lift is real or came from pressing harder.

What does the sum look like in numbers?

Here is an illustration with invented numbers; replace them with your own. A consumer lender has 10,000 accounts a month entering early arrears, the first weeks after a missed payment, averaging 400 euros: 4 million euros at risk. Its collectors reach about a third of those accounts in the first two weeks, and the book recovers 38 percent within 90 days, 1.52 million euros. Complaints run at 3 per thousand contacts, and many of them are about the number of calls.

Now every account gets a conversation within three days, by message first and by call where the message goes unanswered, with offers limited to what the lender pre-approved and a contact cap of two attempts a week. Suppose recovery within 90 days rises to 44 percent: 1.76 million euros, 240,000 more a month. Suppose complaints fall to 2 per thousand contacts, because fewer customers are called four times in a week and more are offered a plan they can keep. More accounts are contacted, so the absolute number of complaints can still rise while the rate falls; the rate is the fair comparison, and the contact cap keeps the absolute number in check. Those two results are not in tension; they come from the same change, which is that more customers got an early, affordable offer.

The number to distrust is a recovery lift with no customer-side figure next to it. If recovery rises to 44 percent and complaints rise with it, or the share of customers who leave within a year climbs, the lift came the old way, and the scorecard shows it in the second column.

Where does the trade-off still exist?

In frequency, mostly. Every extra contact recovers a little more and annoys a little more, and software makes extra contacts cheap, so the temptation to add a third and fourth attempt a week is real. Set the cap before go-live, count every channel against it, and treat the complaint line as the earlier warning; it is the one you can read month to month.

It also exists in what the agent is allowed to say about consequences, and in how long it keeps trying an account that never answers. Both are policy, written by people, and both should be reviewed against the customer column of the scorecard rather than the recovery column alone. The software will do what it is set to do on every call. That consistency is its strength, so the settings deserve more attention than the demo.