AI and Finance

A few years ago, a major bank quietly pulled the plug on a credit-scoring algorithm that had been working perfectly—at least, according to every performance metric they tracked. The model was fast, consistent, and accurate. It never took lunch breaks or had bad days. But then someone asked a simple question: Why did it reject this particular applicant? No one could answer. Not the data scientists. Not the risk managers. Not even the algorithm itself. The bank realized they had built a machine that made decisions they could no longer explain to the humans affected by them. That moment of reckoning—when precision collides with opacity—is where my conversation with Professor Jeroen Rombouts of ESSEC Business School begins.

Jeroen studies something most of us take for granted: how we turn data into decisions. He's spent decades at the intersection of econometrics, machine learning, and finance, and what struck me most in our conversation wasn't his technical fluency—it was his insistence that the future of finance isn't about better algorithms. It's about better judgment. "The people who will succeed," he told me, "aren't the ones who can code the fanciest model. They're the ones who know what questions to ask before the model even runs." That reframing matters. Because right now, we're training a generation of professionals to optimize answers without ever teaching them to interrogate the questions.

Here's the evidence: data quality is treated like a technical problem—something you solve with more engineers or cleaner datasets. But Jeroen flipped that assumption. He argued that data quality is fundamentally a judgment problem. You can have pristine data and still make catastrophic decisions if you don't understand what the data actually represents, where it came from, or what it's silent about. He gave the example of financial models built during low-volatility periods that imploded the moment markets turned chaotic—not because the data was wrong, but because the builders never asked, "What isn't this data showing us?" The mechanism here is crucial: AI systems inherit the worldview of their training data. If that worldview is narrow, convenient, or historically contingent, the system will be brilliant within those boundaries and blind outside them. And in finance, the outside is where the disasters happen.

This is why I keep returning to Jeroen's phrase: "clear thinkers." It sounds almost quaint in an era obsessed with computational power, but it's quietly radical. Clear thinking means being able to step back from a correlation and ask whether it's causal, spurious, or just lucky. It means knowing when a model is a tool and when it's a crutch. It means—and this is the part that makes executives uncomfortable—being willing to override the algorithm when your human judgment tells you something's off. That kind of thinking isn't taught in most data science boot camps. It's not even emphasized in many MBA programs. We've become so dazzled by what machines can do that we've forgotten to cultivate what humans must do: synthesize context, weigh tradeoffs, and take responsibility for decisions that can't be reduced to an optimization function.

What makes this conversation urgent is that finance is increasingly a testing ground for AI at scale. Algorithmic trading. Credit decisioning. Fraud detection. Risk modeling. These aren't experimental use cases—they're mission-critical infrastructure. And yet, as Jeroen pointed out, the professionals entering this space often have deep technical skills but shallow conceptual grounding. They can build a neural network but struggle to explain why a particular feature matters or when a model should be trusted. The gap isn't about math. It's about wisdom. And wisdom, unlike code, doesn't scale automatically. It has to be taught, practiced, and earned through mistakes that we're increasingly delegating to machines that can't learn from failure the way humans do.

So what should leaders do? First, stop treating AI literacy as a purely technical competency. If you're hiring or developing talent in finance, prioritize people who can explain their models to a skeptical board, a worried client, or a regulator who doesn't care about your F1 score. Second, build feedback loops that surface when models are wrong—not just inaccurate, but wrong in ways that reveal flawed assumptions. Jeroen emphasized that the best teams he's worked with treat model failures as learning opportunities, not embarrassments to be buried in post-mortems. Third, create space for dissent. If everyone in the room trusts the algorithm because it's "data-driven," you've built a culture of intellectual outsourcing. Someone needs to be empowered to ask the dumb questions, challenge the consensus, and force the team to articulate why they believe what the data is telling them.

I left this conversation thinking about trust. Not trust in AI—that's the easy part, because machines are seductively confident—but trust between humans navigating a world where AI mediates more and more of our decisions. Jeroen's vision of the future isn't one where finance professionals become programmers. It's one where they become better translators: fluent in both the language of data and the language of judgment, able to move between spreadsheets and messy human reality without losing sight of what actually matters. That's harder than building a model. It requires humility, curiosity, and a willingness to sit with uncertainty instead of automating it away.

The bank that killed its perfect algorithm eventually built a new one—but this time, they staffed the team differently. They brought in people who had worked in branches, who had denied loans and granted them, who understood that behind every data point was a life with context the spreadsheet couldn't capture. The new model wasn't faster. It wasn't even more accurate by traditional measures. But it was explainable. And in a world where we're increasingly governed by systems we don't fully understand, explainability might be the most valuable feature of all.

