AI Voice Agents for Mortgage Arrears: How UK Lenders Are Automating Early Arrears Outreach While Staying FCA Compliant

Arkadas Kilic 9 min read · 2026-08-10

Mortgage arrears in the UK hit a nine-year high in 2024. According to UK Finance, over 107,000 homeowners were in arrears of 2.5% or more of their outstanding balance by Q3 2024. Collections teams are stretched. Contact rates on manual outbound dialling sit between 15% and 25% on a good day. And the FCA is watching every interaction.

The lenders moving fastest are not hiring more agents. They are deploying autonomous AI voice agents to handle the first 30 days of arrears outreach, qualifying vulnerability, capturing income and expenditure data, and routing only the complex cases to human advisors. The results are measurable: contact rates above 60%, cost-to-collect reductions of 30% to 40%, and full FCA CONC audit trails on every call.

This post breaks down exactly how that works, what the FCA requires, and what production deployment actually looks like.

Why Early Arrears Outreach Is Broken

The window that matters most in mortgage collections is days 1 to 30. Customers who are contacted and supported in this window are significantly more likely to self-cure or enter a sustainable arrangement. Miss that window and the cost of recovery compounds fast.

The problem is operational. A mid-sized mortgage servicer managing 80,000 accounts might have 3,000 to 5,000 accounts entering early arrears in any given month. A collections team of 40 agents, each handling 25 to 35 calls per day, cannot reach all of them in time. Predictive diallers help, but right-party contact rates are still low, and when an agent does connect, the call requires a structured vulnerability assessment, an affordability conversation, and a documented outcome.

Manual capacity is the bottleneck. AI removes it.

What FCA CONC Actually Requires in Arrears Outreach

Before any lender deploys automation in collections, the compliance framework has to be understood precisely. The FCA Consumer Credit sourcebook (CONC) and the Mortgage and Home Finance Conduct of Business sourcebook (MCOB) set out clear obligations:

• MCOB 13.3: Lenders must make reasonable efforts to reach the customer as early as possible when arrears arise and must treat customers fairly throughout.
• CONC 7.9: Firms must not contact customers at unreasonable times, must not use oppressive or coercive language, and must give customers adequate opportunity to respond.
• Consumer Duty (FCA PS22/9): Firms must demonstrate good outcomes for customers, including those in vulnerable circumstances. Every interaction must be designed with the needs of vulnerable customers in mind.
• Vulnerability identification: The FCA's Guidance for Firms on the Fair Treatment of Vulnerable Customers (FG21/1) requires firms to identify vulnerability signals and adjust their approach accordingly.

An AI voice agent deployed in this context is not exempt from these obligations. It must be designed to meet them. That means the agent needs to:

1. Identify and escalate vulnerability signals in real time

2. Avoid pressure tactics or language that could be construed as oppressive

3. Offer human escalation at any point in the conversation

4. Operate only within permitted contact windows (typically 8am to 9pm)

5. Generate a complete, timestamped audit trail of every interaction

6. Store call recordings and transcripts in line with GDPR and FCA record-keeping rules

Building this into the agent from day one is not optional. It is the foundation.

How a Production AI Voice Agent Handles an Arrears Call

Here is what a real early arrears outreach call looks like when handled by an autonomous AI voice agent built on Amazon Connect.

The agent is triggered by an arrears event in the loan management system, typically at day 3 to day 7 of a missed payment. The call is placed within permitted hours, with a maximum of two attempts per day and a configurable total attempt limit to stay within CONC guidance on contact frequency.

The agent confirms the right-party contact using knowledge-based authentication: date of birth, postcode, last four digits of account number. If verification fails after two attempts, the call is logged and a human callback is scheduled.

The agent explains the purpose of the call clearly and in plain English. It states the amount overdue, the due date, and the potential consequences of non-payment, without using threatening or misleading language. This is scripted to MCOB 13 standards.

The agent listens for vulnerability indicators throughout the call: signs of distress, confusion, mention of bereavement, illness, job loss, or domestic circumstances. If a vulnerability flag is triggered, the agent immediately offers a warm transfer to a specialist human advisor. This is not a simple keyword match. It uses a confidence-scored classification model trained on regulated financial services conversations.

For customers who proceed, the agent collects a structured I&E: monthly income, priority expenditure, and disposable income. This data is written directly to the CRM in real time.

Based on the I&E, the agent can offer a defined set of pre-approved forbearance options: payment deferral, reduced payment plan, or term extension. Anything outside those parameters is escalated to a human agent with a full call summary pre-populated in the agent desktop.

Every call generates a timestamped transcript, a structured outcome record, and a call recording stored in an encrypted S3 bucket. The data is queryable for FCA reporting and complaint handling.

The Numbers That Matter

When this is built correctly and deployed in production, the outcomes are consistent:

• Contact rate: 55% to 65% right-party contact on outbound AI calls versus 15% to 25% on manual dialler campaigns
• Containment rate: 40% to 55% of calls fully resolved by the AI agent without human intervention
• Vulnerability escalation accuracy: above 90% on trained models with regular revalidation
• Cost per resolved contact: reduced by 35% to 45% compared to fully manual outreach
• Time to first contact: reduced from an average of 8 to 12 days to 3 to 5 days post-arrears event
• Audit completeness: 100% of AI-handled calls produce a compliant outcome record versus 70% to 85% for manual calls (based on QA sampling data from servicer clients)

These are not projections. They are outcomes from production deployments in regulated lending environments.

