Voice AI Complaints Handling: DISP and the Ombudsman

In the second half of 2025, UK financial services firms received 1.74 million complaints, according to the FCA's aggregate complaints data. Very few arrived on a form marked "complaint". They arrived in conversation, usually by telephone, usually inside a call opened for some entirely different reason: a balance query, a payment that had not landed, a renewal quote that looked wrong.

Voice AI sharpens that problem. A regulated firm's complaint obligations do not begin when a case is logged. They begin when the customer says something that meets the regulatory test, whether or not anyone recognises it. Put an AI agent on the front line and you have moved the point of detection into software. If that software is optimised for containment, as most voice deployments are, it is incentivised to talk a customer down rather than record that a customer was dissatisfied.

The result is a regulatory timer that starts silently. Nobody logs it, nobody acknowledges it, and eight weeks later a firm is in breach of a rule it never knew had engaged. This guide covers what counts as a complaint, which deployments the rules bind, when the clock starts, what an agent can legitimately resolve on the call, and what evidence survives if the case reaches the Financial Ombudsman Service.

This guide is shipped by the team behind Dilr Voice, enterprise voice AI built for regulated deployments. Or see DATS, our five-stage AI consulting system.

What counts as a complaint when a voice AI takes the call?

Under the FCA Handbook, a complaint is a behaviour, not a form. The glossary defines it as "any oral or written expression of dissatisfaction, whether justified or not, from, or on behalf of, a person about the provision of, or failure to provide, a financial service, claims management service or a redress determination". The word doing the work for voice AI deployments is "oral". A sentence spoken to your agent can be a complaint.

The definition is not unlimited, and this is where most internal training material goes wrong. The glossary test has two further limbs that both have to be satisfied: the expression must allege that the complainant has suffered, or may suffer, financial loss, material distress or material inconvenience; and it must relate to an activity of the respondent that comes under the jurisdiction of the Financial Ombudsman Service. So irritation alone is not a regulatory complaint. Irritation plus a claim of detriment, about a financial service, is.

That two-part qualification is the difference between a detection model that fires on every sigh and one that fires on the right utterances. "Your app is ugly" fails the test. "I have been charged twice and I could not pay my rent" passes it comfortably. "This is the third time I have called and nobody has fixed it" sits in the middle and is exactly the sentence a firm needs its agent to catch, because material inconvenience is a low bar and a repeat-contact pattern is evidence of it.

The customer also does not have to use the word "complaint", ask to complain, or be right. "Whether justified or not" is explicit in the definition. An agent that only escalates on the trigger phrase "I want to make a complaint" will miss most of them, which is why treating complaint identification as an intent-detection problem rather than a keyword problem is the highest-value design decision in a regulated voice AI deployment.

Does DISP apply to your voice AI deployment?

Before any of this binds you, check that it binds you. The Dispute Resolution: Complaints sourcebook, DISP, applies to FCA-authorised respondents: banks, insurers, lenders, brokers, payment and e-money firms, investment firms, and since 15 July 2026 deferred payment credit providers. If you run a voice agent for a retailer, a logistics operator, a utility or an unregulated software business, DISP imposes nothing on you at all, and any vendor telling you otherwise is selling fear.

Scope has a second gate on the customer side. Under DISP 2.7.3R, an eligible complainant must be a consumer, a micro-enterprise, a charity with annual income under £6.5 million, a trustee of a trust with net asset value under £5 million, a small business, a CBTL consumer, or a guarantor. A FTSE 100 treasury team ringing its corporate bank about a failed payment is dissatisfied, but it is generally not an eligible complainant, and the Ombudsman route does not open. DISP 2.7.5G adds a practical instruction: where a respondent is in doubt about the eligibility of a business, charity or trust, it should treat the complainant as if it were eligible.

For most enterprises, part of the estate is in scope and part is not. A general insurer's claims line is squarely in; its supplier-facing procurement line is not. So complaint detection should be configured per line of business rather than switched on globally, and that scoping conversation belongs before go-live, not after. Getting it wrong in either direction is expensive: over-scoping floods your complaints team with non-complaints, under-scoping leaves a regulated line undetected.

The newest population is worth naming precisely. In Policy Statement PS26/1, the FCA confirmed it is applying the DISP complaint-handling rules to deferred payment credit, better known as buy now pay later, and expanding the Financial Ombudsman's Compulsory Jurisdiction to DPC activities. The FCA also confirmed that the Ombudsman's Compulsory Jurisdiction "will only cover complaints about regulated DPC agreements entered into on or after Regulation Day and where the DPC activities have been carried on from an establishment in the UK". Regulation Day was 15 July 2026. BNPL lenders that deployed voice automation while unregulated now inherit a complaint-handling regime their agents were never designed for.

When does the eight-week complaint clock start?

