Agentic AI in Debt Collection: How UK Contact Centres Can Recover More Without Breaching FCA Consumer Duty

Arkadas Kilic 8 min read · 2026-09-11

Debt and collections contact centres in the UK are operating in a vice grip. On one side: recovery targets that demand higher contact rates, faster resolution, and better right-party contact (RPC). On the other: FCA Consumer Duty, which came fully into force in July 2023 and fundamentally changed what acceptable collections practice looks like.

Most contact centres are trying to solve this with more agents, more compliance training, and more QA overhead. That approach is expensive, slow, and does not scale.

The contact centres that will outperform over the next three years are the ones deploying autonomous AI voice agents that handle outbound and inbound collections calls at scale, while keeping every interaction within FCA guardrails by design, not by hope.

This post is a practical breakdown of how that works.

The Collections Problem in Numbers

Before discussing the solution, it is worth grounding this in the actual operational reality facing UK collections teams.

• Right-party contact rates across the industry average between 15% and 25% on outbound dialling campaigns. That means 75% to 85% of agent time is spent on no-answers, wrong numbers, and voicemails.
• Cost per collected pound in outsourced collections typically runs between £0.18 and £0.35 depending on debt age and product type.
• Agent attrition in collections contact centres runs at 35% to 50% annually in the UK, driven by call difficulty, compliance pressure, and burnout.
• FCA supervisory focus on Consumer Duty has already resulted in enforcement action and skilled persons reviews for firms where collections practices did not demonstrate fair treatment, adequate vulnerability identification, or appropriate forbearance options.

The math is brutal. You need more contacts, higher quality conversations, full compliance documentation, and you need to do it with a workforce that is constantly churning.

What Agentic AI Actually Means in a Collections Context

The word "agentic" gets used loosely. In a collections context, it has a specific meaning.

A traditional IVR or scripted voice bot follows a decision tree. It cannot adapt. It cannot handle objections. It cannot identify vulnerability signals in real time and pivot the conversation accordingly.

An agentic AI voice agent operates differently. It:

• Conducts dynamic conversations rather than following a fixed script, adjusting based on what the customer says
• Makes decisions autonomously within defined guardrails, such as offering a payment plan within pre-approved parameters without escalating to an agent
• Identifies vulnerability signals in speech patterns, tone, and stated circumstances, and routes or adapts accordingly
• Logs every interaction with full transcript, sentiment analysis, and compliance checkpoints for QA and regulatory audit
• Integrates with core systems in real time, pulling account data and writing back outcomes to your collections management system

This is not automation of simple tasks. This is autonomous handling of complete collections interactions, end to end, for a significant proportion of your portfolio.

FCA Consumer Duty: What It Actually Requires from Collections

Consumer Duty is not a checklist. It is an outcomes-based standard. The FCA expects firms to demonstrate that collections practices deliver good outcomes for customers, specifically across four outcome areas:

1. Products and services are fit for purpose

2. Price and value are fair relative to the benefit received

3. Consumer understanding is supported, meaning customers understand their options

4. Consumer support is accessible and effective, including for vulnerable customers

For collections specifically, this translates into several operational requirements that AI agents must be designed to meet:

A well-built agentic AI system is not just compatible with these requirements. It is structurally better positioned to meet them than a human agent workforce, because compliance is built into the system rather than depending on individual agent behaviour.

How We Build Agentic Collections Agents: The Architecture

At Rel8 CX, we build these systems on AWS, using Amazon Connect as the contact centre platform. Here is how a production agentic collections deployment is structured.

Outbound Campaign Orchestration

The agent initiates outbound calls against a prioritised contact list, managed through Amazon Connect outbound campaigns. Call pacing is controlled to comply with Ofcom calling rules and your own dialling policy. No manual dialler configuration required.

