FCA Consumer Duty and AI Voice Agents: What Debt Collection, Mortgage Servicing and Insurance Firms Must Get Right

Arkadas Kilic 8 min read · 2026-07-06

The FCA's Consumer Duty is not a checkbox exercise. Since it came into full force in July 2023 for open products and July 2024 for closed book products, it has fundamentally changed the standard firms must meet when interacting with customers. And now that AI voice agents are moving from proof of concept into production contact centres, the compliance stakes are higher than ever.

For firms in debt collection, mortgage servicing and insurance, the question is no longer whether to deploy AI voice agents. The question is whether the systems you deploy can demonstrably deliver good outcomes under Consumer Duty scrutiny. The FCA has been explicit: the duty applies regardless of the channel. If an AI voice agent is handling a vulnerable customer in arrears, the same obligations apply as if a human agent were on the call.

This post covers the specific requirements firms must build into production AI voice systems to stay on the right side of the regulator and, more importantly, to actually serve customers well.

What Consumer Duty Actually Requires of Automated Voice Interactions

Consumer Duty rests on four outcome areas: products and services, price and value, consumer understanding, and consumer support. All four are directly implicated when you deploy an AI voice agent in a regulated context.

The FCA expects firms to take reasonable steps to understand the needs of customers, including those with characteristics of vulnerability, and to respond appropriately. In a voice interaction, that means the system must:

• Detect signals of vulnerability in real time
• Adapt its communication style and pace accordingly
• Escalate to a human agent when the situation demands it
• Record and retain evidence that good outcomes were pursued

The FCA's Financial Lives survey found that 47% of UK adults show one or more characteristics of vulnerability. In debt collection and mortgage arrears portfolios, that proportion is materially higher. A production AI voice agent that cannot identify and respond to vulnerability is not Consumer Duty compliant, full stop.

The Three Sectors Where the Risk Is Highest

Debt Collection

Debt collection is the highest-risk environment for AI voice deployment under Consumer Duty. Customers in arrears are disproportionately likely to be experiencing financial difficulty, mental health challenges or other vulnerability indicators. The FCA's guidance on collections is explicit about the need for forbearance and appropriate treatment.

An AI voice agent in collections must be able to:

• Recognise distress signals in speech, including pace changes, long silences, emotional language and explicit statements of difficulty
• Pause automated payment prompts when vulnerability is detected
• Route to a specialist human agent within a defined threshold, typically under 30 seconds from detection
• Offer appropriate breathing space referrals under the Breathing Space scheme, which gives eligible customers 60 days of legal protection from creditor action
• Log every interaction with a structured audit trail that maps to Consumer Duty outcome evidence requirements

Firms that deploy collections voice agents without these capabilities are not just taking regulatory risk. They are taking the risk of causing real harm to customers in difficult circumstances.

Mortgage Servicing

Mortgage servicing sits at the intersection of Consumer Duty and the FCA's existing Mortgage Conduct of Business (MCOB) rules. For customers in arrears or approaching the end of a fixed rate, the stakes around communication quality are significant.

The FCA expects firms to contact customers proactively about rate changes and to ensure those customers understand their options. An AI voice agent handling mortgage servicing calls must:

• Confirm customer understanding, not just deliver information. Consumer Duty requires firms to test whether communications actually land, not just whether they were sent.
• Provide clear, accurate information about product options without creating the impression of a personal recommendation unless the firm is authorised to provide advice
• Identify when a customer's circumstances have changed in a way that requires human review, for example a customer who mentions they have lost their job or separated from a partner
• Maintain a complete interaction record that can be produced in response to an FCA data request or a Subject Access Request

The FCA fined a major lender 7.7 million pounds in 2023 partly for failures in mortgage arrears handling. The standard is not theoretical.

Insurance

Insurance presents a specific challenge because of the complexity of policy information and the consequences of misunderstanding. A customer who cancels a policy, makes a claim or queries their cover based on incorrect information from an AI voice agent creates both a Consumer Duty failure and a potential liability.

For insurance firms, AI voice agents must:

• Accurately represent policy terms without simplification that creates material misunderstanding
• Flag when a customer's query requires human review, particularly around claims, exclusions and renewals
• Support fair value assessments by ensuring customers understand what they are paying for
• Comply with the FCA's rules on renewal communications, including the requirement to display the previous year's premium for general insurance customers
• Handle complaints correctly, including logging them as complaints even when the customer does not use the word "complaint"

The Five Technical Capabilities That Separate Compliant from Non-Compliant Systems

1. Real-Time Vulnerability Detection

This is the capability most firms underestimate. Vulnerability detection in a live voice interaction requires acoustic and linguistic analysis running in parallel with the conversation, with a latency low enough to influence the call before harm occurs.

