There's a common assumption in the tech world that banks are the last ones to move. Too regulated, too slow, too conservative to take real bets on new technology. I held some version of that assumption myself, honestly. Then I spent two months training around 100 employees inside one of Portugal's biggest banks on AI agents, and I left with a completely different picture.
I can't share which bank it is, but I can share everything I learned.
The Platform Bet
Let's start with the technical setup, because it sets the tone for everything else.
The bank didn't use any third party agent platform. Not Microsoft Copilot, not any of the managed tools from OpenAI or Anthropic, none of it. They built their own platform from scratch. And when I first saw it, my honest reaction was: this is impressive, and also a massive bet.
I completely understand the reasoning. This is a bank. Data control is not a nice to have, it is existential. They needed to know exactly where data goes, what models see it, and that nothing ends up in any training pipeline they don't control. So they plugged into something like AWS Bedrock to access frontier models in a secure, isolated environment, where they own the data and the models never train on it. Smart. Necessary, even.
But building your own agent platform is a completely different thing from using a hosted model endpoint. These are full blown automation agents with configurable logic, branching, integrations. Complex applications. And complex applications have bugs, require maintenance, and can fall behind fast when the third party tools you chose not to use are shipping major updates every few weeks.
The agent platforms that exist today, the ones we track in this newsletter, are getting better at an extraordinary pace. New features, better reliability, more integrations, better developer experience. Every week. When you build your own, you're betting that your team can keep up with that pace while also running a bank. That's a hard bet to win.
If I owned a bank today, I am genuinely not sure what I would do. The control argument is real. The maintenance burden argument is also real. I lean toward not building my own platform, but I understand why they made this call, and I have a lot of respect for the ambition behind it.
The Connectivity Problem
The other big technical challenge they face has nothing to do with the build vs buy decision. It's about connecting the platform to the bank's internal systems, and this is where regulated industries reveal a complexity that most of us never think about.
When we work with smaller companies or less regulated industries, connecting an agent to internal data is usually a technical challenge. For a bank, it's a security, legal, compliance, and technical challenge all at once. There are systems that different teams simply cannot access. There are data sources that require multi layer authorisation. There are privacy rules that constrain what an agent can even retrieve, let alone act on.
Right now, the platform is useful, but not as useful as it will eventually be once those integrations are in place. The agents work with what they have access to. The potential of the full vision is still ahead of them. Getting there will take time, and probably a lot of careful work that has nothing to do with AI.
What 100 People in a Bank Actually Did With AI Agents
Now, the part that genuinely moved me.
The 100 people I trained came from all kinds of departments. Insurance teams that deal directly with clients. Back office people. Analysts. People who spend their entire day on a computer but have never thought about automation in their life. The investment the bank made to pull these people out of their daily work and put them in a training room is not trivial. Anyone who has worked inside a bank knows how every hour of every person is accounted for. This was a real commitment.
And what happened in those rooms was something I didn't fully expect.
People were genuinely excited. Not politely engaged, actually curious and motivated. They had ideas. Real ideas about problems they deal with every day that an agent could solve. And in the sessions where we actually built things together, divided into small teams, some of those ideas turned into agents that worked.
I remember one session near the end where each team shared what they'd built. One team presented an agent that classified customer complaints by bank area, and for each complaint it pulled the relevant internal procedures and suggested a possible response based on real bank documents, not generic internet content, actual internal documentation. When they finished presenting, three or four people from other teams immediately said "I want that agent. Can you share it?"
It wasn't production ready. It needed more work. But the reaction said everything about what this technology can do for people who spend their days doing repetitive, high stakes, document heavy work. The potential was obvious to everyone in the room, even the people who had never built anything in their life before that day.
What Training Actually Revealed About How People Use Agents
This is the part I keep thinking about.
When people start building agents, they tend to start well. They think carefully about the inputs, the outputs, the instructions. They're specific. They test. The agent works as expected. Then, because agents are remarkably good at following instructions, people start to relax. They get confident. They start taking shortcuts.
The instructions get looser. The logic gets more compressed. "The agent will figure it out." And then, surprisingly, it doesn't. Or it does something slightly off. And the person doesn't understand why, because the agent was doing so well before.
This is a deeply human pattern, and it's one of the hardest things to train out of people. Consistency matters enormously when working with agents. The way you described something yesterday needs to match the way you describe it today. The shortcuts that feel natural to a human, like skipping context because "it's obvious," create ambiguity that compounds over time.
The second thing I noticed is about problem decomposition. The people who got results fastest were the ones who could break a complex problem into smaller, sequential problems and solve one at a time. This is a skill that some people have developed through experience, whether from engineering, project management, or just a particular way of thinking. Others found it genuinely difficult. It's something you can teach, but it takes time.
And then there's a third thing, which is harder to explain but just as important. People need a basic, non mathematical understanding of how a language model actually works. Not the math. Not the architecture. Just enough to understand why a model does what it does, why it sometimes gets things wrong, why the phrasing of an instruction matters, why giving it more context usually helps. Without this mental model, people treat the agent like a magic box, and when it fails they have no idea where to even start fixing it.
These three things, consistency, problem decomposition, and a basic intuition about how models work, are what separate the people who build useful agents from the ones who give up after the first frustrating session.
Banks Are Not Laggards
I want to end on the thing that surprised me most, because I think it matters for how we think about enterprise AI adoption.
The assumption that large, regulated institutions are slow to move on AI is wrong. Or at least it's more complicated than that. What I saw inside this bank was a serious, well funded, strategically important initiative. They built their own platform. They trained thousands of people. They made the investment to do this properly, even knowing it would take time to get the integrations right, even knowing the platform would need continuous work.
The people in those training rooms were not resistant. They were ready. They just needed someone to show them that the thing they were imagining, agents that handle the repetitive, complex, document heavy work of banking, was actually buildable today, with the skills they already had.
The industry is moving. Faster than most people think. And the institutions that crack the training and culture side of this, not just the technology side, will have a real advantage.
That's the part that keeps me thinking long after the training ended.