AI Voice Agents for Mortgage Arrears Servicing: How UK Lenders Are Automating FCA-Compliant Early Contact on Amazon Connect
Mortgage arrears volumes in the UK rose 26% year-on-year in 2023 according to UK Finance data, and servicers are under simultaneous pressure from two directions. The FCA's Consumer Duty rules demand earlier, more empathetic contact with borrowers in financial difficulty. Operational budgets demand that collections teams do more with less headcount. Those two pressures are not compatible when you are relying on human agents to make every early-stage outbound call.
The answer a growing number of regulated UK lenders are reaching is autonomous AI voice agents running on Amazon Connect. Not a pilot. Not a proof of concept. Production systems handling thousands of arrears contacts per day, with FCA-aligned vulnerability detection built into every conversation.
This post covers exactly how those systems are architected, what compliance guardrails are non-negotiable, and what outcomes lenders are achieving in the first 90 days of operation.
Why Early Arrears Contact Is the Right Starting Point for AI Voice
Mortgage servicers typically segment arrears into early-stage (1 to 2 missed payments) and late-stage (3 or more). Early-stage contact is high-volume, relatively low-complexity, and time-sensitive. The FCA's MCOB 13 rules and the Consumer Duty guidance both emphasise that lenders must make contact promptly and treat customers fairly from the first sign of financial difficulty.
In practice, most servicers cannot achieve the contact rates they need with human-only outbound teams. A team of 20 agents making 60-call days generates 1,200 outbound attempts. A production AI voice agent deployment on Amazon Connect can execute 10,000 outbound attempts in the same window, with consistent scripting, full call recording, and real-time transcription feeding compliance dashboards.
The economics shift dramatically. Early-stage resolution rates improve because more customers are reached before arrears deepen. Late-stage caseloads shrink. Human agents are freed to handle the complex, emotionally demanding conversations where their skills are genuinely irreplaceable.
FCA Consumer Duty: What the Compliance Architecture Must Deliver
Before any lender can deploy an AI voice agent in a collections or arrears context, the compliance architecture must address four non-negotiable requirements from the FCA.
1. Vulnerability Detection in Real Time
The FCA's Financial Lives survey consistently shows that around 47% of UK adults display one or more characteristics of vulnerability. In an arrears population, that proportion is significantly higher. Consumer Duty requires firms to identify vulnerable customers and adapt their approach accordingly.
In a production AI voice agent, vulnerability detection operates at the conversation layer using real-time transcription and sentiment analysis. Specific linguistic markers trigger escalation protocols: expressions of distress, references to bereavement or illness, signs of cognitive difficulty, or explicit statements about mental health. When any of these are detected, the AI agent does not attempt to continue a collections workflow. It acknowledges the customer with a scripted empathy response and transfers to a specialist human agent, flagging the vulnerability indicators in the CRM record.
This is not a keyword list. Production systems use classification models trained on thousands of annotated calls to identify vulnerability signals in context. A customer saying they have had a difficult few months is different from a customer saying they have just lost their spouse. The system must distinguish between the two.
2. Informed Consent and Clear AI Disclosure
The FCA has not issued a blanket prohibition on AI voice agents in collections, but Consumer Duty's outcome-focused framework makes it clear that customers must understand what they are interacting with. Production deployments include an explicit disclosure at the start of every call: the customer is told they are speaking with an automated system and is given a clear option to request a human agent at any point.
This disclosure is not just a compliance box. Data from production deployments shows that transparent disclosure does not materially reduce engagement rates. Customers who are willing to discuss their arrears will engage with an AI agent that is competent and respectful. Those who are not willing will not engage with a human agent either.
3. Audit-Ready Call Records
Every interaction must be fully recorded, transcribed, and stored in a way that supports FCA supervision and internal quality assurance. Amazon Connect's native call recording and Contact Lens transcription capabilities provide the infrastructure. The compliance layer on top indexes every call by arrears stage, outcome, vulnerability flags triggered, and payment arrangement details confirmed.
