There is a version of AI in financial services that is very good at finding patterns. It has been trained on years of data, it can process millions of variables, and it will tell you, with impressive confidence, what tends to happen next. What it cannot tell you is why. And in the moment you need to make a decision — to intervene, to change something, to act — “what tends to happen” is the wrong answer to the wrong question.
The moment correlation breaks
Correlation-based models are built on an implicit assumption: that the future will look enough like the past that past patterns will hold. This assumption is reasonable in stable conditions. It breaks precisely when you need it most — during structural shifts, regulatory changes, or market disruptions. More fundamentally, it breaks the moment you intervene. When you change your underwriting criteria, restructure a portfolio, or alter your ESG policy, you are not observing the world. You are changing it. A model trained on observation has nothing principled to say about what happens when you act.
This is not a failure of the model. It is a failure of the question. Correlation can tell you what co-occurs. It cannot tell you what will happen if you force something to change. That requires a different kind of reasoning — one that encodes not just patterns, but mechanisms.
What causal models do differently
A causal model does not just learn that two things tend to move together. It encodes the directional mechanism: this variable drives that one, through this pathway, under these conditions. Once you have that structure, you can ask the question that actually matters for decision-making: if we intervene here, what happens there? Not “what tends to happen when X is high?” but “what would happen if we set X to this value — deliberately, right now?”
For an actuary, this is the difference between a model that describes historical loss patterns and one that can tell you what happens to your loss ratio if you change your pricing structure. For an ESG analyst, it is the difference between a model that shows ESG scores correlating with returns and one that can tell you whether improving your sustainability practices will actually improve your financial performance — or whether both are being driven by something else entirely.
The Bottom Line
The methodology to build these models has existed for thirty years. What has changed is the infrastructure to apply it at scale — and that infrastructure is here now. The organisations that make this shift are not just getting better predictions. They are getting answers to questions that correlation-based systems cannot answer at all. In regulated industries, that is not a marginal improvement. It is a different game.
