Yet, for every success story, there are far more cases where AI initiatives underperform or fail outright. A major European bank scrapped its multi-year AI-based AML (anti-money laundering) system after it failed to outperform rule-based methods. Across the industry, AI pilots often struggle to scale or deliver measurable ROI. What’s causing the gap between AI expectations and real-world performance?

The answer lies in the assumption that general-purpose AI models, those trained on broad, non-financial datasets, can easily be adapted to the specialized, regulated, and data-fragmented finance scenarios. They can’t. Generic AI is not equipped to deal with sparse datasets in emerging markets, the interpretability requirements of financial regulation, or the subtle patterns that signal synthetic fraud in payments systems.