If every institution is feeding similar data into similar AI systems, why would any of them arrive at a different conclusion?
They would not. And that is not a technology limitation. That is the point.
The VaR Precedent
In the years before 2008, virtually every major bank used Value at Risk — VaR — as its primary measure of market risk. The model had a structural flaw that was not secret: it assumed a normal distribution that does not exist in real markets, and it offered no insight into what happened above its confidence interval. You could have a 99.9% probability of risk below 5% of capital, and the remaining 0.1% could still cost 1,000% of it.[1]
Every institution knew this. Every institution used the model anyway. Because the consensus had adopted it. Because regulators had blessed it. Because questioning it meant sitting in a room without a number to defend. The industry did not lack data in 2007. It lacked the institutional willingness to interrogate what the data was actually saying.
As one post-crisis analysis noted bluntly: we missed the build-up of risk despite having all the numbers in front of us.[2]
The Homogenisation Effect
A 2024 study published in Science Advances found that AI enhances individual productivity but systematically reduces the collective diversity of ideas. What appears original at the individual level becomes indistinguishable in aggregate.[3] The mechanism is structural: AI systems are designed to produce what is statistically probable, not what is genuinely different.
Applied uniformly across an industry feeding similar market data into similar models, the output converges. Similar risk assessments. Similar strategic recommendations. Similar blind spots. Produced faster. With greater institutional confidence. Before anyone has questioned whether the right question was asked.
The industry moves together — and calls it progress.
What This Means for Financial Services
The organisations I have seen use data well — across Japan, Hong Kong, and Southeast Asia over two decades — were not the ones with the most data. They were the ones with the sharpest questions. The most instructive example I was part of: a bancassurance operation in Japan where a manual, paper-based authorisation process was taking three months per policy. The question was not "what data do we have?" It was "what data do we actually need, and what architecture surfaces it at the point of decision?" The answer reduced processing time from three months to three minutes. The data had always been there. The architecture to make it decision-relevant had not.
That is data as how and why, not as what. AI handles the what at speed and at scale. The how and why — the question behind the question, the judgment that the model was not trained to make — still require a human capable of originating a thought the consensus has not yet had.
Critical thinking is not a soft skill. It is the practice by which an organisation avoids being right for the wrong reasons. In a regulated industry, being right for the wrong reasons is often more dangerous than being simply wrong.
When your model tells you something, does your organisation have a culture that asks why — or one that asks which slide it goes on?
- On VaR model limitations: standard VaR used Gaussian distribution assumptions and offered no coverage above its confidence interval. Financial Risk Manager Blog, '2008 Crash and COVID-19 Crisis: Impact on Financial Institutions' Risk Management'; Nassim Taleb, The Black Swan, 2007.
- Jon Danielsson, 'Models and Risk,' modelsandrisk.org. Post-crisis analysis of consensus model failure in financial services.
- Doshi, A.R. & Hauser, O.P., 'Generative AI enhances individual creativity but reduces the collective diversity of novel content,' Science Advances, Vol. 10, Issue 28, eadn5290, 2024.