AI pilots in banking rarely fail. They perform well, hit their metrics and generate a positive internal write-up. Then they just sit there, unable to scale beyond the narrow slice of clean data they were built on.
The gap between a successful pilot and something the business actually uses is where most banks are stuck, and the reason has almost nothing to do with AI. It does, however, have everything to do with what lies beneath: decades of legacy infrastructure, and data spread across it with no clear line of ownership.
Digital transformation didn’t eradicate that problem. It simply layered new platforms on top of it.
Why pilots don’t travel
The obstacle isn’t the model. It’s the data feeding the model, and, more specifically, that nobody can say for certain who owns it. In most large banks, that data still exists across a patchwork of old systems, platforms acquired along the way and business unit databases. Each has its own rules, formats and owners. Some of it is duplicated three or four times over, with each copy telling a slightly different story.
An AI model built on that base fails quietly, producing answers that look plausible on the surface but can’t be trusted. Or it needs so much manual cleanup that the efficiency gains vanish.
You can’t buy your way out of this
It’d be easy to frame this as an investment problem. Most large banks have already spent significantly on infrastructure modernisation. Ownership is a different question, though, and it actually matters more.
Data ownership has historically sat wherever it happened to land, split across IT, individual business lines and whichever team built the original system. Nobody set out to design things this way. It’s just the cumulative result of years of mergers, migrations and quick fixes made under deadline pressure.
That fragmented ownership was manageable when the main risk was a reporting error or a slow query. It’s far less manageable when the same data is expected to feed a model making decisions in real time, at scale, across an entire business.
Starting without the baggage
The contrast with these banks is telling. Not because every established institution can replicate their approach overnight, but because it shows what the absence of legacy really buys you.
From day one, these challengers treat data ownership as a design decision. Teams know what data they own, what it means and who’s responsible for ensuring its quality. When something looks wrong, there’s no scramble to work out who can change it or who needs to sign it off.
That clarity is a large part of why they can move faster. In other words, it has less to do with better technology and more to do with fewer unresolved questions about what falls under whose ownership.
Established banks can’t start from a blank slate. But they can borrow the underlying idea. This comes down to assigning clear ownership of data domains, pushing accountability down to the teams closest to the data and building governance into new projects from the outset rather than retroactively.
Fixing the foundation before scaling further
None of this means banks should pause their AI programmes to overhaul their entire data estate first. No. That’s neither realistic nor necessary. But it does mean treating data governance as a prerequisite for scale, not a cleanup job for later.
The banks making genuine progress with AI right now aren’t necessarily the ones with the biggest budgets or the most advanced models. They’re the ones that did the unglamorous groundwork first.
Nobody puts out a press release about it. But it’s the difference between an AI pilot that never leaves that stage and one that has the power to change how the bank operates.
The technology to transform banking is already here. The question every institution now needs to answer honestly is whether its foundations can support it.
Gary Ellison, VP Head of Data & AI, Valtech