When a compliance team or FCA supervisor requests evidence of how a specific customer was treated, the answer is a searchable record with full audio, transcript, sentiment scores, and the exact decision logic the AI agent followed. That is not possible with manual note-taking on human agent calls.
4. Escalation Paths That Actually Work
An AI voice agent that cannot escalate effectively is a compliance liability. Production systems define escalation triggers at the design stage: vulnerability indicators, customer requests for human contact, failed payment arrangement attempts, and any situation where the AI agent's confidence in the correct response falls below a defined threshold.
Escalation routes into Amazon Connect's queue management, with the context of the AI conversation passed to the receiving human agent via the CRM integration. The human agent does not start from scratch. They see the arrears balance, the conversation summary, the vulnerability flags, and the payment options already discussed.
The Amazon Connect Architecture for Mortgage Arrears Servicing
Rel8 CX builds these systems natively on AWS. Here is the production architecture.
Outbound Campaign Orchestration
Amazon Connect's outbound campaigns feature manages the dialling logic. Arrears accounts are segmented by days past due, product type, and previous contact history. The campaign engine applies the FCA's permitted calling windows (broadly 8am to 9pm, with additional restrictions for vulnerable customer flags) and manages retry logic to avoid over-contacting customers.
Account data from the servicing platform is pushed to Amazon Connect via an AWS Lambda function that runs on a scheduled trigger, typically every four hours during business days. This keeps the contact list current without requiring a real-time integration that introduces latency risk.
The Conversational AI Layer
The AI voice agent is built on Amazon Lex for intent recognition and Amazon Polly for voice synthesis, with custom Lambda functions handling the business logic for payment arrangement calculations, arrears balance lookups, and CRM updates. The conversational flow covers the following stages:
Identity verification: The agent confirms the customer's identity using date of birth and postcode, consistent with the lender's existing authentication policy. Failed verification after two attempts triggers a callback scheduling offer rather than a collections conversation. Arrears acknowledgement: The agent confirms the outstanding arrears amount and the number of missed payments. This information is pulled in real time from the servicing platform via a Lambda integration. Circumstances check: A structured but conversational sequence of questions explores the reason for the arrears. This is where vulnerability detection is most active. The agent is designed to listen before it proposes solutions. Payment arrangement options: Based on the customer's stated circumstances and the lender's forbearance policy, the agent presents up to three payment arrangement options. These are calculated dynamically based on the outstanding balance, the customer's indicated affordability, and the lender's arrears management rules. The agent does not offer arrangements outside the policy parameters. Arrangement confirmation: If the customer agrees to an arrangement, the agent reads back the terms, confirms acceptance, and writes the arrangement to the CRM via a Lambda function. The customer receives an SMS confirmation within two minutes via Amazon Pinpoint. Escalation or scheduling: If no arrangement is reached, the agent offers a callback at a customer-preferred time or transfers to a human agent queue.Real-Time Compliance Monitoring
Amazon Connect Contact Lens runs on every call, generating real-time transcripts and sentiment scores. A custom EventBridge rule triggers a Lambda function when vulnerability keywords are detected, updating the CRM record and alerting the compliance team's dashboard. The same pipeline feeds a daily compliance report that tracks vulnerability detection rates, escalation rates, and arrangement completion rates by arrears segment.
What Production Numbers Look Like
Based on production deployments in regulated UK lending environments, here are the outcomes lenders should model when evaluating this architecture.
