Someone who's three weeks behind picks up, hears the balance and says they'll pay on Friday. Your agent logs a promise to pay and the call goes down as a win. Nobody's paid anything yet.
If you run collections, or build the software and AI agents collections teams use, this is about the last minute of that call. Most of the talk about AI debt collection is about whether the agent can hold the conversation. I think the bigger question is what it does once the customer says yes. In one study, AI and human collectors had got about the same share of people to pay fifteen minutes after the call, and were far apart by the end of the day.
A promise to pay is a note on the account
When someone promises to pay, the money's where it was before the call. There's a date on the account now, and when it comes, something has to check whether the money turned up. If it didn't, the chasing starts again.
A payment taken on the call is money. A payment set up on the call for the date they've agreed goes on that day without anyone chasing it, unless something fails.
On the call, a promise and a payment can sound the same. One study of AI collectors shows where they come apart.
The payment gap opened after the call ended
The study is an NBER working paper by James Choi and three co-authors. At a big Chinese online lender, a random sample of late borrowers was assigned an AI or a human caller.
Take the first answered call on day two overdue. Fifteen minutes later, about the same share of borrowers had paid something: 4.8% for the AI and 4.5% for the humans. By the end of the day it was 27% against 45% (Table 5 in the full paper).
So the gap in payments opened in the hours after the call.
Of the borrowers the AI logged as promising to pay that day, about a third (34%) had paid anything by the end of it (Table 6). The AI also logged fewer promises on those calls: 44%, against 65% for the humans (Table 5). That's not quite like for like, because a promise to a human could run to the end of the next day.
As for why, the authors suggest the AI's "lesser ability to extract promises that feel binding may contribute to the performance gap." That's a "may", not a proof.
The AI is also 2021 to 2022 technology at one lender, in a working paper that hasn't been peer reviewed, so it's no verdict on what you can buy today. It's still the clearest look I've seen at what happens to a promise once the call ends.
InDebted and PayNearMe put the payment on the call
InDebted's number goes the other way. It launched an AI voice agent called Voice in November 2026, and its release says that in early testing on its own accounts, customers who agreed a payment plan with Voice "were about twice as likely to keep it as those who spoke with a human agent."
There's no sample size, and no figure for how often plans usually get kept. It's a different promise too: the study's borrowers were promising to pay that day or the next, and InDebted means plans.
It doesn't pin the result on any one feature. It does say customers "can agree a plan and pay by secure link while still on the phone." I don't know how many did, or whether it's why plans held, but that's the part I'd copy.
PayNearMe's AI Servicing and Collections Agent, launched in October 2026, lists "accept or schedule a payment, capture promises to pay" among the things it can do (its release). Accepting or scheduling a payment ends the call with the money taken or set up. Capturing a promise doesn't.
Where the accounts that don't pay end up
Encore Capital and PRA Group buy defaulted debt from banks and other creditors, then collect on it. Their annual reports both say something about when court comes in. PRA considers its legal channel "when customers do not engage with us voluntarily." Encore says that if its calls or emails get blocked, it "would need to pursue collections through our higher-cost legal collections channel."
Neither filing mentions promises. But when a promise from a call gets broken, the chasing starts again, and my guess is some of those accounts end up on the legal side.
Take the money before the call ends
If your agent or team gets a yes, I'd take the first payment before the call ends. If they say Friday, set up Friday's payment there and then. Keep the promise for people who can't set a payment up today, say because they're not sure the money will be there by then. A payment that's going to bounce helps nobody.
Taking a card on a call raises its own questions, like who hears the number and how later instalments get charged. Our guide to secure payment collection for debt agencies covers those.
And if you're judging an AI agent, judge it on money in, paid on the call or on the date that was set, not promises captured. That's what the creditor actually gets. If you've got both numbers for your own calls, the gap between them is where I'd look first.
The video above is Shuttle's own explainer. On Twilio, the caller keys the card into Twilio's <Pay>, Shuttle passes it to your payment provider, and your team programs the hand-back to the agent.
Frequently Asked Questions
What is a promise to pay in debt collection?
It's when someone on a collections call agrees to pay by a set date and the collector records it on the account. No money has moved yet. When the date comes, something has to check whether the money arrived.
Are AI debt collectors worse than humans at getting people to pay?
The evidence is mixed, and the two sources measure different things. In one NBER study at a Chinese online lender, AI callers using 2021 to 2022 technology were level with humans 15 minutes after the call but well behind by the end of the day. InDebted says payment plans agreed with its agent were kept about twice as often as with a human, in early testing on its own accounts, with no sample size given.
Related reading
Pay by text or log in first? What SBT's survey says about paying bills
Dutch regulator: approving a payment plan by software can still be automated decision-making