Databricks co-founder Arsalan Tavakoli: 95% of enterprises still can't get AI agents into production — data governance is the real blocker

Summary

Arsalan Tavakoli, co-founder and SVP of Field Engineering at Databricks, argues that AI adoption is constrained by execution, not technology. While models improve daily, enterprises struggle to move AI agents from pilots to production. Data governance is the primary blocker.

Tavakoli pushes back on treating every model release as a watershed moment. Databricks sees a smoother adoption curve. The technology is getting better, but "a lot of it right now, what holds people back is less the technology, but it's more people learning how to use it, how to deploy it." The learning curve will stretch out adoption, not the algorithms themselves.

Forward-deployed engineers

When Databricks works with customers, the motion splits into two tracks. One is helping organizations get their data in order or execute migrations—foundational work. The other is building end-to-end. Fox's sports app Cleats is an example where Databricks delivered soup-to-nuts.

For AI labs trying to hire thousands of deployed engineers, Tavakoli identifies two non-negotiable traits. First, they must build hands-on using AI tools themselves. "Slideware demos" and pitch decks don't work anymore. Customers want rapid prototypes. Second, they need to talk to customers and decompose business problems into technical pieces. That blend of builder and communicator is harder to screen for than traditional solution engineering interviews.

Ramp time runs 3 to 6 months. People can go in front of customers in the first two weeks, but full operational velocity takes six months. Aggressive hiring without proper onboarding degrades culture and skill.

Data quality, not data scarcity

Tavakoli separates consumer AI from enterprise. In consumer search, you need the entire web corpus, which is roughly fixed. In enterprise, the problem inverts. AI models start at 60% accuracy, and closing that gap requires better data—cleaner, more recent, properly governed.

Data itself is not scarce. Enterprise data is scattered and often stale. "Data that you have from six months ago may no longer be valid right now." The work is not gathering more data. It's governance. Enterprises need to know what they have, ensure it's high quality, and tell the system what's current. That's where enterprises get stuck, not in model selection.

Build versus buy

Tavakoli rejects the notion that vibe coding will replace commercial software. "Build is easy. Maintaining and sustaining it is hard." His rule of thumb: buy unless the gap between off-the-shelf and your needs is too large and the workload is business-critical. At Databricks, field engineering automation has no commercial option, so they build. CRM systems, they buy if they can.