Bank of America: Fix data before chasing AI

Bank of America's stance reflects a pragmatic view that contrasts sharply with industry enthusiasm for frontier AI models and their capabilities. Rather than asking "which institution has built the smartest, most capable model," the bank reframes the question around data reliability and foundation. The core insight is that model quality becomes irrelevant if the underlying data is fragmented or inconsistent—an advanced AI system operating on poor data produces unreliable outputs that cannot be trusted in high-stakes financial decisions.

The bank traces the source of this problem to fragmented data scattered across legacy systems accumulated through years of operations and acquisitions. ERP systems, treasury platforms, bank portals, and acquired business systems speak different languages and cannot easily share information, creating a misalignment that no amount of AI sophistication can overcome. Davies's point about misinvestment captures the core risk: companies feel pressure to deploy AI quickly, leading them to build solutions on unstandardized foundations, which then fail to deliver value or worse, produce costly errors.

Bank of America's own experience with EricaAssist illustrates the alternative: an employee copilot grounded in years of client history, functioning reliably because it operates on a foundation of cleaned, standardized data. This approach aligns with Intuit's strategy for Credit Karma, which built AI assistants on a shared, consolidated view of customer financial life rather than standalone systems. The message is clear: the competitive advantage comes not from the model itself, but from the data infrastructure that allows the model to make sense of information and produce trustworthy guidance.