A leading private bank asked us to build an AI capability to help relationship managers identify emerging risks across large portfolios. The system had access to market data, client information, historical transactions, risk indicators, and relevant policies. It was technically well designed. The retrieval architecture was sound. The recommendations it generated were consistently coherent and data supported. The problem appeared several months into production when two clients received very different recommendations from the same underlying risk signal. Both recommendations were technically defensible. Both were supported by available data. Both would have passed any reasonable audit of the system's outputs. Yet when the decisions were reviewed internally, the question that surfaced was not whether the system had the right information. It was whether the system understood the context in which that information mattered. One client was approaching a major liquidity event with a sho...