The claim
Picture the letter. A customer complains about a decision your firm made three years ago. AI was part of how the case was assessed. The Financial Ombudsman wants to know why it went the way it did.
Most firms I speak to think they can answer that. They have the architecture diagram, the vendor contract, the model risk assessment and the governance paper the committee approved. All of it describes the system.
None of it describes the decision. The question on the table is narrower: on this file, on that day, what did the AI read, what did it flag, what did a person do with it, and who signed it off? If you cannot show that, the rest is background.
Why now
The FCA set out its expectations plainly on 19 October 2026 in its publication on AI in financial services. It says firms "should also consider whether their records, explanations and monitoring are sufficient to support challenge and help them deal with complaints". It lists the complaints sourcebook, DISP, under contestability and redress in the framework that already applies to AI.
The good practice it describes is specific. Firms "kept records of consumer prompts, outputs, model versions, applied filters and escalation decisions", which "helped monitor outcomes, complaints handling, explainability and investigation where a consumer challenged an outcome".
Then there is time. The Financial Ombudsman says a customer usually needs to complain within 6 years of the problem happening, or 3 years of becoming aware they had cause to complain. In six years, the model behind your checks will have changed. So will the prompt, the checklist and probably the vendor. A record written today is the only version of today you will ever have.
What most firms get wrong
They keep the system and lose the decision. Model documentation is versioned and filed. The output on an individual case sits in a screen that was overwritten at the next release, or in a log nobody can tie back to a case reference.
They assume the human review happened. The FCA's example from a smaller firm is the one I would pin to the wall. It recorded "whether staff needed to review AI outputs, whether this happened, and whether they approved the outputs, overrode them or escalated issues". The FCA says this showed "how oversight operated in practice, rather than relying on an assumed control". Most firms rely on the assumed control.
They treat the trail as an IT project. Engineers log what is useful for debugging: latency, errors, token counts. A complaints handler needs something else: the documents the AI saw, the rule each flag sat under, and the name next to the decision.
The strongest objection: "A person made the final call"
The best argument against all this goes like this. Our adviser, underwriter or compliance reviewer made the decision. The AI only suggested. So the complaint is about a human judgement, and we already record human judgements.
I take it seriously, because it is half right. Accountability does sit with the person and the firm. The FCA says smaller firms "are accountable for how they use AI in their business, even where third-party providers are involved".
But the objection breaks on one question an ombudsman will ask: what was in front of that person when they decided? If the AI summarised the file, flagged two issues and missed a third, the reviewer's decision was shaped by that output. Without a record of it, you cannot show whether the person checked the issue the AI missed, or simply trusted a clean result. The FCA's own wording warns that human review can be an assumed control. The record of the AI's part is what turns it into a real one.
What good looks like
For every file an AI check touches, I would want five things you can retrieve by case reference years later:
- The inputs. Which documents and data the check used, as they were at the time.
- The version. Which model and which checklist or rule set were live that day.
- The output. What it flagged, with the rule or criterion behind each flag and where in the file it found the evidence.
- The human step. Whether review was required, whether it happened, and whether the reviewer approved, overrode or escalated, with the reason for any override.
- The sign-off. Who made the final call, and when.
None of this needs new technology. It needs someone to decide, before go-live, that the per-file record is a requirement, not a nice-to-have. At Curvestone we built the check so that every finding comes with its reasoning and source evidence, and decisions, overrides and timestamps are logged for audit.
What to do next
Run a simple test this month. Pick one file from last year where an AI tool played a part. Ask your complaints team to answer the five points above using only what your systems hold today. Time how long it takes, and note what they cannot find.
If the answer is "we would need to ask the vendor", that is your gap. Close it in the contract and in the workflow now, while the cost is a design decision and not a complaint.
An earlier version of this argument appeared in GRIP (Global Relay) (October 2026). For the wider point on trust, see why the AI trust gap is an evidence problem, and for testing before go-live, how to validate an AI checking model.
Frequently asked questions
What should an AI audit trail record for a complaint?
How long should we keep AI decision records?
The AI trust gap is mislabelled. It's an evidence problem.
The AI trust gap in financial services is the distance between what an AI system can do and what a firm can prove it did to a regulator. Few doubt the model can read the file. What they cannot do is prove why it flagged what it did. Trust is not a feeling, it is an evidence trail.
How toHow to validate an AI checking model before you rely on it
Independent validation means your firm tests the AI check on files it chose, against answers your own reviewers reached without seeing the AI's result. A vendor's test report is a useful input, but the FCA says firms stay responsible for AI they procure, so the evidence that the check works on your files has to be yours.

Dawid Kotur
CEO and co-founder, Curvestone
Dawid co-founded Curvestone in 2024 after a decade working at the intersection of financial services and applied machine learning. He writes about the strategic direction of regulated-industry AI, the FCA's evolving approach to model risk, and the operational changes UK lenders are making in response to Consumer Duty. He sits on the FCA Smart Data Accelerator advisory cohort.
LinkedIn