
Open Banking Limited has published “Do LLMs Have Standards?”, the findings from its AI consultation. One line in it deserves more attention than the rest:
Data quality is the biggest driver of AI value, and of AI risk.
I agree, though the paper’s focus is deliberately on the infrastructure: the standards, APIs and plumbing banks use to expose data. For a lender, quality is decided somewhere else. It is decided in the few minutes an applicant spends connecting their bank account, and in what happens to their transactions straight afterwards.
That is where we have spent our time at Credit Canary: using AI to personalise the application journey so the data we collect is complete, then to categorise and enrich it, with people checking the result. The report makes four findings that matter to lenders. For each one, here is the problem behind it and how we solve it.
1. Data quality decides whether AI helps or harms
What the report found. OBL concludes that data quality is the biggest driver of AI value and of AI risk. One respondent told it that open banking transaction data “varies widely in practice”.
The problem. For a lender, the data is often incomplete before it is ever messy. The applicant connects one account, their salary lands in another, and the lender sees half a picture. A model cannot interpret data that was never collected.
How Credit Canary solves it. We use AI inside the journey itself, shaping it around the individual applicant and unifying multiple bank accounts into a single view. Our journey detects income while the applicant is still there, shows it back to them, and prompts for missing accounts before they leave. The applicant guides the lender through the complexities of their accounts, while the figures still come from the transaction data. That gives us context that later analysis cannot recover, without relying on a self-declared answer that could be manipulated.
That journey won the Open Banking Award at the FF Awards 2025, and our consent completion has averaged 83% across 11 consecutive months against an industry median of 62%.
2. AI works best as the last-mile enrichment layer
What the report found. Respondents see AI as a last-mile layer that standardises and enriches transaction data, sitting on top of well-maintained APIs. TPPs added a warning: accuracy matters, and model updates can degrade reliability. Respondents also cautioned against leaning on AI for basic tasks such as simple data standardisation.
The problem. Rules catch the descriptions you have seen before. The long tail is harder: the small employer, the local landlord, the merchant whose reference changes every month. Every transaction left unmatched is a gap in the affordability picture.
How Credit Canary solves it. We use AI to find similarity and likeness between a new transaction and ones we already understand, which lifts match rates where rules run out. Then, when people check the work, human-in-the-loop feedback corrects what the model gets wrong, and those corrections feed back in, so accuracy improves over time.
The discipline is to use AI where it adds something and nowhere else. Sending every transaction to a model is the lazy option: it costs more and it is harder to check. Rules and known matches handle the routine work, and AI is reserved for what they cannot resolve. That keeps cost under control and keeps the AI’s output small enough for people to review.
3. Affordability needs stronger assurance
What the report found. OBL singles out affordability analytics as a higher-risk use that needs stronger assurance than something like documentation search. It describes what that looks like: reconciliation to source, lineage, confidence scoring and an audit trail.
The problem. An affordability decision affects whether someone gets credit and whether they can repay it. The lender remains accountable for that outcome whatever tool produced the number. An AI output that nobody has checked is a liability.
How Credit Canary solves it. Every affordability figure we produce is built from the applicant’s own transactions, across all their connected accounts, so a lender can trace a number back to the payments behind it. We align that picture with the applicant’s credit file, so what the bureau reports and what the bank accounts show are checked against each other. The applicant has already guided us through anything unusual. AI categorises what rules cannot, an underwriter can review and correct the result, and each correction improves the next decision. The lender ends up with a number it can explain and defend.
4. Agents should do the legwork before a person reviews
What the report found. The paper’s longer-term view is agentic: AI that acts on a customer’s behalf within delegated, revocable consent. The FCA’s Mills Review points the same way.
The problem. Agents are mostly discussed as something customers will use. The nearer opportunity sits inside the lender. Cases reach a reviewer incomplete, and skilled people spend their time chasing data when they should be making decisions.
How Credit Canary solves it. Our agents do that work behind the scenes. They collect all the data a case requires, autonomously, before a person reviews it. Cases arrive ready, manual review times stay low, and the reviewer’s time goes on the decision.
Do not wait for the standards
OBL is candid that its 30% efficiency figure is a planning assumption still to be tested. Lenders do not need to wait for that test, or for the standards work to finish.
A lender can act on all four findings today:
- get complete data during the application
- use AI where it earns its cost
- make every figure traceable
- and have agents prepare each case before a person decides.
That is what we have built at Credit Canary. If you want to see what it does to your completion rates and review times, talk to us.
AI-ready standards will help. The harder question for a lender is whether its own data is ready.
← Back to News & Opinion