I spend most of my time in rooms where someone is trying to figure out whether AI is real or just another vendor pitch.
The rooms are usually in Milan, sometimes in London or Zurich. The people across the table are CIOs, Chief Data Officers, Heads of Innovation. They run banks that manage billions. And most of them are stuck.
Not because they don't believe in AI. They do. Every single one of them has an "AI strategy." The slide deck exists. The board has seen it. The innovation team has been hired.
But here's what I actually see, week after week: the gap between the strategy deck and production is enormous. And it's getting wider, not smaller.
The pattern
It usually goes like this. A bank decides to "do AI." They hire a team or engage a consultancy. Six months later, there's a proof of concept. It works in a sandbox. Everyone is excited.
Then reality hits. Compliance needs to approve it. IT needs to integrate it. The data isn't clean. The legacy systems don't talk to each other. The vendor's API doesn't work with the bank's infrastructure. The internal team is stretched across four other projects.
Twelve months later, the POC is still a POC. The board is asking for results. The team is burned out. And the competitor down the street just went live with something that actually works — because they bought it instead of building it.
I've seen this happen at least a dozen times in the past two years.
What actually works
The banks that are deploying AI in production — and there are some, they're just quieter about it — tend to share a few traits.
First, they start small. Not "small as in a pilot," but small as in: one specific problem, one specific user, measurable outcome. An advisor who needs to prepare for a client meeting in half the time. A compliance team that needs to generate reports that used to take three days.
Second, they're honest about what they should build and what they should buy. The ones that try to build everything in-house usually end up building nothing. Banks keep buying AI they never deploy partly because of this exact mistake. The ones that buy everything end up with a Frankenstein of disconnected tools. The smart ones figure out where their real competitive advantage is and own that. The rest, they partner.
Third, they involve compliance from day one. Not as a gatekeeper, but as a design partner. The banks that treat compliance as a blocker are the ones that get stuck at the POC stage forever. The ones that treat it as a feature — "our AI is compliant by design" — are the ones that actually ship. The EU AI Act is accelerating this by giving everyone a clear framework to build against.
What I think about when I fly back to California
I split my time between Milan and Palo Alto. The contrast is something I think about a lot.
In Silicon Valley, the default assumption is that AI works and you should ship it yesterday. In European banking, the default assumption is that AI might work but you should study it for another year before doing anything.
The truth is somewhere in the middle. Silicon Valley underestimates how complex regulated industries are. European banks underestimate how fast the technology is moving and how much ground they're losing every quarter they wait.
The banks that will win this decade are the ones that figure out how to move at Silicon Valley speed with European rigor. That's a small club right now. But it's growing.
Why I write this
I'm not writing this to sell anything. I'm writing it because I have a front-row seat to something that most people only read about in reports. I talk to the people making these decisions every day. I see what works and what doesn't.
And I think the conversation about AI in banking is too abstract. Too many think pieces, not enough field reports. Too many "AI will transform banking" and not enough "here's what actually happened when a bank tried to use AI last Tuesday."
That's what I want to write about. What I actually see. Not what I think should happen — what's happening. Because innovation is not a department — it's what happens when real problems meet real solutions in the field.
