AI Voice Agents for Mortgage Arrears: How UK Lenders Are Automating FCA-Compliant Early Arrears Outreach on Amazon Connect

Arkadas Kilic 10 min read · 2026-06-11

Rel8 CX builds autonomous AI voice agents for UK mortgage servicers on Amazon Connect, delivering production deployments in 4 to 6 weeks. This post covers what those deployments actually look like, the FCA compliance architecture that makes them viable, and the numbers we see in live production.

If you're evaluating whether AI can handle mortgage arrears conversations without creating regulatory exposure, this is the post for you.

The Problem with Early Arrears Outreach at Scale

Most UK mortgage servicers know the math. A borrower who misses one payment and receives a meaningful conversation within 72 hours is dramatically more likely to self-cure than one who gets a letter on day 14. The FCA's MCOB 13 rules and Consumer Duty obligations make early, empathetic contact not just good practice but a regulatory expectation.

The operational reality is harder. A mid-sized mortgage book of 80,000 accounts might generate 600 to 900 early arrears cases per month. Staffing a team to call every one of those within 72 hours, at the right time of day, with a compliant script, and with live access to the account system is expensive. Most servicers triage. The borderline cases get letters. The letters get ignored. By the time a human calls, the arrears are deeper and the customer is more distressed.

This is where an autonomous AI voice agent changes the economics entirely.

What "FCA-Compliant" Actually Means in This Context

Before getting into architecture, let's be precise about what compliance requires here. Vague claims about "FCA-aligned" AI don't hold up when the FCA reviews your complaints data or your Consumer Duty board report.

The specific obligations that shape how we build these agents:

Building compliance in means the agent architecture has to handle all of these, not as a post-build checklist, but as core logic.

The Architecture: How We Build This on Amazon Connect

Here's the production stack we deploy for mortgage arrears agents.

Outbound Campaign Orchestration

Amazon Connect's outbound campaigns feature (powered by Amazon Pinpoint) handles the dialler logic. We configure campaign queues with:

• Time-of-day restrictions (no calls before 8am or after 8pm, aligned to MCOB 13 and FCA guidance on contact times)
• Attempt limits per account per day (typically 3, configurable)
• Answer machine detection with a compliant recorded message that does not disclose arrears details
• Priority weighting so accounts at higher delinquency risk get earlier call attempts

The campaign pulls from a daily feed from the loan management system (LMS). We've integrated with Phoebus, Nivo, and Mortgage Brain in past builds. The feed includes account balance, days past due, previous contact history, and any existing payment arrangement flags.

The Conversational Agent

The voice agent runs on Amazon Lex with a custom Lambda integration layer. The conversation flow handles:

1. Identity verification: We use a two-factor approach combining date of birth and postcode. For higher-risk conversations (where an arrangement is being set up), we add a third factor. This is configurable based on the lender's existing ID&V policy.

2. Vulnerability screening: Before any arrears conversation, the agent runs a structured vulnerability check. If the customer indicates financial hardship beyond the immediate arrears, recent bereavement, health issues, or difficulty understanding, the agent flags the account and transfers to a human specialist queue. We track this flag in a DynamoDB table that syncs back to the LMS.

3. Arrears disclosure and reason capture: The agent states the arrears position clearly, asks for the reason for the shortfall, and captures a structured reason code (income reduction, unexpected expense, payment processing issue, etc.). This feeds directly into the LMS and informs the arrangement offer.

4. Payment arrangement negotiation: The agent has access to the lender's arrangement policy via a Lambda function that queries the LMS in real time. It can offer a range of arrangement options within defined parameters: full payment now, partial payment with catch-up schedule, or payment holiday if the account is eligible. It cannot offer anything outside the policy matrix. This is a hard constraint, not a soft one.

5. Arrangement confirmation and consent capture: If the customer agrees to an arrangement, the agent reads back the full terms, captures verbal confirmation, and writes the arrangement to the LMS via API. The call recording timestamp is logged as the consent record.

6. Signposting: Every call, regardless of outcome, ends with a signpost to free debt advice (MoneyHelper, StepChange). This is a Consumer Duty requirement and we treat it as non-negotiable.