Real-Time Conversation Handling

When a call connects, the AI agent conducts the conversation using a combination of natural language understanding and a structured decision framework. The agent:

• Confirms right-party contact using identity verification questions
• States the purpose of the call in plain language
• Presents the account balance and last payment information drawn live from your collections system
• Offers payment options within pre-approved parameters
• Handles common objections including "I cannot afford this", "I dispute this debt", and "I need more time"
• Identifies vulnerability signals and either adapts the conversation or transfers to a specialist agent with full context

Compliance Guardrails Built In

Every conversation operates within a compliance framework that is configured at build time and cannot be overridden at runtime. This includes:

• Mandatory forbearance disclosure at defined points in the conversation
• Free debt advice signposting (StepChange, National Debtline, Citizens Advice) triggered by specific customer statements
• Call frequency controls enforced at account level, not just campaign level
• Breathing space check against the national breathing space register before contact is attempted
• Vulnerability flag escalation that routes to a human agent with a warm handoff and full transcript

Full Interaction Logging

Every call is logged with:

• Full audio recording
• Word-level transcript
• Sentiment scores at conversation turn level
• Compliance checkpoint pass/fail for each required disclosure
• Outcome code written back to your collections management system

This gives your compliance team a complete audit trail and enables automated QA across 100% of interactions, not the 2% to 5% sample that manual QA typically covers.

What Outcomes Look Like in Production

Based on production deployments in regulated collections environments, these are the performance ranges we see:

The RPC-to-PTP conversion rate is comparable to human agents because the AI agent is not worse at the negotiation. It is consistent, patient, and never has a bad day. The contact rate improvement is where the compounding effect happens. More right-party contacts at the same or lower cost per interaction means materially more recoveries without adding headcount.

The Vulnerability Question: Can AI Handle This?

This is the question we get most often from compliance and operations leaders, and it is the right question to ask.

The honest answer is: an AI agent should not attempt to conduct a full vulnerability assessment or provide debt counselling. That is not what it is built for. What it is built for is systematic vulnerability detection and escalation, which is something human agents do inconsistently.

The AI agent is trained to detect signals including:

• Explicit statements of financial hardship, illness, bereavement, or job loss
• Emotional distress indicators in speech patterns
• Statements suggesting the customer does not understand the situation
• Requests for more time to seek advice

When any of these signals are detected, the agent pauses the collections conversation, acknowledges the customer's situation, signposts to free debt advice, and offers a warm transfer to a specialist. The transfer includes the full conversation transcript so the specialist agent does not require the customer to repeat themselves.

This is more consistent than human agents. It happens on every call, not on the calls where the agent remembers to follow the vulnerability protocol.

Timeline: From Decision to Production

One of the reasons collections teams hesitate on AI is the assumption that deployment takes 12 to 18 months. That is the timeline for bespoke enterprise software projects. It is not the timeline for a focused, well-scoped agentic AI deployment.

We put production systems live in 4 to 6 weeks. Here is what that looks like:

This is a production deployment, not a proof of concept. At the end of week 6, the system is handling real calls on real accounts.

What This Is Not

It is worth being direct about the boundaries.

Agentic AI voice agents are not a replacement for your entire collections operation. Complex disputes, legal proceedings, insolvency situations, and high-value accounts will always require human expertise. The AI handles the volume, the routine, and the first contact. Your agents handle the complexity.

This is also not a technology you buy off a shelf and configure yourself. The compliance framework, the conversation design, the integration architecture, and the vulnerability logic require practitioners who understand both the technology and the regulatory environment. Getting this wrong in a regulated collections context is not a minor issue.

The Compliance Dividend

There is a framing that compliance teams sometimes resist: AI as a compliance tool, not just a collections tool.

Consider what 100% interaction logging with automated compliance checkpoint scoring means for your Consumer Duty attestation. Instead of sampling 3% of calls and hoping the sample is representative, you have evidence across every interaction. When the FCA asks you to demonstrate that your collections practices deliver good outcomes, you have the data to show it.

That is not a secondary benefit. For firms operating under close FCA supervision, that audit trail capability is worth significant risk reduction.

Getting Started

If you are running a collections contact centre in the UK and you are trying to improve recovery rates without increasing compliance risk, the starting point is a structured assessment of where AI agents can be deployed in your specific workflow.

Not every portfolio segment is the right first deployment. Not every integration is straightforward. The right approach is to scope carefully, build correctly, and scale from a production baseline.

We build enterprise-grade agentic AI systems for regulated collections environments. We go live in 4 to 6 weeks. And compliance is built in from day one, not added on at the end.