Complaints handling

• Every call transcript, chat, email and letter is checked for dissatisfaction, so complaints are caught and logged on the day they're made, with the DISP clock started correctly.
• Each complaint is classified by product, cause and the FCA's reporting categories, and routed to the right team, with possible vulnerability flagged to a specialist.
• Before the handler opens the case, the evidence is gathered: account history, contact history, the terms and policies in force at the time, and similar past decisions, including relevant published ombudsman decisions.
• The handler gets a drafted final response in plain English, with every statement linked to the evidence behind it. They check it, change it and sign it. Nothing is sent without them.
• Root causes are found across thousands of complaints, not one at a time, and each one is given to a named owner in the business to fix.
• Consumer Duty monitoring gets clean, consistent complaints data by product and customer group, including vulnerable customers.

• FCADISP 1: complaint handling rules
  →

  Complaints must be investigated competently, diligently and impartially, and answered fairly. A final response, or a written explanation of the delay, is due within eight weeks (DISP 1.6), or within 15 business days for payment services and e-money complaints, extendable to 35 in exceptional cases. Complaints resolved by the end of the third business day can be closed with a summary resolution communication (DISP 1.5). The workflow, not the AI, runs these clocks.

• FCADISP 1.3: root cause analysis
  →

  Firms must look for recurring or systemic problems in their complaints and fix them, including for customers who didn't complain. The root-cause analysis in this design exists to meet that rule, and every finding goes to a named owner.

• FCAConsumer Duty (PRIN 2A)
  →

  There's no AI-specific rulebook. The FCA applies the Duty to AI like any other change. Final responses must be clear and fair, complaints data feeds outcomes monitoring by product and customer group, and the board report shows what was found and fixed.

• Financial Ombudsman ServiceDISP 2 and 3, and the 2026 redress reforms
  →

  Every final response tells the customer they can go to the ombudsman within six months. Since 1 October 2026 the ombudsman has new grounds to dismiss complaints, and its fair and reasonable test now looks at the standards in force at the time of the act or omission. So the evidence pack must show the terms and policies that applied then, not today's.

• FCAVulnerable customers (FG21/1)
  →

  Signs of vulnerability in a complaint, such as illness, bereavement or financial difficulty, are flagged to a trained specialist, never acted on automatically, and never used to treat the customer less favourably.

• ICOUK GDPR, the Data Protection Act 2018 and automated decisions
  →

  Complaints often contain health and financial hardship details. That means a DPIA before the pilot, processing in UK regions, a clear retention period, and a contract that stops providers training their models on your data. The AI never decides a complaint, so there's no solely automated decision, and every adverse outcome is a handler's decision.

On a phone, the table shows the platform you pick. Sovereign / on-prem means open, portable components you can run in your own data centre or with a UK or European provider; IONOS is shown as one example, and others exist. What sovereign cloud means →

• A complaints quality team that samples calls and chats every week to check what was and wasn't logged as a complaint.
• Override rules, so handlers record why they changed a classification or a draft, and those reasons are fed back.
• A named owner for the policy library and the past-decisions index, so retired policies are dated, not deleted.
• A monthly root-cause forum where product owners accept or reject each theme, with fixes tracked to closure.
• A capacity plan for complaint spikes, with the classification for known redress issues set up in advance.
• Handlers trained before go-live, including what the AI can't see and why the signature is theirs.

1. 3 weeks

   Discovery

   Map today's process across every channel, measure the baseline (including unlogged complaints in a call sample), check data access and pick the first product.

2. Gate · must be true to passBaseline agreed, data access confirmed, DPIA started, compliance and the Consumer Duty champion involved.

3. 6–8 weeks

   Pilot

   One product, such as credit cards, in shadow mode: the AI classifies and drafts, handlers work as normal, and the two are compared.

4. Gate · must be true to passClassification accuracy and draft quality within agreed limits, no customer group worse off, compliance sign-off.

5. 4–6 weeks

   Production

   Drafts and classification go live for the first product, with monitoring, support and a rollback plan.

6. Gate · must be true to passMonitoring, support model, rollback plan and handler training in place.

7. Ongoing

   Scale

   Add products one at a time, then switch on root-cause analysis across the whole book and feed it into Consumer Duty reporting.

8. Gate · must be true to passEach new product passes the same pilot gate.

R ResponsibleA AccountableC ConsultedI Informed· not involved

Where I'd fit

1. 1

   Goes wrong

   The pilot only used logged complaints, so it never tested the hardest part: spotting the ones nobody logged.

   How the design avoids it

   Include a sample of ordinary calls and chats in the test set, labelled by your quality team.

2. 2

   Goes wrong

   The assistant quotes today's terms for a complaint about something that happened three years ago.

   How the design avoids it

   Index policies and terms by effective date, and make the retrieval filter on the date of the events.

3. 3

   Goes wrong

   Drafts read well, so handlers start signing them without checking the evidence.

   How the design avoids it

   Show the source behind every statement, track edit rates, and sample signed responses every week.

4. 4

   Goes wrong

   Root-cause themes go into a slide deck and nothing changes.

   How the design avoids it

   Every theme gets a named owner in the business and a measure that shows whether the fix worked.

5. 5

   Goes wrong

   A redress spike arrives and the classification has no category for it.

   How the design avoids it

   Set up streams for known issues before they peak, and keep the taxonomy owned and versioned.

1. Early 2027
   Motor finance: the Upper Tribunal hears challenges to the FCA's redress scheme. Depending on the outcome, commission complaints are handled through the scheme or individually within the normal eight weeks.

   Keep commission complaints as their own stream, and plan capacity for both outcomes.

2. From 1 January 2027
   The FCA's new single complaints return, COM001, replaces five existing returns, with fixed half-year reporting periods starting January to June 2027 (PS25/19).

   Map the classification to the new return's fields before go-live, so reporting comes straight from the case data.

3. Q1 2027
   The FCA's policy statement on the scope and proportionality of the Consumer Duty (consulted on in CP26/23).

   Keep complaints MI split by product and customer group, so it fits the board report whatever its final shape.

4. 2026–2027
   Further ombudsman reform: a registration stage piloted from October 2026, starting with fraud and scams, a consultation on case fees, and a 10-year time limit for complaints in the Financial Services and Markets Bill now before Parliament.

   Keep evidence packs complete and dated, so a case can move to the ombudsman quickly and stand up on the standards in force at the time.

5. 2026–2027
   The ICO's code of practice on AI and automated decision-making, required by regulations in force since May 2026.

   Keep the classification log, the drafts, the edits and the signatures in a form you could show the ICO.