Part of the AI on Trial: The Burden of Proof campaign series
The Defence Brief: Mortgages
A broker takes a call at 9.14am. The customer is self-employed, buying his first home, with a deposit part-saved and part-gifted. He mentions, halfway through, a county court judgment from 2021 he had forgotten about. The broker has an AI assistant taking notes, drafting the fact-find, pulling lender criteria, and suggesting which products fit. By 9.47am, three product recommendations and a draft suitability letter are with the customer.
Now multiply that by every regulated mortgage adviser in the UK. Every day. Every call.
If the FCA asks how any one of those recommendations was generated, can the firm answer? Not in principle. With the file. With the AI’s contribution clearly labelled, the adviser’s edits captured, and the suitability reasoning retrievable.
Mortgage advice has always carried specific record-keeping obligations. Those obligations did not loosen when AI entered the room. They sharpened.
The FCA spent 2025 making one position very clear. Consumer Duty and the Senior Managers and Certification Regime already apply to AI. The regulator does not need new rules to enforce expectations on AI-assisted mortgage advice. It has the rules it needs.
In December 2025, the FCA published feedback statement FS25/6 and its Mortgage Rule Review roadmap. The document frames AI as a tool to help brokers give better and faster advice, with the adviser keeping responsibility for the final decision. It also flags industry concerns about AI hallucinations, errors and incomplete data influencing recommendations. The FCA’s working assumption for the years ahead is that intermediaries will increasingly act as interpreters and validators of AI outputs.
That is the regulator’s position. The adviser owns the recommendation. The firm owns the evidence.
This is what makes mortgages the sector where the AI evidence question is no longer optional.
Why mortgages face a specific AI compliance problem
Mortgage advice sits under MCOB, the FCA’s Mortgages and Home Finance Conduct of Business sourcebook. The rules require firms to assess and record suitability for every regulated mortgage recommendation. They require firms to retain records for at least three years after the advice is given. They require those records to show what the customer needed, what products were considered, and why the recommendation was made.
Consumer Duty layered another set of obligations on top. Firms have to evidence good outcomes across products, price, understanding and support. Principle 12 requires firms to act to deliver good outcomes for retail customers. Inferring quality from a sampled subset of cases is not what the FCA has signalled it wants. It wants evidence tied to specific interactions.
Three regulatory regimes. One mortgage recommendation. Every interaction.
Now add the volume. A single broker can handle forty cases a month. A medium-sized intermediary firm can clear thousands. The cases look similar on the surface. The differences sit in the detail. Self-employed income. Recent CCJs. Gifted deposits. Help to Buy contributions. Affordability stretches. Vulnerability indicators that the customer does not announce.
This is the territory where AI is being deployed hardest, and where the audit trail is most likely to break.
Charge II: the audit trail problem nobody talks about
Most mortgage firms deploying AI in 2026 have a defensible-looking governance story at the policy layer. There is a model risk policy. There is a vendor due diligence file. There is an AI use case register signed off by a senior manager.
The gap usually opens further down. At the interaction layer.
A broker uses an AI tool to draft a fact-find. The AI extracts what it heard. The broker edits it. The fact-find goes into the file. Six months later, a compliance reviewer pulls the case. They can see the final fact-find. They cannot see what the AI originally produced, what the broker changed, or why. They cannot see whether the AI missed a vulnerability indicator the broker spotted, or whether the AI surfaced something the broker overrode.
That is a chain of custody gap. And in mortgage advice it has a particular intensity, because the volume of cases is high and the cases look similar on the surface. A repeat error in how the AI interprets self-employed income, or how it treats a recent CCJ, or how it handles Help to Buy contributions, can replicate across hundreds of cases before anyone notices. By the time a sampling-based QA programme catches it, the firm has shipped a hundred suitability letters built on the same hidden mistake.
The MCOB suitability record has to show what the AI contributed. The Consumer Duty evidence has to show that the customer understood it. The SMCR liability chain has to show who reviewed it.
If the firm cannot produce that record case by case, the audit trail is broken before the regulator even arrives.
For a deeper look at the chain-of-custody requirements for AI-assisted advice, see Count II: AI Advice Without an Audit Trail.
Charge IV: advice consistency at volume is a different problem
The other distinctive challenge in mortgage advice is consistency. A single broker handling forty cases a month is naturally going to vary. Different days. Different customers. Different lender appetites. The traditional QA response has been file reviews, scorecards and feedback loops. Mortgage advice has run on this model for years.
AI changes the consistency picture in two directions at once.
In one direction, AI lifts consistency. The same prompt, the same lender data, the same regulatory framework, applied across every case. In theory, the floor rises. The least experienced broker in the firm gets pulled closer to the most experienced.
In the other direction, AI introduces a kind of inconsistency the firm has not had to manage before. The same broker, asking the AI the same question on a Tuesday and a Friday, can get materially different answers. The same case, run through the AI twice, can produce two different product shortlists. That is how large language models work, not a flaw specific to one tool.
Consumer Duty requires firms to evidence that customers in similar circumstances are getting similar quality of advice. If the AI is contributing meaningfully to the recommendation, the firm has to show the AI is contributing consistently. Not on average. Per case.
This is where the FCA’s June 2025 discussion paper DP25/2 raised concerns about AI hallucinations and incomplete outputs influencing recommendations. Intermediaries, the FCA said, will increasingly need to act as validators of AI outputs. Validation needs evidence. Evidence needs visibility into what the AI produced before the human got hold of it.