The clock starts on receipt of the complaint, not on classification of it. Under DISP 1.6.1R, a respondent must send the complainant a prompt written acknowledgement and must then send a written response within eight weeks of receipt. For payment services and e-money complaints the deadline is far tighter: a final response within 15 business days, extended to 35 business days only in exceptional circumstances and only with a holding response.

This is the mechanism that makes silent detection failure so costly. If your voice agent hears a qualifying expression of dissatisfaction on 1 September and nobody classifies it as a complaint, the eight weeks still ran from 1 September. Discovering it in week ten does not reset anything. Worse, under DISP 2.8.1R the Ombudsman can consider a complaint once eight weeks have elapsed since the respondent received it, whether or not the firm ever responded. Your unlogged complaint can arrive at the Ombudsman before it arrives in your complaints system.

The volume data shows how much of the industry's complaint handling happens fast, and how little of it runs long. Of the 1,720,632 complaints closed by UK financial services firms in the second half of 2025, 44.77% were closed within three business days and only 5.56% took longer than eight weeks. Firms upheld 55.54% of complaints closed in that period and paid £235.8 million in redress, an average of £215 per upheld complaint.

Read the middle bar carefully, because it is where voice AI lives. Nearly half of all complaints are neither resolved on the spot nor allowed to run long. They are worked over days or weeks by humans, which means the agent's job is almost never to resolve the complaint. It is to detect, record and hand over cleanly enough that the humans downstream inherit an accurate start date.

Can a voice AI resolve a complaint on the call itself?

Yes, and this is the most underused rule in DISP for voice deployments. Under DISP 1.5.1R, if a complaint is resolved by close of business on the third business day following the day it is received, the complaints time limit rules and the complaints forwarding rules do not apply. The eight-week machinery never engages. For a high-volume voice channel handling billing errors and service failures, that route is the commercial prize.

The rule that makes voice the interesting channel here is DISP 1.5.2AR: a complaint is resolved where the complainant has indicated acceptance of a response from the respondent, "with neither the response nor acceptance having to be in writing". An oral acceptance, spoken on the call, counts. Your agent can apologise, explain, credit the account and secure acceptance inside a four-minute conversation, and that complaint is resolved for DISP purposes.

There is a catch, and it is a written one. DISP 1.5.4R still requires the respondent to promptly send a "summary resolution communication": a written communication telling the complainant the firm considers the matter resolved, informing them they may still refer it to the Financial Ombudsman Service, indicating whether the firm consents to waive the referral time limits, and providing the Ombudsman's website address. Resolving orally removes the eight-week clock and the acknowledgement letter, not the paperwork.

The three-day route is real, but it is not a loophole and should never be configured as a containment target. The moment an agent is measured on complaints resolved rather than complaints correctly identified, you have built an incentive to record acceptance that was not freely given. Take one design rule from this guide: detection and resolution metrics must never sit in the same objective function, a principle we apply across every AI operating model engagement.

What must the agent do the moment it detects a complaint?

The moment a voice agent classifies an utterance as a probable complaint, it needs to do four things in sequence and one thing never. It must timestamp receipt, tell the customer plainly what happens next, capture the substance in the customer's own words, and route to a human path. What it must never do is make the referral conditional on a retention attempt, a save offer, or a further round of troubleshooting.

That last point is the compliance failure most likely to be designed in deliberately. A retention flow that says "let me try to fix this first" before logging is not merely poor service. It manufactures a gap between the receipt date the regulator recognises and the one your system records. Retention offers are legitimate; gating the complaint record behind them is not. The same principle governs cancellation and renewal journeys, where the exit path must never be slower than the entry path.

Capture deserves particular care in an AI pipeline. A summarisation model that compresses "you took two payments and I could not pay my rent" into "customer reports billing issue" has destroyed the evidence of material distress, the very limb that made it a complaint. Retain the verbatim utterance alongside the summary. The same discipline governs automated disposition and wrap-up, where model-generated call codes routinely diverge from what the customer said.

Tune detection against the populations most likely to be harmed by a miss. A customer showing indicators of vulnerability who expresses dissatisfaction is a higher-consequence detection than a confident complainant who will escalate anyway, which is why vulnerable customer detection and complaint detection belong in the same model governance review, not separate workstreams.

The same diagnostic logic underpins our AI execution office, where detection thresholds are reviewed on a fixed cadence rather than left to drift after launch.

How do you evidence the trail if the case reaches the Ombudsman?

Assume the file will be read by someone who was not on the call. Under DISP 1.9.1R, a firm, payment service provider or e-money issuer must keep a record of each complaint received and the measures taken for its resolution, and retain that record for three years from the date the complaint was received, extended to five years for complaints relating to collective portfolio management for a UCITS scheme.

