FOXNUT AI

Definition

How should you use AI for debt collections?

The evidence supports using AI for debt collections to send and track messages, with a person on written disputes and requests to stop.

By the Foxnut team · Updated

What can AI for debt collections do?

The evidence supports using AI for debt collections to automate the sending and the tracking, with a person kept in the loop for exactly the two moments the law treats as hard stops: a written dispute, and any request to stop using a channel or to stop entirely. Everything between those two moments is a rule with a stated boundary. Software enforces that kind of rule more reliably than a rota of people working from memory and a spreadsheet.

We are Foxnut AI, a senior-only practice in Bengaluru and Paris that builds AI workflows into commercial operations and hands them over running. There’s a link to talk to us at the end of this guide. Let’s get the rules for your collections straight first.

Important note: this page covers federal rules only.

By debt collections we mean a business chasing its own customers for money already owed: reminders, calls, texts and letters about an unpaid invoice or account, sent by the company that is owed the money or by whoever it has authorised to send them. Before the rules below, the transfer standard for a system that talks to customers is worth reading first. A collections build is one of the few processes in this library that reaches past the company’s own systems and puts a message in front of a person who never chose the vendor or agreed to the schedule.

What changes when a machine does the sending is not the arithmetic. A computed balance is no harder to work out here than anywhere else. What changes is the audience. A robot that texts a customer at 3am about a balance is not a technical achievement, it is a lawsuit that has not been filed yet.

Most processes we cover move a number between systems the company itself operates. This one addresses a person directly, and where that person owes a personal debt and not a business one, a body of federal law written specifically for it governs the contact.

The rules that follow bind a script exactly as they bind a person reading from one. Automating the process does not automate away the obligation. For this process, exact compliance is closer to the whole design problem than to a footnote on it.

Which rules apply to AI debt collections?

Two facts about a collections message decide which federal regime, if either, governs it, and neither fact is about what the message says. The first is who is sending it. The second is what is owed.

Regulation F, the CFPB’s rule implementing the Fair Debt Collection Practices Act, defines a “debt collector” as a business whose principal purpose is collecting debts, or anyone who regularly collects debts “owed or due, or asserted to be owed or due, to another.” It specifically excludes “any officer or employee of a creditor while the officer or employee is collecting debts for the creditor in the creditor’s name.” So a company reminding its own customer about its own invoice, under its own name, generally sits outside the Act altogether.

The exclusion has a named limit. A creditor that, “in the process of collecting its own debts, uses any name other than its own that would indicate that a third person is collecting” is a debt collector regardless of who actually owns the account. The name on the message, not the software behind it, moves a build in or out of the Act’s reach.

On what is owed, Regulation F defines “debt” as an obligation “primarily for personal, family, or household purposes,” meaning a consumer’s obligation. An invoice one business owes another does not meet that definition on any reading of it. A business-to-business collections process sits outside the FDCPA and Regulation F entirely, no matter who sends the reminder or how often.

Neither exclusion reaches the Telephone Consumer Protection Act. The FCC’s implementing rule opens with “No person or entity may” place a call using an automatic telephone dialing system or an artificial or prerecorded voice to a cellular number without the called party’s prior express consent. That restriction attaches to the call itself, regardless of who is calling or what is owed.

Important note for voice AI for collections: a build can sit entirely outside Regulation F, because it is a first-party creditor collecting a commercial invoice under its own name, and still sit inside the TCPA the moment its outreach is an autodialer or a synthesised voice reaching a mobile number.

Regulation F / FDCPATCPA
Original creditor, own name, consumer debtGenerally excluded (creditor-in-own-name exclusion)Applies if the call is autodialed or uses an artificial or prerecorded voice to a cell number
Original creditor, a name suggesting a third party, consumer debtApplies - treated as a debt collectorApplies under the same condition
Third-party collector or debt buyer, consumer debtApplies in fullApplies under the same condition
Any of the above, commercial (business-to-business) debtExcluded - the debt itself is outside the Act’s definitionApplies under the same condition; the TCPA carries no consumer-debt limit

What does an AI collections agent have to enforce?

