When LLMs first hit the scene, the consensus was clear: you need a semantic layer for reliable, accurate results. The benchmarks proved it. The research confirmed it. I believed it too (and was one of the first voices saying so).
Then the models got better. And the consensus stopped being true.
Today, a semantic layer doesn't give your agent accuracy. It gives your agent a ceiling. It limits flexibility, constrains the questions your agent can answer, and forces you to anticipate every question in advance. That's not how data in enterprises actually works.
This talk covers where the semantic layer falls short, what we built instead, and how to architect agents that gather and create business context dynamically, without sacrificing governance or trust.
