Banks are training AI on decades of their own data. The problem is everything that data leaves out.
By Joey Rault, Chief Revenue Officer
A year ago, when I talked to commercial banking leaders about AI, the conversation stayed in the realm of ambition. I’d hear things like “here’s what this could do for us one day”. That conversation has since moved on.
It’s no longer a question of whether AI has a role to play. Every bank I talk to has already rolled out meaningful capabilities across operations, servicing, and internal workflows. Most of it sits in the back office. The harder question is the one they’re asking now: how do we turn this into something clients actually feel the benefit of?
The answer has less to do with models than most people expect, and a great deal to do with the underlying data. Specifically, a paradox that has always existed in commercial banking and that AI is about to make impossible to ignore.
The data paradox
The average commercial bank holds decades of transaction history, battle-tested risk models, and industry benchmarks. By any reasonable measure, the problem is no longer one of scarcity. And yet, there’s still a gap.
Banks can see exactly what happens inside their own walls with incredible precision: what cleared, what a client’s average balance was over the last two years, and so on. That’s all incredibly important. But it’s the shadow a business casts inside the bank, not the business itself.
For a long time, banks just dealt with that gap. It meant slower reporting and a lot of manual busywork, but they lived with it. AI changes the stakes considerably, because AI doesn’t paper over weak foundations. It exposes them.
Why AI brings the paradox into sharp relief
Consider what happens when a relationship manager’s AI assistant is fed a thin picture of a client. It will still produce an answer, and it will do so confidently. The response will be fast, fluent, and totally convincing. While a person working from the same incomplete file might hesitate, sensing something is missing, the model does not. It applies a partial picture to the next client, and the next.
What had been an inconspicuous gap in the data becomes the bank’s confident, automated opinion of its own customers.
This is the data paradox in action. Except the gap that used to cost banks extra paperwork and slow reporting now sits at the center of every AI investment they’re making.
We’ve all heard the trope: AI is only as good as the context behind it. It’s true for commercial banking. Fed internal records alone, it automates the bank’s existing blind spots. Fed the complete picture, it scales a banker’s best judgment across thousands of clients.
Where the missing piece actually lives
The missing context lives in the client’s accounting or ERP software, not the bank’s systems.
While a bank sees a cleared check, the ERP holds the intent behind it: cash flow patterns shifting, new supplier terms, margins moving in ways that won’t reach a bank statement for months.
I’ve come to think of this as a question of proportion. ERP data is a sliver of total data by volume, but nearly all of it by value. But getting it to work for a banker is a three-part problem.
Reaching the data. ERP data belongs to the client, not the bank, and getting to it requires their explicit consent and active participation. Plenty of banks stall here and never move past it. Those that do have cleared a real hurdle, because no other source paints as rich a picture of how a business operates. And no model can compensate for its absence. AI works with what it’s given. If the records showing a customer paying late, or a contract that quietly lapsed, never reach the bank, the model won’t spot them either.
Making it usable. Every ERP records the same concept a little differently. Turning that into something consistent and trustworthy isn’t a mapping project a team finishes once. It has to run continuously, across the whole book of business. The tempting shortcut is to hand this job to the AI itself, but it doesn’t work. A model asked to interpret inconsistent records will make a reasonable guess, and a slightly different one each time, with no way to trace how it got there. Standardization has to produce the same answer every time, and show its working. That takes deterministic, governed infrastructure underneath the model. Skip that layer, and a bank ends up back at the confident, unverifiable answer it was trying to avoid.
Getting it in front of the right person. Data that sits in its own system, however well-structured, is one more place a busy banker has to remember to go. Most won’t, because behavior doesn’t change just because a better dataset exists somewhere in the bank. The insight has to travel to where the banker already works, which for most is fast becoming the LLM itself: where they ask questions, draft client notes, and increasingly expect answers to meet them.
What it changes when the picture is complete
The thing I find myself telling banking leaders most often is that none of this has to be built from scratch anymore. Reaching consented client data, keeping it current and governed, and delivering it inside the tools bankers already use has become its own layer of infrastructure, one a bank can put in place rather than assembling piece by piece internally. That’s the kind of banking-grade foundation Codat was built to provide.
Once that foundation is there, every AI investment the bank has already made becomes more valuable, because each one is drawing on the same complete client picture.
The copilots and servicing tools already in place don’t need replacing, and neither do the portfolio analytics dashboards sitting alongside them. They need better context underneath them.
Which brings me back to the question I started with: how do we turn these AI investments into something clients actually feel? I’ve watched what it looks like when it works. A banker walks into the conversation already understanding the business. They raise a working capital issue before the client has noticed it, or flag a supplier dependency that’s quietly becoming a risk, and they do it with a specific number attached. That’s the moment AI leaves the back office and shows up in the relationship.
The best bankers I know have wanted to work this way their whole careers. The desire was never missing, the data was. And for the first time, that’s a solvable problem.
Ultimately, the banks that pull ahead from here won’t be the ones with the most advanced models. They’ll be the ones that refuse to act on half a story.