At Sierra Ventures' 21st Annual CXO Summit, Ion Stoica, Co-Founder of Databricks and Anyscale and Professor at UC Berkeley, and Patrick Wendell, Co-Founder and VP of Engineering at Databricks, made a clear case: model intelligence is no longer the hard part. Security, context and cost discipline are what turn AI demos into business value.
Wendell, who leads AI adoption inside Databricks, sees safety and governance as the biggest bottleneck. What counts as acceptable AI behavior differs by company, so guardrails and sandboxing must reflect each organization's own rules. Stoica added that surprising AI behavior usually traces back to gaps in the context it was given, and that many of the fixes are security practices enterprises have used for decades.
For most enterprise work, today's models reason well enough. What they lack is the implicit knowledge employees absorb every day: who does what, how the org works, what lives in company data. Wendell framed that collective knowledge as a company's true IP, and the central challenge as putting it to work with AI while keeping it private.
Stoica expects enterprise AI to be built from many smaller, task-specific models rather than one giant one, since a fine-tuned model can often beat a frontier model on a defined task at lower cost. Standardized APIs make swapping models easy. Wendell urged leaders to judge AI by return on investment, not token spend, starting with tracking where costs go and what they produce.