The best finance leaders won't be the ones who trust AI the most—they'll be the ones who know exactly when not to.# AI and Finance

A few years ago, a major bank quietly pulled the plug on a credit-scoring algorithm that had been working perfectly—at least, according to every performance metric they tracked. The model was fast, consistent, and accurate. It never took lunch breaks or had bad days. But then someone asked a simple question: Why did it reject this particular applicant? No one could answer. Not the data scientists. Not the risk managers. Not even the algorithm itself. The bank realized they had built a machine that made decisions they could no longer explain to the humans affected by them. That moment of reckoning—when precision collides with opacity—is where my conversation with Professor Jeroen Rombouts of ESSEC Business School begins.

Jeroen studies something most of us take for granted: how we turn data into decisions. He's spent decades at the intersection of econometrics, machine learning, and finance, and what struck me most in our conversation wasn't his technical fluency—it was his insistence that the future of finance isn't about better algorithms. It's about better judgment. "The people who will succeed," he told me, "aren't the ones who can code the fanciest model. They're the ones who know what questions to ask before the model even runs." That reframing matters. Because right now, we're training a generation of professionals to optimize answers without ever teaching them to interrogate the questions.

Here's the evidence: data quality is treated like a technical problem—something you solve with more engineers or cleaner datasets. But Jeroen flipped that assumption. He argued that data quality is fundamentally a judgment problem. You can have pristine data and still make catastrophic decisions if you don't understand what the data actually represents, where it came from, or what it's silent about. He gave the example of financial models built during low-volatility periods that imploded the moment markets turned chaotic—not because the data was wrong, but because the builders never asked, "What isn't this data showing us?" The mechanism here is crucial: AI systems inherit the worldview of their training data. If that worldview is narrow, convenient, or historically contingent, the system will be brilliant within those boundaries and blind outside them. And in finance, the outside is where the disasters happen.

This is why I keep returning to Jeroen's phrase: "clear thinkers." It sounds almost quaint in an era obsessed with computational power, but it's quietly radical. Clear thinking means being able to step back from a correlation and ask whether it's causal, spurious, or just lucky. It means knowing when a model is a tool and when it's a crutch. It means—and this is the part that makes executives uncomfortable—being willing to override the algorithm when your human judgment tells you something's off. That kind of thinking isn't taught in most data science boot camps. It's not even emphasized in many MBA programs. We've become so dazzled by what machines can do that we've forgotten to cultivate what humans must do: synthesize context, weigh tradeoffs, and take responsibility for decisions that can't be reduced to an optimization function.

What makes this conversation urgent is that finance is increasingly a testing ground for AI at scale. Algorithmic trading. Credit decisioning. Fraud detection. Risk modeling. These aren't experimental use cases—they're mission-critical infrastructure. And yet, as Jeroen pointed out, the professionals entering this space often have deep technical skills but shallow conceptual grounding. They can build a neural network but struggle to explain why a particular feature matters or when a model should be trusted. The gap isn't about math. It's about wisdom. And wisdom, unlike code, doesn't scale automatically. It has to be taught, practiced, and earned through mistakes that we're increasingly delegating to machines that can't learn from failure the way humans do.

So what should leaders do? First, stop treating AI literacy as a purely technical competency. If you're hiring or developing talent in finance, prioritize people who can explain their models to a skeptical board, a worried client, or a regulator who doesn't care about your F1 score. Second, build feedback loops that surface when models are wrong—not just inaccurate, but wrong in ways that reveal flawed assumptions. Jeroen emphasized that the best teams he's worked with treat model failures as learning opportunities, not embarrassments to be buried in post-mortems. Third, create space for dissent. If everyone in the room trusts the algorithm because it's "data-driven," you've built a culture of intellectual outsourcing. Someone needs to be empowered to ask the dumb questions, challenge the consensus, and force the team to articulate why they believe what the data is telling them.

I left this conversation thinking about trust. Not trust in AI—that's the easy part, because machines are seductively confident—but trust between humans navigating a world where AI mediates more and more of our decisions. Jeroen's vision of the future isn't one where finance professionals become programmers. It's one where they become better translators: fluent in both the language of data and the language of judgment, able to move between spreadsheets and messy human reality without losing sight of what actually matters. That's harder than building a model. It requires humility, curiosity, and a willingness to sit with uncertainty instead of automating it away.

The bank that killed its perfect algorithm eventually built a new one—but this time, they staffed the team differently. They brought in people who had worked in branches, who had denied loans and granted them, who understood that behind every data point was a life with context the spreadsheet couldn't capture. The new model wasn't faster. It wasn't even more accurate by traditional measures. But it was explainable. And in a world where we're increasingly governed by systems we don't fully understand, explainability might be the most valuable feature of all.

The best finance leaders won't be the ones who trust AI the most—they'll be the ones who know exactly when not to.