AI voice for debt recovery changed when FCA Consumer Duty came into force. The obligations are no longer soft expectations about treating customers fairly — they are binding outcome requirements, with evidencing standards, vulnerable-customer identification duties, and a four-outcome framework that every automated touchpoint must demonstrably serve.
For debt recovery operations, this creates a paradox that most operators have not yet resolved. The traditional argument against AI voice in collections was that it lacked the human judgment to navigate difficult conversations. Consumer Duty does not rehabilitate that argument. What it does is raise the evidencing bar so high that the human argument becomes harder to sustain: can you evidence consistent fair treatment across 50,000 calls per month made by agents under contact-rate pressure, at the end of a shift, without commission bias? AI voice can, if you build it correctly.
This post sets out the Consumer Duty-compliant design for AI voice in debt recovery: what the Duty requires, what AI voice can evidence that human-only operations struggle to, where the architecture must be hardened, and what failure modes break compliance before the first FCA supervisory review.
This guide is published by the team behind Dilr Voice — enterprise voice AI deployed in regulated financial services, collections, and customer operations. Or see our AI operating model consulting for FCA-regulated deployments.
What Consumer Duty actually requires from debt recovery operations
FCA Consumer Duty is not a checklist. It is a principle-based framework structured around four outcomes that firms must deliver and evidence at every customer interaction — including automated ones.
The four outcomes, applied to debt recovery:
| Outcome | What it means for debt recovery | FCA's evidencing expectation |
|---|---|---|
| Products and Services | Repayment plans and forbearance arrangements are designed for the customer's actual circumstances, not optimised for recovery speed | Data showing individual circumstances were assessed, not just applied a blanket plan |
| Price and Value | Charges, interest, and fees are proportionate and clearly communicated; no hidden cost escalation through the recovery process | Disclosure logs, cost communication records at each contact |
| Consumer Understanding | Customers understand what they owe, what options are available, and what the consequences of each choice are — in language they can follow | Comprehension confirmation at call, complaint rates, repeat-enquiry rates on same topic |
| Consumer Support | Customers can reach support at the right stage, especially when vulnerable; no channel barriers between the customer and the help they need | Escalation rate data, time-to-human metrics, vulnerability flag outcomes |
These are not aspirational standards. The FCA's April 2026 response to the Treasury Select Committee explicitly named evidence of fair treatment at scale as a supervisory priority — and singled out AI deployments as a focus area for checking whether automated processes produce consistent outcomes or introduce systematic bias.
Debt recovery operations that deploy AI voice without designing for each of these four outcome areas are not Consumer Duty compliant. The technology does not confer compliance — architecture does.
Why AI voice can actually help Consumer Duty compliance
The instinctive reaction among many compliance teams is that AI voice in debt recovery is a Consumer Duty risk to be managed. That is half right. Poorly designed AI voice is a Consumer Duty risk. Correctly designed AI voice solves several of the hardest Consumer Duty challenges that human-only operations face.
Consistent treatment at scale. Consumer Duty requires firms to evidence fair treatment across their entire customer base, not just in a sample of monitored calls. A human agent delivering 40 calls per day, five days per week, under contact-rate pressure produces a dataset that is structurally inconsistent — the tone changes, the offer framing changes, the escalation threshold varies by agent mood and shift position. AI voice applies the same script logic, the same affordability routing, and the same forbearance triggers on call 1 and call 50,000.
100% call recording and transcript. Consumer Duty evidencing demands that firms can retrieve the full record of an interaction — what was said, what was offered, what the customer said in response. Human call recording at 100% rate is expensive and creates storage and retrieval overhead. AI voice deployments are already generating transcripts, call summaries, and disposition codes on every single call as a native output. That transcript is the evidencing artefact.
Vulnerability detection at scale. The Duty's vulnerability provisions require firms to identify and respond to customer vulnerability. For human agents, this depends on individual training, alertness, and willingness to record a vulnerability flag when contact-rate pressure argues against it. AI voice with properly engineered vulnerability detection applies the same detection logic across every call and generates an auditable record of what was detected and what response was triggered.
Outcome analytics from the full call population. Consumer Duty monitoring requires firms to track outcomes — not just process metrics like contact rate. AI voice programmes generate structured call-level data (intent, outcome, escalation, customer response) that feeds outcome analytics. A human-only operation must QA-sample to approximate this; a well-designed AI voice programme produces it as a byproduct of every call.
AI voice does not pass Consumer Duty by existing. But it creates the data infrastructure that Consumer Duty evidencing requires — provided the conversation design and the escalation architecture are built for it from the start.
For the broader commercial context of AI voice in UK financial services collections, see our guide to AI voice in fintech collections and KYC verification — which covers the pre-Consumer Duty collections landscape and the FCA regulatory baseline that Duty now extends.
The vulnerability detection architecture
Consumer Duty's vulnerability obligation is the hardest to operationalise at scale. The FCA's guidance is clear: firms must identify signs of vulnerability and respond to them. For voice AI, this means building a detection layer that identifies vulnerability signals during the call and triggers an appropriate response in real time — not as a post-call QA function.