An AI collections agent has to hold, on purpose, the limits a person used to keep by habit. A person collecting by hand rarely calls the same number eight times in a week. A queue generated by a scheduler will, unless something stops it.

Regulation F sets a specific bound. A collector is presumed to comply if it places no more than seven calls about a particular debt in seven consecutive days, and none within seven days of an actual conversation about that debt, with the conversation date restarting the count. A person following a hunch stays inside that bound by accident. A program has to hold the count per debt, across channels, and reset it correctly the moment a call connects.

Time zone is the same problem in a different shape. The rule presumes any contact before 8 a.m. or after 9 p.m. “at the consumer’s location” to be inconvenient, so the sender’s working hours are irrelevant and the recipient’s clock governs.

The CFPB’s supervisory findings describe exactly this failure. Examiners found that collectors “sent payment reminder emails to the consumer before 8 a.m. in the consumer’s time zone,” which is the mistake a system makes when it schedules from the sender’s calendar and does not resolve each recipient’s local time first.

Consent and disclosure carry the same per-message weight. Every communication after the first has to disclose that it is from a debt collector. The same findings note this failed at “service providers” acting on collectors’ behalf and sending confirmations by text, which is the vendor layer failing a disclosure rule a trained human agent repeats by rote.

A request to stop using one channel binds only that channel under the rule, so a system needs a channel-level opt-out per account and not one blanket flag. A request to stop entirely triggers a near-total halt with a short, named list of exceptions. None of this is a judgment call once it is written down. All of it is state that has to stay correct, per account, per channel, continuously.

What decides whether debt collections is worth automating?

Six structural dimensions decide whether a process is worth automating. Debt collections reads unusually on most of them, because the thing being measured is a communication rule and not a transaction error, and the rule is written down to a level of detail most processes never receive.

Debt collections: process profile
DimensionWhat it reads on debt collectionsSource
Exception varianceDefined by channel and pace, not by the amount owed. The exception a system throws here is rarely a wrong balance; it is a cadence or consent rule crossed - more than seven calls about one debt inside seven days, contact before 8 a.m. or after 9 p.m. at the recipient's own clock, or a message sent through a channel the customer already asked the company not to use. A 2025 supervisory review found this exact failure already occurring at scale: collectors' own emailed reminders going out before 8 a.m. in the recipient's time zone.12 CFR 1006.6(b)(1), 1006.14(b)(2); CFPB, Fair Debt Collection Practices Act CFPB Annual Report 2025
VolumeHigh by design, and the design goal is also the risk. The CFPB received about 207,800 debt collection complaints in 2024, seven percent of every complaint it received that year. Among complaints naming a communication problem, the largest single sub-issue by a wide margin was frequent or repeated contact - 51% of communication-tactics complaints and 58% of electronic-communication complaints. A program built to raise the number of touches per account is tuning the exact variable most complaints describe.CFPB, Fair Debt Collection Practices Act CFPB Annual Report 2025
Cost of an errorPriced per communication rather than per account. FDCPA civil liability runs to actual damages plus up to $1,000 in an individual action, or the lesser of $500,000 or one percent of the collector's net worth in a class action, within a one-year limitations period. A bona fide error defense exists, but only where the collector maintained procedures reasonably adapted to avoid the violation - documented automation is the shield the statute itself describes. The TCPA prices an unauthorised autodialed or prerecorded call to a mobile number separately, at $500 per violation, trebled to a court's discretion up to $1,500 if willful - by the call, not by the campaign.15 U.S.C. 1692k(a)-(d); 47 U.S.C. 227(b)(3)
ReversibilityA sent message cannot be unsent, and the law treats a written dispute as a hard stop rather than a note to fix later. Once the validation notice is sent, a 30-day validation period runs, and a dispute lodged in writing within it requires the collector to cease collection until it sends verification. An automated program has to be able to halt itself on that one written signal, mid-sequence, for that one account, immediately - the fix is not a better message next time, it is nothing further until a document exists.12 CFR 1006.34(b)(5), (c)
Regulatory exposureDepends on who is calling and what is owed, more sharply than elsewhere in this pillar. Regulation F binds only a defined 'debt collector,' and that definition excludes a creditor collecting its own debt in its own name; it also covers only debt owed for personal, family or household purposes, so a business-to-business invoice is outside the Act regardless of who calls. The TCPA carries neither carve-out - its restriction on autodialed or artificial-voice calls to a cell number applies to any person or entity, consumer debt or not, the instant the call reaches a mobile number. The same automated dialer can be fully outside one regime and fully inside the other, on the same call.12 CFR 1006.2(g)-(i); 47 CFR 64.1200(a)(1)
Vendor market maturityRegulator-shaped rather than market-shaped, unusually for this pillar. Regulation F does not merely tolerate automated contact, it specifies the shape it must take: a fixed, closed list of what a compliant voicemail may say, called a limited-content message, and documented safe-harbor procedures for exactly when a collector may text or email an address without a live, channel-specific consent record. A regulator writing down the permitted content of an automated voicemail is a level of specificity most processes in this pillar never receive. The 2025 supervisory findings treat vendors acting 'on behalf of' collectors as an established, examined layer of that market, not a novelty.12 CFR 1006.6(j), 1006.6(d)(3)-(5); CFPB, Fair Debt Collection Practices Act CFPB Annual Report 2025

