AI agents for complaints handling: assist and evidence, never decide

In the second half of 2025, firms regulated by the FCA closed 1,720,632 complaints, and 5.56% of them were closed more than eight weeks after receipt — roughly 95,000 cases that ran past the outer deadline the FCA Handbook sets (FCA, Aggregate complaints data: 2025 H2, published 28 April 2026). That figure is the right place to start when scoping an AI agent for complaints handling in financial services. Complaints are the most deadline-bound, highest-risk workflow a bank or insurer runs, which is precisely why the agent's job is to assist and to evidence — not to decide.

What the FCA's DISP rules actually require

Any automation built on top of complaints has to respect three separate clocks, all of them in the FCA Handbook's Dispute Resolution sourcebook. Getting these wrong is not a service-level miss; it is a rule breach.

• Close of the third business day. DISP 1.5.1R disapplies the complaints time limit rules and the complaints forwarding rules where a complaint is resolved by close of business on the third business day after receipt. The firm is not off the hook: DISP 1.5.4R still requires a written summary resolution communication stating that the complaint is considered resolved, telling the complainant they may be able to refer it to the Financial Ombudsman Service, and providing the Service's website address.
• Eight weeks. DISP 1.6.2R requires the respondent, "by the end of eight weeks after its receipt of the complaint", to send the complainant a "final response" which "encloses a copy of the Financial Ombudsman Service's standard explanatory leaflet", "provides the website address of the Financial Ombudsman Service" and "informs the complainant that if he remains dissatisfied... he may now refer his complaint to the Financial Ombudsman Service".
• Fifteen business days for payments and e-money. DISP 1.6.2AR is tighter: a PSD or EMD complaint gets a final response "by the end of 15 business days after the day on which it received the complaint", extending to 35 business days only in exceptional circumstances and only with a holding response sent inside the original 15.

Behind all three sits DISP 2.8.2R: the Ombudsman cannot normally consider a complaint referred more than six months after the date the firm sent its final response or summary resolution communication. The final response letter is therefore not just an answer — it is the document that starts the referral clock. Automating the letter without automating the audit trail behind it is how firms lose FOS cases they should win.

The volumes justify the investment; the uphold rate explains the caution

Complaints at this scale are an industrial process. In 2025 H2, firms received 1.74m complaints, down from 1.83m in 2025 H1. The percentage upheld by firms fell from 57.88% to 55.54%, and total redress came to £236m, against £284m the half before (FCA, Aggregate complaints data: 2025 H2). The same dataset shows 44.77% of complaints closed within three days and 49.67% between three days and eight weeks — so roughly half of all cases are already living inside the window where an agent's deadline tracking earns its keep. The next release, covering 2026 H1, is scheduled for 22 October 2026 (FCA, Complaints data).

Downstream, the Financial Ombudsman Service received 214,600 new complaints in 2025/26, a 30% reduction on the 305,700 received in 2024/25, and resolved over 224,000 (Financial Ombudsman Service, 21 May 2026). Hire purchase (motor) was the most complained-about product at 37,700, followed by current accounts at 32,900 — of which 18,900 concerned fraud and scams. The Service upheld 30% of complaints in the consumer's favour, down from 34%. Note also that consumers recorded as vulnerable rose from 10% (32,000) to 25% (52,900) of cases. A quarter of your referred caseload involves someone the rules say needs different handling. That is not a population to hand to an unsupervised model.

Where an AI agent for complaints handling in financial services legitimately helps

The defensible scope is everything that produces evidence for a human decision, and nothing that substitutes for one.