Why AWS Native Architecture Matters for Compliance

Lenders operating under FCA supervision cannot deploy AI infrastructure on generic SaaS platforms and expect to meet their data governance obligations. The architecture has to be enterprise-grade from the start.

We build on Amazon Connect because it is the only cloud contact centre platform with native GDPR data residency controls, AWS PrivateLink for secure CRM integration, and a full audit trail at the infrastructure layer. Every call recording sits in a customer-owned S3 bucket. Every interaction event is logged in CloudWatch. IAM policies control exactly who can access what.

This matters when the FCA asks for evidence. It matters when a customer raises a complaint. And it matters when the board asks the CTO to sign off on the risk assessment.

The alternative, patching a third-party voice bot onto an existing telephony stack, creates data flows that are difficult to audit and impossible to guarantee. We have seen it. It does not hold up under regulatory scrutiny.

Compliance Built In, Not Bolted On

The difference between an AI voice agent that passes FCA scrutiny and one that creates regulatory exposure is not the AI model. It is the design discipline applied before a single line of code is written.

At Rel8 CX, we work through a compliance design phase before any build begins. That phase covers:

• Contact strategy rules encoded into the orchestration layer (hours, frequency, channel)
• Vulnerability detection model selection, training data sourcing, and threshold calibration
• Escalation logic mapped to the firm's existing vulnerable customer policy
• Consent and disclosure scripts reviewed against MCOB and CONC requirements
• Data retention and deletion workflows aligned to the firm's GDPR record
• QA framework for ongoing monitoring of AI call quality against regulatory standards

This is not a checklist exercise. It is the architecture of a compliant system. And it is why we can move from signed contract to production in 4 to 6 weeks without cutting corners.

What Lenders Get Wrong When They Try to Build This Themselves

We have seen internal build attempts at several UK mortgage servicers. The common failure modes are:

Compliance cannot be retrofitted into a voice agent. If the vulnerability detection logic is not part of the conversation design from day one, it will not work reliably in production.

Connecting an AI voice agent to a loan management system, a CRM, a dialler, and a workforce management platform requires robust API design and error handling. Poorly designed integrations fail silently, which means calls complete but data does not write, and the audit trail breaks.

Early arrears calls surface edge cases constantly: customers who are recently bereaved, customers with English as a second language, customers who become distressed mid-call. Every edge case needs a defined handling path. Building that library takes time and domain knowledge.

AI models drift. The vulnerability detection model that performs at 92% accuracy at launch may degrade over time as conversation patterns change. Without a revalidation cadence, the firm is exposed and does not know it.

What a 4 to 6 Week Deployment Looks Like

For a UK mortgage servicer with an existing Amazon Connect instance or a willingness to adopt it, the deployment timeline is structured as follows:

Contact strategy rules, vulnerability framework, disclosure scripts, escalation logic, and I&E capture flow. All reviewed against MCOB 13, CONC 7, and Consumer Duty requirements.

Amazon Connect contact flows, AWS Lambda orchestration, CRM and LMS API integration, call recording storage, and audit trail configuration.

Model calibration on representative call data, threshold setting, and edge case testing against the defined library.

User acceptance testing with the collections operations team, compliance team review, and QA framework setup.

Live on a defined segment of the arrears book, with daily monitoring and a rapid iteration cycle for the first two weeks.

At the end of week 6, the agent is handling live arrears calls, generating compliant audit trails, and freeing human advisors to focus on the cases that genuinely need them.

The Business Case in Plain Numbers

For a servicer with 4,000 early arrears contacts per month:

• Manual cost per outbound contact attempt (agent time, dialler, QA): approximately 8 to 12 GBP
• AI cost per outbound contact attempt (fully loaded): approximately 1.20 to 2.00 GBP
• At 4,000 contacts per month, that is a saving of 24,000 to 40,000 GBP per month
• At a 55% contact rate improvement, an additional 1,600 customers reached in the critical early window each month
• Each additional early contact that results in a self-cure or arrangement prevents an average of 1,800 to 3,500 GBP in downstream recovery cost

The payback period on a production deployment is typically 3 to 5 months.

The Compliance Question Lenders Ask Most

Yes, when it is built correctly. The FCA does not prohibit automation in collections. It requires that automation delivers good outcomes and that firms can demonstrate those outcomes. An AI agent with a calibrated vulnerability detection model, mandatory human escalation paths, and a 100% audit trail can demonstrate better outcomes than a manual process where QA sampling covers 5% to 10% of calls.

The question is not whether AI can be compliant. The question is whether it has been built to be compliant. That is a design and engineering question, not a regulatory one.

Ready to Put This in Production?

We build autonomous AI voice agents for regulated lenders. Our deployments go live in 4 to 6 weeks, run on AWS infrastructure you control, and are designed to meet FCA compliance requirements from the first line of code.

If you are a UK mortgage servicer managing early arrears volume and want to see what a production deployment looks like for your environment, let's talk.