The regulatory-exposure row reframes the rest. The same automated dialer can be lawful under one regime and unlawful under another on the same call, because Regulation F and the TCPA are triggered by different, independent facts: who is collecting and what is owed, against how the call is placed and where it lands.

The vendor-maturity row says the regulator has already done unusually specific design work on the automation question itself, down to what an automated voicemail is permitted to say. Between them, these rows describe a process where the compliance boundary is not a matter of interpretation once the three triggering facts are known. Most of the engineering is keeping the system honest about which of those facts is true for any given account.

What do AI debt collections programs get wrong?

The first error is reading the frequency and timing rules as a ceiling on effectiveness and not as the actual design brief. Seven contacts a week is not a target to approach. The CFPB’s own complaint data names frequent or repeated contact as the largest single communication complaint by a wide margin, so the rule and the most common complaint describe the same behaviour from opposite sides. A program tuned to maximise touches per account is optimising the variable customers complain about most.

The second error is treating automation as a way to shrink the disclosure and consent burden, when it multiplies it. Every message a program sends after the first one still needs the debt collector disclosure attached. Every autodialed or synthesised-voice call to a mobile number still needs consent captured and kept, per number. A build that sends ten times as many messages needs the same compliance check run ten times as often, not once at the top of the design.

The third error is assuming the FDCPA’s protections travel with the word “collections” and not with the debt. A business-to-business collections process, chasing one company’s unpaid invoice from another, sits outside the Act’s definition of debt entirely. Borrowing Regulation F’s rules there answers a question nobody asked, and the actual design problem for that case sits closer to receivables and cash application than to communications compliance.

The fourth error is treating the bona fide error defence as something a business already has, when a system has to earn it. The defence does not forgive a mistake. It forgives a mistake made despite procedures reasonably adapted to avoid it. That makes the documentation of the automated program’s rules, its logs and its overrides the difference between an error and a court finding intent, not an afterthought layered on top of the build.

Should you use AI agents for collections?

On the evidence, yes for the sending and the tracking, and no for the two hard stops. The evidence supports keeping a person in the loop for a written dispute and for any request to stop using a channel or to stop entirely. Everything between those two moments, the count, the clock, the disclosure line and the consent record, is a rule with a stated boundary and not a judgment call.

