AI debt collection: what to automate, what regulators expect

Every month the same accounts fall into arrears, and your collectors work through them in whatever order the dialler serves them up. The account worth four figures gets the same thirty seconds as the one worth forty pounds, and the accounts nobody reaches at all simply roll into the next cycle.

AI debt collection is where most lenders start, because the volume is high, the conversations repeat, and the business case is easy to put in front of a CFO. It is also where the regulatory exposure concentrates: the FDCPA and Regulation F in the US, TCPA rules covering calls and texts, FCA CONC and the Consumer Duty in the UK.

This guide covers consumer debt collection, not accounts receivable or commercial recovery. It sets out what AI agents handle in arrears work today, what belongs with your people, how an AI collections agent runs a contact from first dial to CRM note, and which rules govern each of those decisions. For the rollout sequence that follows those decisions, see our step-by-step rollout for lending and collections.

What AI debt collection means in consumer lending

AI debt collection is the use of autonomous agents to run contact and resolution on overdue consumer credit accounts across voice, SMS, email, and chat. It spans the whole arrears timeline rather than one moment in it: pre-arrears reminders before a payment is missed, early arrears in the first 1–30 days, later arrears where the balance has aged, and recovery work on accounts that have already defaulted.

Two distinctions matter before you scope any of this work.

First-party and third-party collection are regulated differently. The FDCPA applies to third-party debt collectors and debt buyers, not generally to a creditor collecting its own debt in its own name. A lender running an in-house arrears team sits outside much of the FDCPA, though UDAAP, state law, and the TCPA still apply. An agency running the same conversation sits squarely inside it. The same AI agent, deployed on either side of that line, inherits a different rulebook, so settle which one you are before you write a single procedure.

Consumer debt collection is not accounts receivable. Searches for collections automation drift quickly into B2B invoicing and dunning software, which solves a different problem for a different buyer with none of the consumer protection rules attached. Everything below concerns consumer credit: loans, cards, income advance, and BNPL.

For the wider arrears vocabulary, from arrears management through to collections automation, our glossary covers the terms that come up most often in arrears work.

The rules that govern AI debt collection in the US and UK

Put this section in front of your compliance team early, because it decides the shape of everything you build afterwards. None of these rules ban automation. The CFPB's Regulation F is typical of the set: it constrains how often you make contact, when, through which channel, what you say, and what you must disclose, without once asking whether a human or an agent is doing the work.

One rule bites AI deployments specifically. In February 2024 the FCC ruled that AI-generated voices in robocalls count as "artificial" voices under the TCPA, which means a synthetic-voice collections call to a mobile number needs the same prior express consent as a prerecorded one. Vendors who treat AI voice as a neutral delivery mechanism have this backwards, and it is worth asking any vendor how they handle consent capture and scrubbing before you see a demo.

The UK equivalent of that pressure point is contact strategy. CONC 7.3 and the Consumer Duty together make call cadence a conduct question rather than a performance one. An agent that can dial ten thousand numbers an hour is a liability unless the cadence rules sit above it, which is why contact frequency belongs in the agent's guardrails and not in a dialler setting somebody can override on a slow Tuesday.

What to automate: outbound collections, payment plans, and inbound queries

The work that automates well in collections shares three traits: it repeats at volume, the right answer follows from account data rather than negotiation, and the regulatory boundaries can be written down in advance. That covers more of the book than most teams expect.

• Outbound contact at the right moment: segmenting by delinquency stage and dialling or messaging when that borrower is most likely to answer. AI agents for outbound collections make this economic on low-ticket debts in the £0–200 range, the segment most teams cannot justify chasing by hand
  
  

• Securing a promise to pay: confirming the full arrears amount and a date, then logging the commitment against the account. This is the core early-arrears outcome, and it is covered in detail on our secure promises to pay page
  
  

• Taking the payment and setting up the mandate: a commitment is not money. The agent takes payment on the call, sends the direct debit or autopay link live, and confirms the mandate before it hangs up
  
  

• Handling the inbound calls outbound creates: every outreach campaign generates callbacks, and those callers reach inbound collections queries with their full history already loaded and stricter identity checks applied
  
  

• Post-contact admin: the CRM note, the confirmation email or SMS, and the audit trail recording every data point referenced and action taken. This is unglamorous and it is where human collectors lose a quarter of their day
  
  

Our overdue payment collections page walks through how a single outbound contact runs end to end, and the lenders' use case guide sets collections alongside the rest of the borrower journey.

What stays with your human collectors

Drawing this line precisely is what separates a deployment your compliance team signs off from one they stall. The boundary is set by whether a regulator expects a person to exercise judgement on that decision, not by how difficult the conversation is.