• Intake and triage. Recognising that an inbound email, call transcript or webform is an expression of dissatisfaction, opening the case and starting the DISP clock on the correct date. Missed intake is the root cause of most eight-week breaches.
• Categorisation and root-cause tagging. Consistent product and cause coding across channels, which is what makes the firm's own complaints MI — and its regulatory return — worth reading.
• Evidence gathering. Pulling the transaction history, call recordings, prior contacts, product terms and policy documents into one pack, with provenance attached to every item. This is the highest-value task and the one humans spend most of their time on.
• Drafting the final response. A structured draft that sets out the complaint, the evidence considered, the reasoning and the required FOS wording — presented to a case handler for amendment and sign-off, never sent automatically.
• Deadline tracking. Per-case countdowns against the three-day, eight-week and 15-business-day clocks, escalating before the breach rather than reporting it afterwards. This is the least glamorous capability and the one that most reliably reduces FOS referrals caused by silence.

All of it depends on the agent working inside your case management system, against your records, under your access controls — which is a procurement question before it is a technical one. Our security questions to ask an AI agent vendor before signing covers the ground that matters here.

Where the agent must not act alone

Four decisions stay with a named human, every time: whether to uphold or reject; how much redress to pay; whether the complainant has characteristics of vulnerability and what that changes; and the final wording of anything that affects referral rights. Add a fifth in practice — any case already close to the eight-week boundary should be routed to a person, not a queue.

The reason is not squeamishness about automation. It is that a wrong final response is not a bad customer experience; it is a rule breach with a six-month referral tail and a regulatory return attached to it.

The Consumer Duty raises the bar from "answered" to "good outcome"

Under the consumer support outcome, PRIN 2A.6.2R requires firms to design and deliver support that meets the needs of retail customers, including those with characteristics of vulnerability, and to ensure customers do not face "unreasonable barriers (including unreasonable additional costs)" when they, among other things, "submit a claim" or "make a complaint". PRIN 2A.6.3G adds that unreasonable barriers include steps that are "unreasonably onerous or time consuming" or "complex", and requesting "unnecessary information or evidence".

Read that as a design constraint on Consumer Duty complaints AI. An agent that reduces handling time by demanding three more documents from the complainant has improved your internal metric and created a regulatory problem. The correct instinct is the opposite: the agent should retrieve from your own systems what a human would otherwise ask the customer to send.

SM&CR means the accountability never moves to the model

The FCA has been explicit that it is not writing a separate rulebook for this. "We do not plan to introduce extra regulations for AI. Instead, we'll rely on existing frameworks, which mitigate many of the risks associated with AI," it says, adding that "our rules emphasise accountability for senior managers and are relevant to the safe use of AI" (FCA, AI and the FCA: our approach, updated 13 February 2026).

The harder question is what discharging that accountability looks like. Sheldon Mills put it directly in a speech on 28 January 2026: "Accountability under the Senior Managers and Certification Regime (SM&CR) still matters – but what does 'reasonable steps' look like when the model you rely on updates weekly, incorporates components you don't directly control, or behaves differently as soon as new data arrives?" (FCA, Designing for the unknown). In complaints, the practical answer is a logged decision trail: which evidence the agent retrieved, what it drafted, what the case handler changed, and who approved the response that went out. For the wider regulatory picture, see our guide to which UK regulator governs your AI agent.

The same pattern applies under other regulators' complaint regimes

The clock changes; the architecture does not. In the United States, companies receiving a complaint through the Consumer Financial Protection Bureau "generally respond in 15 days", with a final response in 60 days where the matter is still in progress (CFPB, Complaint process). A firm operating across both regimes is running the same workflow against different deadlines and different referral wording — which is an argument for making the deadline logic configurable rather than hard-coding a jurisdiction's rules into the agent.

Working with Braivex

Braivex builds Service-as-Software for financial services firms internationally: secure, context-aware agents deployed into the systems you already run, priced on outcomes rather than seats. For complaints automation in a bank or insurer, that means an agent that opens the case, gathers the evidence, tracks every DISP deadline and drafts the response — while your case handlers keep the decision and your senior managers keep the audit trail. The same pattern maps onto other regulators' complaints regimes; only the clocks and the referral wording change. If you want to see how it would sit against your own complaints volumes, get in touch.