The regulatory-exposure row should set the design before any other row is considered. Whether Regulation F attaches at all turns on two facts fixed before the first message is drafted: whose name is on the communication, and whether the amount owed is a consumer’s personal debt or one business’s invoice to another.

The TCPA attaches or does not on a third, independent fact: whether the outreach is an autodialed or synthesised-voice call to a mobile number. That fact is indifferent to the first two. A design built against only one of the three questions will pass its own tests and still be wrong on the ones it never asked.

How can you check your own collections program?

Count, and skip the demonstration. Pull the messages sent to one account in its worst week of contact and set them against the actual rule: no more than seven calls in seven days about that debt, nothing before 8 a.m. or after 9 p.m. at the recipient’s own clock, the collector disclosure present on every message after the first, and a record of consent behind every autodialed or synthesised call to a mobile number.

An account that fails any one of those on paper says more about what an automated program has actually been doing than any dashboard will.

A collections system sends messages under your name to people who owe you money, so one wrong rule can repeat across every account before anyone sees it. The people who sign our proposal are the people who build yours. There is one tier, we made and launched our own products first, and your team holds the count, the clock and the consent record once we have handed the system over.

The first step is a free expert audit: a senior practitioner reviews your setup, with no RFP and no pitch deck. You can read about Foxnut Consulting’s AI enablement work first. We would rather help you build the version that passes that check than watch you find out the hard way it did not, so tell us where your program stands.

Sources

  1. Consumer Financial Protection Bureau, Regulation F, 12 CFR Part 1006 'Debt Collection Practices' - section 1006.2 definitions of debt collector, creditor and debt; section 1006.6(b)(1) time-of-day presumption and section 1006.6(d) rules for third-party and electronic communications; section 1006.6(j) limited-content message; section 1006.14(b)(2) telephone call frequency presumptions; section 1006.14(h) medium-specific opt-out; section 1006.34 validation notice and the validation period. Read as the eCFR enhanced-content rendering of the current partRetrieved
  2. Federal Communications Commission, 47 CFR 64.1200 'Delivery restrictions' - paragraph (a)(1), the prohibition on initiating a call using an automatic telephone dialing system or an artificial or prerecorded voice to a cellular telephone number without the prior express consent of the called party. Read as the eCFR enhanced-content rendering of the current sectionRetrieved
  3. 15 U.S.C. 1692k, 'Civil liability' (Fair Debt Collection Practices Act) - subsection (a) actual and statutory damages, including the $1,000 individual cap and the class-action cap of the lesser of $500,000 or one percent of the debt collector's net worth; subsection (c), the bona fide error defense conditioned on the maintenance of procedures reasonably adapted to avoid the violation; subsection (d), the one-year limitations period. Read via Cornell Law School's Legal Information InstituteRetrieved
  4. 47 U.S.C. 227(b), 'Restrictions on use of automated telephone equipment' (Telephone Consumer Protection Act) - paragraph (b)(3), the private right of action and its statutory damages of $500 per violation, trebled to a court-discretionary maximum of $1,500 for a willful or knowing violation. Read via Cornell Law School's Legal Information InstituteRetrieved
  5. Consumer Financial Protection Bureau, 'Fair Debt Collection Practices Act CFPB Annual Report 2025', November 2025 - section 2.1's complaint count and share (approximately 207,800 debt collection complaints in 2024, seven percent of total complaints that year); section 2.2's communication-tactics and electronic-communication sub-issue shares; section 3.1.3's examination finding that debt collectors sent payment reminder emails before 8 a.m. in the consumer's time zone; section 3.1.6's examination finding that collectors' service providers failed to disclose, in communications sent on the collectors' behalf, that the communication was from a debt collector. Read as text extracted from the Bureau's own report PDFRetrieved

Foxnut AI takes on work like this from Bengaluru and Paris. If you want to see how we do it before you talk to anyone, here is how an AI engagement is scoped and priced.