• Affordability assessments and arrangements to pay: a promise to pay is the full arrears balance on a date, and an agent can take one. An arrangement to pay is partial and recurring, which requires an income and expenditure assessment, and that belongs with a person. Lenders who blur these two create exactly the affordability gap the FCA looks for
  
  

• Hardship and forbearance decisions: the agent can gather the income and expenditure answers and open the case, which is how our hardship assessment and forbearance flow works, but the forbearance decision itself goes to a specialist
  
  

• Any signal of vulnerability: bereavement, illness, mental health, and coercion all route to a human on detection, mid-conversation, without finishing the collections script first
  
  

• Hard-stop account categories: Individual Voluntary Arrangements, Debt Management Plans, and Breathing Space registrations must never receive a collections contact at all, which makes this a pre-dial screening job rather than an in-conversation one
  
  

• Complaints and high-value settlement negotiation: both carry commercial or conduct consequences that justify a person's time, and the economics support it once the agent has absorbed the low-ticket volume underneath
  
  

The pattern across all five is the same: the agent does the gathering and the admin, a person makes the call that carries regulatory weight.

How an AI collections agent runs a single contact

Inside a single contact, a well-built collections agent runs a fixed sequence, and every step of it is auditable afterwards.

1. Segment and schedule. The agent selects who to contact based on delinquency stage, prior contact history, and channel preference, then checks the contact against cadence limits and hard-stop flags before anything dials
   
   

2. Verify identity. Identity checks run before any account detail is discussed, and the threshold is stricter on inbound than outbound, because an inbound caller has not been authenticated by the fact that you dialled them
   
   

3. Load live account data. Balance, payment history, prior promises, eligible repayment options, and settlement figures come from the servicing system during the conversation rather than from a nightly export
   
   

4. Run the conversation and present options. The agent works the approved repayment options for that borrower's stage and circumstances, in their language and local accent, without putting anyone on hold to look something up
   
   

5. Take the action. Payment taken, mandate set up, promise to pay captured, callback booked, or hardship case opened, depending on where the conversation lands
   
   

6. Close the loop. The CRM note is written, the confirmation sent, and the full transcript with decision rationale logged against the account
   
   

Guardrails run across all six steps rather than at the end. Customer-side guardrails watch for complaints, vulnerability, and financial difficulty and reroute the moment one appears. Agent-side guardrails catch false promises, disallowed terminology, tipping-off, and out-of-bounds advice, editing the draft before the borrower ever hears or reads it. Gradient Labs builds AI agents for financial services, and our Lending Agent runs more than twenty lending-specific guardrails of this kind on every turn, pre-configured for US, UK, and EU consumer protection rules, including those above.

What good looks like in production

Collections teams do not judge an agent on resolution rate, and a vendor who leads with one has misread the buyer. The scorecard here is recovery rate, payment rate, and contact rate, with promise-to-pay maturity sitting underneath as the number that tells you whether any of it was real.

SteadyPay, an FCA-authorised embedded lending provider, runs outbound voice collections, payment plan negotiation, and identity verification through the agent.

"Before Gradient Labs, we needed to find a way to reach our growing number of customers effectively. Now we make 33,000 calls a month, converting 60% of engaged customers to committed repayment dates, all within FCA compliance standards. It has fundamentally changed how we manage the collections layer of our lending infrastructure."

Violeta Filip, Head of Customer Experience, SteadyPay

The compliance standard and the recovery number sit in the same sentence there, because in collections neither one counts without the other. On the card side, Yonder now has 85% of its collections processed by AI agents, with the frontline and back-office agents holding case context between them.

Three measurement habits separate the deployments that hold up from the ones that look good for a quarter.

• Track promise-to-pay maturity, not promise-to-pay volume. Commitments are easy to generate and worthless if they do not convert. Measure the proportion that turn into cleared funds, by delinquency stage
  
  

• Track contact compliance as a first-class metric. Calls per account per week, time-of-day distribution, and hard-stop breaches belong on the same dashboard as recovery rate, not in a quarterly audit
  
  

• Set the comparison against your own prior baseline. Collections performance varies enormously by book quality and segment, which makes any cross-company benchmark close to meaningless
  
  

Trust matters more in collections than almost anywhere else in the borrower journey, and AI reputation for lenders covers how the quality of these conversations feeds back into the brand.

If you are further along and comparing suppliers rather than scoping the work, our ranked guide to lending AI vendors covers that decision instead.

Want to see how an AI agent handles your arrears book inside your own rules? Book a demo.