In mortgage firms, poor guidance quality scaling unchecked looks like the same systematic mis-recommendation appearing in twenty suitability letters before anyone connects them. The broker did not notice. The QA sample did not catch it. The customer signed the agreement.
This is the territory Count IV covers in depth.
For more on guidance quality and vulnerable customer obligations, see Count IV: When AI Guidance Goes Wrong at Scale.
What good evidence looks like in a mortgage file
A defensible AI-assisted mortgage advice file in 2026 needs to answer a specific set of questions when the regulator opens it.
What did the customer say. Captured from the conversation, not reconstructed by the adviser from memory after the meeting.
What did the AI generate. The first draft of the fact-find, the suggested product shortlist, the initial wording of the suitability letter, captured before the adviser made any changes.
What did the adviser change, and why. The differences between the AI’s output and the final record, with the adviser’s reasoning attached.
What did the AI miss, and what did the adviser catch. Vulnerability indicators, unusual income patterns, credit history flags, family circumstances. If the adviser surfaced these and the AI did not, that is evidence the human-in-the-loop control is working. If the AI surfaced them and the adviser overrode, that is evidence the adviser is exercising judgement. Either is defensible. The gap is only damaging if it is invisible.
What outcome did the customer get, and was that outcome consistent with similar cases. Suitability does not exist in isolation. It exists in relation to what the firm is recommending to comparable customers across its book.
This is the structured record Consumer Duty is asking firms to be able to produce. It is the chain of custody the FCA has signalled it expects when AI is in the advice process. In most mortgage firms today, it is also not currently being produced by any single system.
How Aveni helps mortgage firms build the defence brief
Aveni’s products are built for this evidence problem.
Aveni Assist supports brokers and advisers during the advice meeting itself. It transcribes the conversation, extracts the fact-find data, generates the suitability documentation, and creates a structured record that links the AI’s outputs back to the inputs that informed them. Each contribution from the AI is labelled. Each adviser edit is captured. The integrations with the platforms mortgage advisers already use, including Intelliflo Office and Xplan, mean the record lands in the systems the firm already runs compliance against. The chain of custody is built as the meeting happens, not reconstructed afterwards.
Aveni Detect monitors every interaction the firm runs, not a 3% sample. It surfaces patterns across cases that no sampling-based QA could catch. The same misinterpretation of self-employed income appearing in eight files. The same missed vulnerability cue across twelve calls. The same product over-recommendation cluster in a specific broker’s caseload. Detect closes the loop between individual case files and firm-wide advice consistency.
The next layer of Aveni’s roadmap, Guidance Agents, extends this further into real-time monitoring of guidance quality as advisers work. Aveni is designing the capability to flag advice that falls outside suitability parameters during the conversation, surface vulnerability indicators the AI missed, and intervene before a poor recommendation reaches the customer. Guidance Agents is in development as part of Aveni’s 2026 roadmap, not yet deployed.
Across the three, the proposition is straightforward. Mortgage advisers keep their judgement. Mortgage firms get a defensible evidence record. Compliance teams get visibility into the AI’s contribution at every case. The FCA gets the proof it has signalled it expects.
Aveni’s products are designed to be the expert witness for the defence. Not the prosecution.
What Heads of Compliance and Heads of Advice should check before the next FCA visit
Five questions the senior team should be able to answer in writing.
Can you retrieve the AI’s contribution to any specific mortgage recommendation, for any case in the last three years? If the answer is “we have the final file but not the AI’s drafts,” the audit trail is incomplete.
Can you show how AI-assisted advice for similar customers compares across your broker network? Consistency is now an evidence question, not a sampling exercise.
Can you identify which suitability records were AI-influenced and which were entirely human-generated? Consumer Duty and MCOB both require firms to understand the provenance of their advice.
Can you demonstrate that vulnerable customer indicators are being caught when AI is in the process? If the AI missed them and the adviser caught them, that is the evidence the FCA wants to see.
Can a senior manager named under SMCR explain the AI’s role in any specific mortgage recommendation, six months after it was made? If not, the personal accountability story breaks at the interaction layer.
If any of these answers is “we are working on it,” the AI compliance gap is already real. The Mortgage Rule Review roadmap stretches into 2027. The Consumer Duty evidence question is live now.
The firms that build the evidence infrastructure ahead of the next supervisory cycle will be the ones who get to use AI to grow their advice business with confidence.
Related reading from the AI on Trial campaign
- Count II: AI Advice Without an Audit Trail — the chain of custody requirements for AI-assisted advice under Consumer Duty.
- Count IV: When AI Guidance Goes Wrong at Scale — meeting Consumer Duty standards for AI-influenced guidance quality and vulnerable customer outcomes.
- The Defence Brief: Wealth & Advice — sister sector spoke on suitability evidence and adviser workflow.
- The Defence Brief: Insurance — sister sector spoke on vulnerable customer obligations and claims handling.
- Hub: AI Governance in Financial Services — the full campaign and the wider governance framework.
Sources
- FCA Mortgage Rule Review and Feedback Statement FS25/6, published 15 December 2025.
- FCA Discussion Paper DP25/2, June 2025, on the future of the UK mortgage market.
- FCA Handbook, MCOB 4 (Advising and Selling Standards).
- FCA Consumer Duty, Principle 12 and the four outcomes (products, price, understanding, support).
- FCA Senior Managers and Certification Regime (SMCR).