Three-tier detection protocol:
Caller mentions health condition, carer responsibility, or financial hardship. Agent adjusts tone, offers payment helpline reference, logs advisory flag.
Caller exhibits confusion, distress, or crisis language. Agent pauses recovery conversation, offers immediate referral to specialist team, logs alert flag with call segment.
Caller mentions harm, self-harm ideation, or acute crisis. Agent stops immediately, warm-transfers to human with live context, escalation logged as priority incident.
What the detection layer must cover:
The FCA's vulnerability guidance categories four types: health, life events, resilience, and capability. For voice AI, the detection architecture should cover each:
- Health signals: explicit mention of illness, medication, mental health conditions; speech pattern changes (unusual pace, confusion, repetition)
- Life event signals: keywords around bereavement, job loss, relationship breakdown, housing instability
- Resilience signals: expressions of overwhelming stress, being "at breaking point," inability to cope
- Capability signals: repeated misunderstanding of the same point, requests for re-explanation, explicit statement of difficulty understanding
The detection logic does not have to be perfect — it has to be better than inconsistently trained humans under contact pressure. The FCA does not require zero false negatives. It requires evidence that the system is designed to identify vulnerability and that the response protocol is documented and consistently applied.
Post-call vulnerability tagging is equally important. Every call should exit with a vulnerability flag (none / advisory / alert / critical) attached to the account record and the call transcript. This creates the monitoring dataset that Consumer Duty supervisory review will ask to see: what proportion of calls triggered vulnerability detection, what was the response, and what happened to those customers' accounts subsequently.
Conversation design for Consumer Duty: what the script must do
Consumer Duty is not only an architecture problem — it is a conversation design problem. The script logic for a Consumer Duty-compliant debt recovery voice agent must actively serve the four outcomes, not simply avoid prohibited language.
Fair treatment in the conversation structure:
The FCA's Consumer Support outcome requires that customers can access the support they need at each stage. For a debt recovery AI voice agent, this has specific structural implications:
-
Speak-to-human must be available at every point. The script cannot suppress the human-handover option until a contact target is met. The option must be genuine, accessible without multiple confirmations, and routed to a team that can actually handle complex conversations — not just a general helpline queue.
-
Affordability enquiry must precede plan offers. A Consumer Duty-compliant script offers a repayment plan only after a genuine (not perfunctory) affordability conversation. "Can you afford £50 per month?" is not an affordability assessment. The agent must ask about income, essential expenditure, and other commitments — and the script logic must route based on the answers, not ignore them.
-
Forbearance triggers must fire consistently. If the script identifies hardship signals (job loss, health emergency, Universal Credit reference), the forbearance offer must trigger automatically — every time, regardless of whether the account is flagged as a priority recovery case. Consumer Duty does not allow priority cases to be structurally excluded from fair treatment.
-
Language must be accessible. The Consumer Understanding outcome requires that customers actually understand what is being communicated. Jargon-heavy script language — "default sum," "contractual arrears," "statutory demand" — fails this standard when used without plain-language explanation. The script should state the total owed in plain figures, the options in plain language, and the consequences of each choice clearly before requesting a decision.
What the script should not do:
- Optimise call handle time at the expense of comprehension confirmation
- Apply urgency language that increases customer distress without serving a legitimate information purpose
- Offer a single repayment option without indicating alternatives exist
- Route complaints into a "call back" loop that exceeds the FCA's complaint handling timelines
For the broader FCA compliance framework that governs AI voice in financial services, see our deep dive on FCA AI governance for voice AI in 2026 — and our guide to the FCA Code of Conduct extension to AI-assisted communications from September 2026, which adds a further compliance layer to financial services voice AI deployments.
The audit trail architecture FCA supervision will ask for
When FCA supervisors examine a Consumer Duty programme, they will ask for evidence of outcomes — not evidence of process. The audit trail architecture for AI voice in debt recovery must produce evidence at the outcome level, not just the activity level.
The five-layer evidence stack:
- Layer 1 — Call recording Full audio, indexed by account, searchable by date
- Layer 2 — Call transcript Verbatim text, speaker-labelled, linked to call recording
- Layer 3 — Disposition coding Structured outcome per call: arrangement made / referred / escalated / no contact / vulnerability flag
- Layer 4 — Outcome tagging Post-call account outcome: arrangement active / broken / complaint raised / forbearance applied
- Layer 5 — Monitoring aggregate Consumer Duty dashboard: outcome rates, vulnerability detection rate, complaint linkage, escalation rate
Layer 5 — the monitoring aggregate — is what Consumer Duty governance actually requires at board and ExCo level. The FCA expects firms to monitor whether their AI deployments are producing the four outcomes across the full customer population, not just in a monitored sample. An AI voice programme that generates call recordings and transcripts but does not aggregate outcome data into a Consumer Duty monitoring view is structurally underequipped for supervisory review.
