Something in AI shifted over the past year, quietly enough that the moment was easy to miss. The tools stopped just answering questions and started taking actions. In financial services that shift is further along than most people realize, and a new survey of security teams across the industry put numbers to what I had felt from the inside. Firms are deploying AI that acts faster than they are learning to govern it. That gap is the whole problem, and the agents are the least of it.

Today I am launching XipHub. It exists to solve a problem I came to know well from the inside. Almost every regulated business is trying to use AI right now, and almost none of them are getting real, dependable value from it. I had become convinced the reason is structural, and that closing the gap would take a real platform, not another tool. XipHub is that platform, and I am here to lead it. This is what it is.

At the end of March, I stepped down as the bank's Chief Data Officer to start a company. I have been turning over how to write this, because the easy version sounds like a story about leaving, and that is not how it feels from the inside. It feels like the opposite. Four years inside a bank, first running technology and then its data, taught me the industry. The problem that actually pulled me out is younger than that. It took shape over the last couple of years, as AI went from a curiosity almost no one took seriously to something every institution suddenly had to make real. I could see it clearly and could only ever work around it, and at some point working around it stopped being enough.

I have been stuck on a problem lately. Almost everyone has AI now, and almost none of that activity turns into something a business can lean on, because the value leaks out through everything the tools cannot see. What pulls at me is the next question. If scattered, ungrounded AI is the problem, what would it actually take to run AI at scale inside a regulated business? Not which model to pick. The structure around the models. The deeper I go, the more it comes down to a handful of things, and not one of them is another tool to buy.

Spend any time around how companies actually use AI right now and one thing jumps out. It is everywhere. Someone is drafting copy in one tool, someone else is summarizing a contract in another, a third person is pasting notes into a chatbot between meetings. By every obvious measure, adoption already happened. So it took me a while to work out why all that activity adds up to so little that a business can lean on. The longer I sat with it, the clearer it got that the problem is not the AI. It is everything around it.

I came into banking from the outside, a builder used to moving fast and questioning why things worked the way they did. I carried a stack of assumptions with me. Most of them turned out to be wrong, or at least incomplete. Looking back at where I started and where I am now, the gap between the two is the truest measure I have of what this stretch taught me.

There is a vision that floats around about the fully autonomous bank, AI handling everything end to end, people barely in the loop. Having spent real time building with these tools and working through what it takes to govern them, I do not believe in that destination, and not only for the reasons people expect. The more useful question is not how autonomous a bank could become. It is where autonomy should stop, and why part of that line is permanent rather than just a limit of today's technology.

Watching the open banking rule play out has been a lesson in what regulatory uncertainty does to a small bank. The rule was finalized, challenged in court almost immediately, and eventually sent back toward the drawing board. From inside a community bank, I followed it less as a policy drama and more as a planning problem, because that is what it became.

The unsettling part of this moment in fraud is not that banks finally have powerful AI to defend themselves. It is that the people attacking them have the exact same tools. I have watched the threat change shape over the past couple of years, and what stands out most is the asymmetry. AI helps the defender, but it helps the attacker at least as much, and in some ways more.

I have built enough with AI agents on my own projects to tell the demo magic apart from what actually holds up. The hype says agents will soon run everything. What I have found is narrower and more useful. They are very good at a specific shape of problem and unreliable outside it. So the interesting question for banking is not whether to use agents, but where they genuinely earn their place, and how to box them in so they stay useful.