Production systems we build on Amazon Connect use real-time transcription combined with intent classification to flag vulnerability indicators within 3 to 5 seconds of detection. The system does not wait for the end of the call to analyse sentiment. It acts during the call.

The categories to detect include:

• Explicit statements of financial difficulty or mental health challenges
• Confusion or repeated requests for clarification
• Emotional distress markers in speech patterns
• Mentions of bereavement, job loss or relationship breakdown
• Cognitive difficulty indicators such as very slow speech or inability to follow simple instructions

2. Structured Escalation with Context Transfer

Escalation is not just a routing event. Under Consumer Duty, the human agent who receives an escalated call must have the context to continue the interaction without making the customer repeat themselves. Repeating sensitive information, particularly about vulnerability, causes harm and creates a poor outcome.

Production systems must pass a structured summary to the receiving agent in real time, including the vulnerability indicators detected, the stage of the conversation reached, any commitments made by the AI voice agent and the customer's stated preferences. This is not optional functionality. It is a Consumer Duty requirement.

3. Interaction Logging at the Outcome Level

Most firms log calls. Fewer log calls in a way that maps to Consumer Duty outcome evidence requirements. The FCA expects firms to be able to demonstrate, not just assert, that good outcomes were delivered.

That means logging must capture:

• What information was provided and when
• Whether the customer confirmed understanding
• What options were offered
• What the customer chose and why
• Any vulnerability indicators detected and how the system responded
• The outcome of the interaction against the firm's defined good outcome criteria

This data must be retained for a minimum of five years for most regulated products and must be retrievable in a format that supports regulatory review.

4. Guardrails Against Harmful Outcomes

AI voice agents in regulated environments must have hard guardrails that cannot be overridden by the model's own judgement. These include:

• No collection of payment commitments from customers who have indicated vulnerability without a human review step
• No presentation of credit products to customers who have indicated financial difficulty without an affordability check
• No call termination by the AI agent when a customer is in distress
• Mandatory signposting to free debt advice services (StepChange, National Debtline, Citizens Advice) in collections interactions where the customer cannot agree a payment arrangement

These guardrails are implemented at the orchestration layer, not left to the model's discretion.

5. Audit-Ready Infrastructure

The FCA can request interaction data, outcome evidence and system documentation. Firms must be able to respond within the FCA's standard information request timelines, which can be as short as 14 days.

Production AI voice systems must be built with audit readiness as a design requirement, not a retrofit. That means:

• All interaction data stored in encrypted, immutable logs
• Clear data lineage from customer interaction to outcome classification
• System documentation that describes how the AI makes decisions, sufficient to satisfy the FCA's explainability expectations
• Regular outcome monitoring with defined thresholds that trigger human review

What the FCA's Supervisory Approach Means in Practice

The FCA has been clear that it will use data to identify firms where customer outcomes are poor and investigate proactively. The Consumer Duty requires firms to produce an annual board report assessing whether they are delivering good outcomes. That report must be based on evidence, not assertion.

For firms with AI voice agents in production, that means the monitoring infrastructure must be in place before the agents go live. Deploying first and building compliance infrastructure later is not an acceptable approach under Consumer Duty. The FCA's position is that firms must be able to demonstrate compliance from the point of deployment.

Firms that cannot produce outcome evidence for their AI voice interactions are in a materially worse position than firms that have no AI voice capability at all. Deploying a non-compliant system at scale creates evidence of systemic failure.

The Build Approach That Gets This Right

At Rel8 CX, we build production AI voice systems for regulated financial services firms on AWS and Amazon Connect. The compliance architecture is not an add-on. It is built into the system from day one.

Our standard engagement for a compliant AI voice agent in debt collection, mortgage servicing or insurance runs 4 to 6 weeks to production. That timeline includes:

• Vulnerability detection and escalation logic
• Consumer Duty outcome logging
• Guardrail implementation at the orchestration layer
• Audit trail infrastructure on AWS
• Integration with existing CRM and case management systems
• Testing against defined good outcome criteria before go-live

Firms that have attempted to build this capability internally typically spend 6 to 12 months and still lack the outcome monitoring infrastructure the FCA expects. The difference is that we have built these systems before, in regulated environments, and we know where the compliance gaps appear.

The Bottom Line

Consumer Duty does not prohibit AI voice agents in regulated financial services. It requires that those agents deliver good outcomes, treat vulnerable customers appropriately and generate the evidence to prove it.

Firms that get this right will have a significant operational advantage: lower cost per interaction, faster resolution, consistent compliance and the ability to serve customers at scale without sacrificing quality. Firms that get it wrong will face regulatory action, customer harm and reputational damage.

The technology is ready. The compliance framework is clear. The question is whether you build it correctly.