Contact rate improvement: AI voice agents operating on outbound campaigns typically achieve contact rates of 35 to 45% on early arrears populations, compared to 18 to 25% for human-only outbound teams working the same lists. The difference is volume and consistency. The AI agent does not have good days and bad days. Self-service arrangement rate: Of customers who are successfully contacted and complete the full conversation flow, 40 to 55% agree to a payment arrangement without human agent involvement. This varies by arrears depth and product type. First-payment defaults on buy-to-let products have lower self-service rates than owner-occupier accounts that are one payment behind. Average handling time: The AI agent completes a full arrears contact including identity verification, circumstances check, and arrangement confirmation in an average of 4.5 to 6 minutes. Human agents handling equivalent calls average 12 to 18 minutes, largely because of note-taking and system navigation time. Vulnerability escalation rate: In production deployments, between 8 and 14% of calls trigger a vulnerability escalation. This is consistent with FCA expectations for an arrears population and provides documented evidence that the vulnerability detection system is functioning. Time to production: Rel8 CX builds and deploys production-ready mortgage arrears AI voice agents in 4 to 6 weeks. That includes the Amazon Connect configuration, the conversational AI build, the CRM integration, compliance testing, and the first wave of live calls.The Integration Layer: What Has to Connect
A mortgage arrears AI voice agent is only as good as its integrations. The three systems that must be connected for a production deployment are the loan servicing platform, the CRM, and the telephony infrastructure.
Loan servicing platform: The agent needs real-time read access to arrears balances, payment history, and account status. It needs write access to record arrangement outcomes. In UK mortgage servicing, common platforms include Phoebus, Nivo, and legacy systems running on SQL databases. Rel8 CX builds the integration layer using AWS Lambda and API Gateway, with the data model normalised to a standard schema regardless of the source system. CRM: Vulnerability flags, call outcomes, and escalation triggers must be written to the CRM in real time. For most UK lenders this means Salesforce Financial Services Cloud or Microsoft Dynamics 365. Both have well-documented APIs that support the required write operations. SMS confirmation: Amazon Pinpoint handles outbound SMS for arrangement confirmations. This requires a UK sender ID registration and compliance with the ICO's PECR rules for transactional messaging.Common Objections and How Production Experience Answers Them
"Our customers will not engage with an AI agent for something this sensitive."
Production data does not support this concern for early-stage arrears. Customers who are one payment behind are often embarrassed and prefer the lower-stakes interaction with an AI agent to a conversation with a human collector. Engagement rates for early-stage AI outbound are consistently higher than for late-stage human outbound on the same population.
"We cannot get FCA sign-off on an AI system."
FCA approval is not required for deploying AI in arrears servicing. What is required is demonstrable compliance with Consumer Duty outcomes, MCOB 13, and the FCA's guidance on fair treatment of customers in financial difficulty. A production system with full audit trails, documented vulnerability detection, and clear escalation paths is more defensible to an FCA supervisor than a manual process with inconsistent agent notes.
"Our IT team cannot support an Amazon Connect deployment."
Amazon Connect is a managed cloud service. The operational burden on an internal IT team is significantly lower than for on-premises telephony infrastructure. Rel8 CX manages the build and provides documentation and training for internal teams to operate the system post-deployment.
What to Do Before You Start
Before engaging a build partner, lenders should complete three internal steps.
First, document the current arrears contact process in detail, including the scripting used by human agents, the escalation criteria for vulnerability, and the forbearance options available at each arrears stage. The AI agent will encode this logic, so it must be explicit before the build starts.
Second, audit the data quality in the loan servicing platform. The AI agent's effectiveness depends on accurate arrears balances and current contact details. A data quality exercise before the build saves significant rework during integration testing.
Third, identify the compliance sign-off process internally. The legal and compliance team will need to review the conversation flows, the vulnerability detection criteria, and the disclosure language before the system goes live. Building this review into the project timeline from day one prevents delays at the end.
The Case for Building Now
The FCA's Consumer Duty implementation deadlines have passed. Lenders that cannot demonstrate proactive, empathetic early contact with customers in financial difficulty are exposed to supervisory scrutiny. At the same time, arrears volumes are rising and collections headcount is not growing proportionally.
An autonomous AI voice agent on Amazon Connect is not a future-state aspiration. It is a production system that regulated UK lenders are running today. The architecture is proven, the compliance framework is documented, and the outcomes are measurable.
We build these systems. We have done it in regulated lending environments. We can have a production system running in 4 to 6 weeks.
Book a discovery callIs your pilot going to reach production?
Fifteen questions, three minutes, no cost. You get a score against the ten checks we run every deployment through, and a straight answer on what is blocking yours.
Find out what is blocking you