The Escalation Logic

The agent escalates to a human in any of these conditions:

• Customer requests to speak to a person (immediate, no resistance)
• Vulnerability indicators detected
• Customer disputes the arrears amount
• Account has an existing legal hold or litigation flag
• Three failed ID&V attempts
• The requested arrangement falls outside the policy matrix
• The customer mentions mental health, self-harm, or uses distress language (detected via Amazon Comprehend sentiment analysis running in parallel on the transcript)

Escalations route to a specialist queue in Amazon Connect with a screen-pop showing the full conversation transcript, the account details, and the vulnerability flag if set. The human agent picks up with full context. No repetition required.

Compliance Architecture: The Bits That Actually Matter

Here's a comparison of a typical bolt-on AI approach versus the production-grade compliance architecture we build:

The dual-control approach on contact times is something we added after a client audit. If the campaign configuration has an error, the agent itself checks the current time before proceeding. Two failure modes have to align for a non-compliant contact to happen.

What the Numbers Look Like in Production

We don't publish client names, but here are the metrics from live mortgage arrears deployments.

The FCA Consumer Duty Angle: Board-Level Reporting

One thing lenders consistently underestimate is the Consumer Duty board reporting requirement. Firms have to demonstrate, with evidence, that they're delivering good outcomes for customers in financial difficulty.

An AI voice agent on Amazon Connect generates structured data that makes this reporting tractable:

• Every call produces a transcript, a structured outcome code, a vulnerability flag (yes/no/reason), and a debt advice signpost confirmation
• Arrangement terms are written to the LMS with a timestamp and call reference
• Escalation reasons are categorised and reportable
• Contact attempt patterns are auditable against MCOB 13 requirements

When the FCA asks for evidence of early and meaningful engagement with customers in arrears, this data is your answer. A spreadsheet of call recordings is not.

Why Amazon Connect Is the Right Platform for This

We're AWS builders. That's not a preference, it's a deliberate choice based on what the platform actually delivers for regulated UK lenders.

What the 4 to 6 Week Build Actually Covers

Lenders sometimes ask whether 4 to 6 weeks is realistic for a regulated deployment. Here's what that timeline includes:

Full rollout happens after the pilot cohort validates the numbers. No firm goes live at scale without evidence from their own book.

Who Should Be Reading This

If you're a Head of Mortgage Servicing, a Collections Director, or a Chief Risk Officer at a UK lender, this is a production-ready capability, not a proof of concept. The compliance architecture exists. The Amazon Connect integration patterns are proven. The FCA regulatory alignment is built in, not bolted on.

If you're running a team of 20 outbound collectors who spend 40% of their time on early arrears calls that could be handled autonomously, that's a workforce planning conversation worth having.

If you're preparing your Consumer Duty board report and your evidence base for early arrears engagement is thin, this gives you structured, auditable data from day one.

Frequently Asked Questions

Rel8 CX is an AWS Advanced Partner specialising in autonomous AI voice agents for regulated UK contact centres. We build production deployments in 4 to 6 weeks with FCA compliance architecture built in.

Yes, within a defined policy matrix and with proper consent capture. The arrangement terms are read back to the customer, verbal confirmation is captured, and the arrangement is written to the loan management system via API with a timestamped audit record.

Rel8 CX deploys production-grade mortgage arrears agents in 4 to 6 weeks, including LMS integration, compliance review, UAT, and a pilot cohort before full rollout.

A properly built agent can satisfy the key Consumer Duty obligations: early contact, clear communication, vulnerability identification, debt advice signposting, and auditable outcomes. The architecture has to be built for compliance, not retrofitted.

The Bottom Line

Mortgage arrears is one of the highest-stakes conversations a lender has with a customer. It's also one of the most process-intensive, compliance-constrained, and volume-sensitive. Those three characteristics make it exactly the right candidate for an autonomous AI voice agent.

The firms that get this right will have faster first contact, higher self-cure rates, better Consumer Duty evidence, and lower cost to serve. The firms that wait will be staffing for a problem that doesn't need to be a staffing problem.

We build this. In production. In 4 to 6 weeks.