Complaint linkage is a specific requirement that many debt recovery operations overlook. When a complaint is raised — about an AI voice call specifically, or about the debt management process generally — the audit trail must be able to surface the relevant call recording, transcript, and disposition code. If the CRM complaint record cannot link back to the AI voice call record, the evidencing chain is broken and the firm cannot demonstrate that the complaint was investigated with reference to the actual interaction.
Data retention adds a further dimension. Call recordings, transcripts, and disposition codes for debt recovery interactions under Consumer Duty must be retained for the duration of the relationship plus six years. The AI voice deployment must integrate with data retention infrastructure that enforces this — not a separate manual process that will decay.
Common deployment failures that break Consumer Duty compliance
Deploying AI voice in debt recovery without Consumer Duty architecture is not neutral. It actively creates supervisory risk. These are the failure modes FCA reviewers will find.
Treating AI as exempt from Duty obligations. The Duty applies to any firm that delivers financial services to retail customers in the UK. It does not have an exemption for automated processes. A voice agent that calls a customer in arrears is a firm communication subject to Consumer Duty, regardless of whether a human or an algorithm initiates and delivers it. Any advice from a vendor or internal legal team that AI voice is outside Consumer Duty scope is wrong, and acting on it creates regulatory exposure.
No vulnerability detection layer. The most common architecture failure in existing AI voice debt recovery deployments is treating vulnerability detection as a future phase. It is not. Under Consumer Duty, vulnerability detection at every customer touchpoint is a baseline requirement — not a nice-to-have enhancement. A deployment that goes live without it is non-compliant from day one.
Scripts optimised for contact rate rather than fair treatment. AI voice deployments that inherit scripts from historical human diallers often inherit contact-rate optimisation logic: short affordability windows, urgency framing, single-option plan offers. These patterns may have been acceptable under TCF. They are not acceptable under Consumer Duty, where the Consumer Understanding and Consumer Support outcomes require that the script genuinely serves the customer's need to understand their options.
Inadequate escalation from AI to human. A Consumer Duty-compliant escalation is not a routing to any available agent. It is a warm transfer — with the call summary, the vulnerability flag, the account context, and the reason for escalation passed to the receiving agent before the caller is connected. A blind transfer destroys the fair treatment architecture that the AI voice programme was building. See our guide to voice AI warm transfer and context handoff for the technical design.
Missing outcome data. Consumer Duty monitoring requires outcome data, not just activity data. A debt recovery AI voice programme that records calls and tracks contact rate but does not capture arrangement outcomes, vulnerability flag outcomes, and complaint linkage is not generating the evidence the Duty requires.
Building the Consumer Duty-compliant deployment: a checklist
- Conversation design reviewed against all four Consumer Duty outcomes — not just TCF compliance holdover
- Three-tier vulnerability detection layer — advisory, alert, and immediate escalation, each with defined response logic
- Post-call vulnerability tagging — flag attached to account record and call transcript on every call
- Affordability routing before plan offer — genuine income/expenditure conversation, branching on the answers
- Forbearance trigger logic — fires consistently on hardship signals, not selectively by account type
- Human escalation accessible at every point — warm transfer with context, not blind routing
- Five-layer evidence stack — recording, transcript, disposition coding, outcome tagging, monitoring aggregate
- Complaint linkage — CRM complaint record links to the AI voice call record
- Retention architecture — call data retained per regulatory timeline, not isolated in AI voice platform only
- Consumer Duty monitoring view — board/ExCo-level dashboard showing outcomes across the full call population
The FCA's September 2026 AI code extension
From 1 September 2026, the FCA's Code of Conduct extends to AI-assisted communications in financial services. For debt recovery operations using AI voice, this adds a formal notification and documentation obligation on top of the Consumer Duty requirements already in force. Firms deploying AI voice in collections must have their Consumer Duty architecture in place before this deadline — not because the September extension introduces the Duty (it already applies), but because the extension creates a fresh supervisory hook for FCA reviewers to examine AI voice deployments specifically.
This means the window to remediate Consumer Duty gaps in existing AI voice deployments is narrowing.
Ready to build this? Start with Dilr.ai's AI placement diagnostic — a four-to-six-week scoping engagement that maps your current debt recovery voice operations against Consumer Duty requirements and produces a ranked remediation roadmap. Or explore our AI operating model consulting for regulated financial services. For live-platform evaluation, try Dilr Voice — enterprise voice AI with Consumer Duty-aligned vulnerability detection and audit trail architecture built in.
Build the evidence stack before the FCA asks for it.
30-min scoping call. We'll assess your current voice operations against Consumer Duty outcomes and identify where the architecture gaps are — before a supervisory review does.
Written by the Dilr.ai engineering team — practitioners who deploy enterprise AI in regulated industries including FCA-regulated financial services. Follow us on LinkedIn for deployment notes, or subscribe via the RSS feed.