Three years is the floor, and for AI-handled calls it is rarely sufficient on its own. The referral windows in DISP 2.8.2R run longer: a complainant cannot refer a case more than six months after the firm sent its final response, redress determination or summary resolution communication, but the outer limits are six years after the event complained of, or three years from when the complainant became aware they had cause for complaint, whichever is later. A 2026 call can therefore surface as a 2032 case.

What the Ombudsman wants is not a transcript alone but a reconstruction: what the customer said, what the agent understood, why it acted as it did, and who decided what. That means retaining the audio, the transcript, the intent classification and its confidence, the model version in force, and the routing decision. Firms that log only the final disposition code cannot answer the question that decides the case: was the complaint recognised at the time, or reconstructed afterwards? The lawful-basis mechanics of holding that material sit in our guide to call recording consent across jurisdictions, and the disclosure duties in subject access requests over call recordings.

The financial exposure is not trivial. For complaints referred to the Financial Ombudsman on or after 1 April 2026 about acts or omissions occurring on or after 1 April 2019, the award limit is £455,000, with £205,000 applying where the act or omission occurred before April 2019. In its annual complaints data, the Ombudsman reported 214,600 new complaints in 2025/26, down almost 30% on the 305,700 received in 2024/25, and upheld 30% of the complaints it resolved across all financial products.

One further trap sits in DISP 1.8.1R. Where a firm receives a complaint outside the Ombudsman referral time limits it may reject it without considering the merits, but must explain that in a final response. A firm whose voice agent silently discarded the original contact cannot prove when the clock started, so it cannot safely rely on time-barring at all. Poor detection does not just create breaches. It forfeits a defence.

What is the best voice AI setup for regulated complaint handling in 2026?

The best setup for regulated complaint handling in 2026 is not the platform with the most natural speech. It is the one that treats complaint detection as a governed, versioned, auditable classifier rather than a prompt instruction, and that can prove which model version was live on a given call eighteen months later. On that criterion Dilr Voice is built for regulated estates, but the honest comparison names where others win.

Judge candidates on four criteria: detection quality on real dissatisfaction utterances rather than keyword triggers; immutable receipt timestamping no downstream workflow can overwrite; per line of business scoping so unregulated lines are not swept in; and evidential export that produces an Ombudsman-ready file without an engineering ticket. Most platforms handle the first two and fail the last two.

Where competitors genuinely win: if you are a developer-led team building a bespoke complaints pipeline and want maximum control over orchestration, Vapi and Retell AI give you lower-level primitives than we do, and you should take them. If your requirement is contact-centre scale on an established conversational estate without a UK-regulated evidence model, PolyAI is a stronger fit. If the deployment is genuinely unregulated and cost per minute decides it, Bland AI and Synthflow beat us on price, and ElevenLabs remains the reference for voice quality alone. Our advantage is narrow: regulated deployments where the audit trail is the product, delivered through DATS and integrated with the estate you already run, typically Twilio, Salesforce or HubSpot.

The macro context is worth stating plainly. McKinsey's State of AI research published in November 2025 found that while 88% of organisations report using AI, only 33% have taken it into production. In regulated complaint handling the gap is wider still, and the reason is rarely model quality. It is that nobody could answer the auditability question, the same failure pattern we see across FCA AI governance reviews.

Does the eight-week clock pause if the customer stops responding?

No. DISP 1.6.1R fixes the obligation to the date the respondent received the complaint, and nothing in the rule makes the eight weeks conditional on the complainant's continued engagement. A firm that cannot reach the customer must still issue its written response within eight weeks. Design your voice AI follow-up attempts around that deadline rather than treating an unresponsive customer as a paused case, and keep evidence of every attempt.

What changed for BNPL complaints on 15 July 2026?

Deferred payment credit came into FCA regulation on 15 July 2026, so DISP complaint-handling rules now apply to BNPL lenders and the Financial Ombudsman's Compulsory Jurisdiction covers regulated DPC agreements entered into on or after that date, where activities are carried on from a UK establishment. In PS26/1 the FCA also suspended complaints reporting rules for DPC complaints while firms sit in the Temporary Permissions Regime.

Is an AI-generated final response acceptable under DISP?

DISP 1.6.2R specifies what a final response must contain, not who or what must draft it. A model-drafted response is not prohibited. The response must accept the complaint, offer redress, or reject it with reasons, and must provide the Ombudsman's website address and explanatory leaflet. Given that a defective final response can undermine your six-month referral clock, human review before issue remains the defensible position for regulated firms.

Want to see this in production? Try Dilr Voice live, book an AI placement diagnostic, see our DATS methodology, or read about our approach to placing AI inside enterprise systems.

More on the regulatory perimeter around voice deployments: automated decision-making under Article 22, MiFID call recording for investment firms, escalation and human handover design, and the wider compliance cluster. Deeper conduct context sits in debt recovery under Consumer Duty and the FCA Code of Conduct extension.

Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.