How to use AI in debt collection: what to automate first, and what to never hand over

The right way to use AI in debt collection is to automate in order of risk: start with the schedule (deciding who gets contacted today), then the drafting (writing the message for that specific account), then multi-channel sending, and add voice last if at all. Keep four things human: disputed accounts, negotiating settlement terms, the decision to escalate legally, and any customer whose relationship is worth more than the invoice.

That ordering is not caution for its own sake. It follows the value. The scheduling step is where nearly all of the recovery gain lives and it carries almost no downside risk, while the voice step carries the most relationship risk and the least incremental gain for a business chasing its own commercial invoices. Most teams do this backwards, get excited about an AI voice agent, and never fix the reason their invoices aged in the first place.

Why your invoices actually age

Before automating anything, it is worth being precise about the failure you are fixing. In almost every small and mid-sized business we look at, overdue invoices are not sitting there because someone chased them badly. They are sitting there because nobody chased them at all for several weeks.

The pattern is consistent. An invoice goes out. It passes its due date during a busy stretch. The person responsible for receivables is also doing payroll, or selling, or delivering the work. By the time anyone opens the aging report, the invoice is 45 days old, the conversation is now awkward, and the awkwardness delays it further. Three months later it is a collections problem.

This matters because it tells you what to buy. Industry recovery norms put collection above 70% on invoices under 90 days past due, and often below 15% once an account passes 180 days. The single highest-value thing software can do is move first contact from week six to week one. Every feature that does not serve that is secondary.

Step 1: automate the calendar before anything else

The first thing to hand to software is the decision about who gets contacted today. Not the writing, not the sending. The deciding.

Write down your escalation policy once: what happens three days before due, on the due date, at day 7, day 15, day 30, day 60 and day 90. Most businesses have never written this down, which is precisely why it does not happen. Once it exists as configuration, the software builds the day's contact list every morning without anyone opening a spreadsheet.

This step alone typically produces the majority of the improvement, and it is essentially risk-free. You are not changing what you say to customers. You are changing whether you say it. If you do nothing else on this list, do this.

Step 2: let AI draft, but set the tone yourself

The second thing to automate is the drafting. A three-day courtesy nudge on a $400 invoice and a 75-day escalation on a $38,000 balance should not be the same email, and in practice most businesses send a version of the same email to both because writing a fresh one every time is unrealistic.

A language model handles this well. Give it the account context (balance, age, payment history, what has already been sent, whether they have broken a promise to pay) and it will produce a message that fits the situation. The quality bar is not literary. It is that the message references the right invoice number and the right amount, and that the firmness matches the stage.

Two guardrails matter here. Set the tone explicitly, because default AI collections copy tends toward a generic corporate register that does not sound like you, and your customers know what you sound like. And review the first twenty messages before you let them send unattended. You are checking for factual errors about the balance far more than for style.

Step 3: add channels in the order your customers actually respond

Once the calendar and the drafting are running, add channels. Email first, because it is written, cheap, and creates the record you will want later. SMS second, for the short nudge that gets read, and only for customers who have given you a mobile number in a business context.

Voice comes last, and for B2B collections it should be used narrowly. AI voice agents have improved fast, and the same underlying technology now handles automated calls that qualify prospects and book meetings in sales teams. In collections the economics are strongest for high-volume consumer books, where a per-call cost far below a human collector transforms the early delinquency stage.

For a business chasing its own commercial invoices, the calculation is different. The person you need is a specific accounts payable clerk at a company you want to keep selling to. An automated call handled badly costs you a relationship worth more than the invoice. Use it as a scheduled reminder on accounts that have ignored two written contacts, in business hours, with clear identification and an immediate route to a human.

The compliance guardrails to configure on day one

No US federal rule prohibits using AI to collect a debt. The FDCPA, Regulation F and the TCPA regulate the contact itself, not the technology that produced it, and the CFPB has been clear that a company is responsible for what its AI systems do just as it is for its employees. So the rules you already have to follow are the rules your agent has to follow.

Configure these before the first message goes out:

• Calling hours. Consumer-debt calls run between 8:00 a.m. and 9:00 p.m. in the debtor's local time zone. Set the agent to the customer's time zone, not yours.
• Frequency caps. Regulation F presumes harassment above seven calls about a particular debt in seven days, and prohibits a call within seven days of a live conversation about that debt. The count is per debt, not per customer, and your cap needs to apply across every channel.
• Automated disclosure. A voice agent should say plainly that it is automated and should not mislead anyone about what it is.
• Hard stop on disputes. Any dispute or request for debt validation stops the sequence immediately and routes to a person. This should not be a setting someone can forget to switch on.
• Exclusion list. Strategic accounts, customers in an active payment plan, and anything already with an attorney come out of the automated flow.

If you are collecting your own B2B invoices, the compliance picture is genuinely simpler. The FDCPA covers debt incurred primarily for personal, family or household purposes and primarily regulates third party collectors, so a business collecting its own commercial debt in its own name is generally a first party creditor outside it. The TCPA still applies to calls and texts, a personal guarantee pursued against an individual can pull you into consumer territory, and California SB 1286 extended Rosenthal-style protection to certain covered commercial debt from July 1, 2025. This is information, not legal advice, and state law varies. Build to Regulation F standards anyway. It costs you nothing and it is the posture you want if anything is ever questioned.

What to never hand to AI

Four things belong with a person, and the discipline to keep them there is what separates a collections program that works from one that creates problems.

Disputed accounts. When a customer says the work was wrong, the delivery was short, or they already paid, that is a claim to investigate. No follow-up sequence resolves it, and continuing to send reminders into a live dispute is how a receivable becomes a complaint.

Settlement negotiation. Deciding to take 70 cents on the dollar, or to accept a six-month payment plan, is a commercial judgment about your cash position and your read of the customer. Let the agent surface the request. Let a person answer it.

Legal escalation. The decision to send a final demand, place with an agency, or file in small claims is yours. Software should force that decision onto someone's desk when a sequence exhausts itself, which is genuinely useful, because the normal failure mode of manual collections is that nobody ever decides anything and the account just goes quiet.

Relationship accounts. Your three largest customers do not need a sequence. They need a call from someone they know. Configure them out and stop worrying about it.

How to tell whether it is working

Measure two things and ignore the rest for the first quarter. First, days sales outstanding, which is the number that reflects whether cash is actually arriving sooner. Second, the age of first contact: how many days past due, on average, before the customer hears from you. That second number is the one automation moves immediately, and it is the leading indicator for the first.

Be skeptical of the percentage lifts in vendor case studies. They compare an automated program against whatever the client was doing before, which is often nothing systematic, so the measured improvement includes the effect of simply having a process. That effect is real and it is worth paying for. It is just not evidence that a particular model is smarter than another one.

Predictive treatment models do produce genuine gains on portfolios large enough to test across thousands of accounts. On a few hundred open invoices there is not enough signal for a model to beat a well-designed calendar, and a vendor claiming otherwise is selling you something.

Where to start this week

Open your aging report and write down what should happen at day 3, day 15, day 30, day 60 and day 90. That document is the actual asset. Everything else is execution, and execution is what AI debt collection software is for.

Then load your open invoices, set that calendar as configuration, and let the routine follow-up run in your own name while you handle the exceptions. If the balance is already badly aged and you have written it off internally, that is the case where placing the file with a collection agency makes sense despite the 25% to 50% commission. For everything younger than 90 days, consistent contact you did not have to remember to send is the whole game.