FCA guidance on AI in collections: a readiness check

The FCA published its first good and poor practice examples for AI in financial services on 19 October 2026. A collections director reading the 20 case studies will recognise most of them. Vulnerability detection, affordability assessments, letter drafting, chatbots that signpost free debt advice: the report keeps returning to the work collections teams do every day, which makes it the closest thing we have to supervisory guidance on AI in collections.

This post takes each finding that touches collections and turns it into a question to put to your own team, with the evidence you'd need to answer it.

What the FCA published

The publication covers three areas: Consumer Duty and customers in vulnerable circumstances, governance and accountability, and operational resilience. It draws on supervisory work and the FCA's AI Lab, and sets out 20 anonymised case studies alongside lists of good practice and areas for improvement.

The FCA's position is that the existing rulebook already covers AI: the Consumer Duty, the Senior Managers regime, systems and controls rules and the complaints sourcebook. The report adds no obligations. It shows how supervisors expect existing ones to be met when AI is involved, and that's the standard firms will be measured against.

The FCA also says AI used well can lower costs and make support more accessible and more tailored. Its criticism falls on firms that adopted AI carelessly, and on firms that avoided it so cautiously that customers lost out.

Where AI in collections already lives

Five use cases cover most of what lenders are doing or trialling today:

• Spotting payment stress and vulnerability from contact notes, payment behaviour and how a customer moves through a digital journey
• Suggesting arrangement terms and supporting affordability assessments
• Drafting and simplifying letters, emails and text messages
• Chatbots and assistants inside the self-service portal
• Quality checking agent conversations and complaint responses

Every one of these appears in the FCA's case studies. The report asks firms to decide, use case by use case, where AI should support a person, where it should automate and where it should hand over. It then asks them to build the checks and monitoring into the system itself, so the controls hold without someone having to catch problems at the end.

Six findings, translated for collections

Our Ofgem Consumer Outcomes readiness review asked one question of every outcome: can we prove it? The same test applies here.

1. Switching AI off for vulnerable customers can be the unfair option

Some firms excluded anyone with a vulnerability indicator from AI journeys to reduce risk. The FCA accepts this can be appropriate, and warns that it becomes unfair when the alternative route gives those customers worse support or a smaller benefit.

Picture a portal that hides the self-serve arrangement journey from flagged customers and sends them to a phone queue. The customers who most need to sort things out at 11pm are the ones told to call back in office hours.

Ask: what happens to a flagged customer, and is it as good?

Evidence: the journey design for flagged customers, and outcome comparisons between flagged and unflagged groups.

2. Human review needs its own evidence

The FCA warns against relying on staff review to compensate for weak model governance. In one case study, reviewers grew used to accepting AI recommendations and challenged them less, even when the output carried low confidence. In another, a firm using AI for affordability assessments had reviewers judge each case before seeing the AI's proposed answer, so it could compare the two.

Agents approving AI-suggested arrangement terms will approve more of them as the months go by. That drift is predictable, and the regulator expects you to measure it.

Ask: what's your override rate, has it fallen, and what triggers an escalation?

Evidence: a record per case of whether staff reviewed, approved, overrode or escalated, and why. The FCA cites a smaller firm that logs exactly this.

3. Vulnerability detection needs more than words

Firms placed too much confidence in a language model's ability to detect vulnerability from text alone, particularly temporary vulnerability. The firms praised in the report combined text signals with payment behaviour and signs of a customer struggling in a digital journey. Frontline staff helped define the indicators, and the list of proxy terms was kept live and reviewed.

Ask: which signals does the model use, who defined them, and when were they last reviewed?

Evidence: the indicator list with its version history, and the support that followed each flag.

4. Records that would survive a complaint

The FCA links auditable records directly to complaints handling. Good practice means logging prompts, outputs, model versions, filters applied and escalation decisions.

A customer says the arrangement was unaffordable or the letter was misleading. To answer, you need what they saw, which model version produced it and what the agent did next, in one place.

Ask: could you rebuild one AI-touched case end to end this afternoon?

Evidence: AI interactions logged in the same record as the collections activity they relate to, rather than in a vendor dashboard.

5. Customers shouldn't repeat themselves when they change channel

The FCA praised firms that let customers choose between AI-enabled and human support and move between them without starting again. It criticised automated journeys that gave unclear answers and offered no way out.

A customer who leaves the portal assistant halfway through setting up an arrangement should reach an agent who can already see what they'd entered.

Ask: does a customer who abandons the bot start from scratch on the phone?

Evidence: one record shared by the portal and the agent desktop.

6. You own the AI you buy

Firms remain accountable for AI supplied by third parties. The FCA flags untested vendor assurances, dependence on a single provider with no exit plan, and one firm that had to redesign a product when the model underneath it was retired. The better-run firms asked AI-specific questions during procurement and secured contractual notice of model changes.

Put those questions to your collections software supplier. Put them to your own engineering team too, for anything built in-house on a general-purpose model.

Ask: for every AI component in the collections journey, who owns it, what happens when it changes, and what's the fallback?

Evidence: an inventory of AI use cases and the models behind them, supplier assessments, and a tested contingency.

If you're a credit union or a newer lender

The FCA gives smaller firms their own examples, and the bar is proportionate. It describes smaller firms adding vulnerability markers to a sample of test cases and comparing outputs, mapping each AI use case to a named senior manager with an agreed escalation route, and starting with one narrow use case before widening.

The minimum it expects is a named owner, a threshold for when to stop, and a record showing whether anyone reviewed the outputs.

Most of the gaps are evidence gaps

Work through the six questions and most of the weak answers trace back to where the evidence lives. A vulnerability flag raised in one tool, a letter drafted in another, an arrangement agreed in a third and a complaint logged in a fourth takes a manual trawl to reconstruct for a supervisor, or for the customer.

Flexys builds collections software that holds the complete record of collections activity: every communication, arrangement, action and decision, with a full audit trail, in one system, alongside the customer self-service portal and the agent desktop. Whatever a lender decides AI should support, automate or hand over in its collections journey, that record is where the FCA's evidence expectations land.

Book a demonstration to see how Flexys evidences collections activity across the customer journey, from first contact to sustained arrangement.