By Joshua Summers, EnFi
Bank boards across the UK and Europe are mandating the same thing: increase your usage of AI. And banks are listening. According to the Cambridge Centre for Alternative Finance’s 2026 Global AI in Financial Services Report, 52% of financial firms already have agentic AI in active use. But as AI works its way deeper into lending, one worry keeps surfacing: accuracy. In a survey by Arizent, 80% of bank executives said they were very or somewhat concerned about the risk of nonsensical or inaccurate information in their generative AI. But I believe there’s a bigger concern they’re overlooking: context.
Yes, AI makes mistakes. But the biggest threat to successful AI adoption is poor data quality and siloed legacy systems. Think about spreading where you’re pulling numbers off income statements, tax returns and rent rolls into a model. Try to automate it and someone might object, saying, “if the AI gets one number wrong I’ll have to check it by hand anyway”. But your analysts get numbers wrong too. The difference is that when a human does it, we just assume we can trust it and we have no idea where the mistakes actually are. The question was never whether AI makes mistakes. It’s whether you can verify and trace the work. And that’s exactly where most banks fall down.
A real credit decision isn’t ten documents you drop into a chatbot. It can be a thousand documents, plus data sitting in third-party systems and years of judgment that lives in someone’s head and was never written down. Point a general-purpose tool at ten files and it’ll happily write you a credit memo, but it can’t reconstruct what it never saw or give you the depth that credit demands. What you actually need is a cardiac surgeon for credit: a system that knows what a borrower, a guarantor, a covenant or an amendment is, and how they all connect. What we call the ontology of credit. You wouldn’t hire a primary care physician to be your next underwriter, and a general-purpose model is exactly that hire.
This matters more in our world than almost any other, because the rules are coming. SR 26-2, the Fed’s April 2026 model-risk framework, was written for old-school credit models and barely touches the generative and agentic systems banks are rushing to deploy. In Europe, meanwhile, the AI Act now treats creditworthiness assessment as high-risk. If you’re putting AI anywhere near a credit decision, three things are non-negotiable: provenance, explainability and auditability.
Provenance: you can trace every data point back to the document, the page, and the moment in time it came from. Explainability: the system doesn’t just hand you an answer, it shows how it arrived at that decision, so when an auditor sits down next to you and asks for the details, it’s all there. And auditability: every action, human or AI, recorded and reversible, so it’s always clear who did what.
Putting Copilot, Claude, or ChatGPT on every desktop is not an AI strategy. There’s no doubt they are excellent tools, but nobody in a bank is a prompt engineer. With no structure underneath, every analyst will feed in a different context and get a different answer, and you won’t find this out until the next review cycle. And if you haven’t authorized AI tools, there’s a very real chance your employees are doing it anyway using private access. Your private data ending up in a public model without your IT knowing is a huge risk. All concerns that most banks haven’t priced in yet. Which is why the unglamorous part has to come first. In most institutions the data is a mess scattered across documents, spreadsheets and systems that were never built to talk to each other. Before AI can help you, that knowledge has to be gathered, structured and connected into a single, traceable view of the borrower and the deal.
Get that right and it powers everything on top of it: your tools, your agents, your people. And to be clear, this isn’t really about loan losses, which remain historically low for now despite a wall of commercial real estate maturities approaching. It’s about efficiency and speed, the metric banks genuinely live and die by.
The lenders who can underwrite in hours instead of weeks will win the best borrowers, and in a consolidating market they’ll end up buying the ones who can’t. The edge won’t come from digitizing a few documents or bolting on another tool. It’ll come from turning all that fragmented data into intelligence you can actually trust and defend.
Because in banking, the most dangerous AI mistake isn’t making up an answer. It’s making a decision without the full context.
