The Berkeley Benchmark on Data Agents, the first rigorous evaluation of AI agents on realistic enterprise data tasks, found that the best frontier models only work 39% of the time on real-world queries. The models aren't the problem. The benchmark shows agents fail at planning and reasoning, not at finding the right data, and that even throwing 50 attempts at a hard query barely moves the needle. So if it's not the model, what is it? Shared context and shared data. Not a semantic layer you build once, but a living system that captures how your business actually works; assembled continuously from the questions asked, corrections made, and decisions debated across your team. When AI has that, it compounds. Without it, you're just running expensive experiments. This talk shows what getting there looks like in practice.