Why Enterprise AI Keeps Failing—and It's Not the Company's Fault
Arsalan Tavakoli-Shiraji, a co-founder of the data company Databricks, will speak at TechCrunch Disrupt 2026 in San Francisco on October 13-15. His talk has a striking title: "The Enterprise Isn't Broken. Your Assumptions About It Are."
The title signals something important. It's pushing back against a common story in the technology world: that big companies are slow, resistant, or outdated when it comes to AI. Tavakoli-Shiraji is suggesting the real problem is different. The people trying to bring AI into these companies often misunderstand how those companies actually work.
Where Enterprise AI Projects Go Wrong
Right now, many large organizations are struggling with AI. They start with a promising pilot project—a small test that looks good in the lab. But when they try to scale it up across the whole company, things fall apart. The AI system that worked fine in a controlled setting suddenly bumps up against real-world complications: messy data, strict rules about who can see what information, security requirements, and ways of working that have been in place for years.
A few things typically go wrong. The AI system doesn't match what the business actually needs. The data isn't organized the way the AI expects. And the company's rules about data privacy and security create obstacles that weren't anticipated. These aren't signs of a broken organization. They're signs that the people building the AI didn't fully understand the constraints a large company has to work within.
A Pattern We've Seen Before
This reminds me of something that happened twenty years ago, when cloud computing was new. Early cloud companies blamed big enterprises for being stuck in their ways and afraid of change. But that wasn't really what was going on. The real issue was that cloud services were designed for a different set of problems. Large companies needed security and compliance rules that cloud providers hadn't built in yet. Once vendors understood what enterprises actually needed—rather than trying to force them to change—cloud adoption took off.
Something similar is happening with AI now. Databricks, the company Tavakoli-Shiraji co-founded, was built on this insight. Instead of telling enterprises to rip out their existing systems and start fresh, Databricks designed its technology to work alongside the systems that are already there.
What Tavakoli-Shiraji Brings to This Conversation
His job at Databricks puts him in a position to see what works and what doesn't across many different companies. When an enterprise tries to deploy AI and it either succeeds or stalls, he's in the room. That ground-level view matters.
The company's approach has been to build AI infrastructure that respects how enterprises operate—with their governance rules, security requirements, and existing data systems—rather than asking them to reshape themselves around the technology. His presentation at Disrupt will likely focus on how AI people need to rethink their assumptions about what enterprises need.
Why This Matters Now
Big companies have real constraints. They can't move as fast as a startup because they have to manage risk, follow regulations, and keep their operations consistent. From a pure technology standpoint, these constraints can look like obstacles. But they exist for a reason: they keep the company safe, compliant, and stable.
The broader context here is that enterprise AI has been treated like a technology problem when it's partly an organizational one. When someone talks about how well an AI model performs in tests, that's useful information. But whether that AI actually gets deployed and creates value depends on how well it fits into the way the company works.
Many enterprise AI projects fail not because the technology is weak, but because nobody checked whether the setup would work in reality. That's where Tavakoli-Shiraji's title comes in: the enterprise isn't broken. The assumptions were.
What This Means for Companies Trying AI
For any large organization looking at AI right now, this is a useful lens to look through. Instead of asking "Why is our company too slow to adopt this?", ask "Are we building this in a way that works with how our company actually operates?" The answer might be no—and that might be why the project stalled.
The conference is happening in October, when enterprises are planning their budgets for next year. It's a strategic moment for these conversations to happen. And having someone explain why the problem isn't enterprise resistance, but misaligned expectations, might change how companies approach their AI plans.
