Databricks Co-Founder Challenges How Tech Companies Think About Enterprise AI

Arsalan Tavakoli-Shiraji, co-founder and senior vice president of field engineering at Databricks, will speak at TechCrunch Disrupt 2026 (October 13-15 in San Francisco) on a topic that cuts against much of the current AI industry chatter: "The Enterprise Isn't Broken. Your Assumptions About It Are."

The talk signals a direct challenge to a popular narrative. Many in the AI world blame enterprises for being slow or resistant to AI adoption. Tavakoli-Shiraji's framing suggests the problem is different—that AI teams and vendors often bring misguided expectations about how big organizations actually work.

Why Enterprise AI Projects Often Fail

Right now, many companies struggle to move AI from early testing to real-world use. Proof-of-concept projects show promise, but when it comes time to deploy at scale, things break down. The problems typically fall into three buckets: AI capabilities that don't match real business workflows, data infrastructure that wasn't ready, and rules and oversight systems that can't accommodate AI systems.

Tavakoli-Shiraji's position gives him firsthand insight. As head of field engineering, he works with customers during implementation—the point where things either come together or unravel. His vantage point spans many different industries and company sizes.

The common industry take blames enterprises themselves—they're legacy-minded, bureaucratic, change-resistant. Tavakoli-Shiraji's argument flips this. He's saying the issue is that AI practitioners and vendors often misunderstand what enterprise operations actually need, then treat those needs as obstacles instead of requirements to build around.

A Pattern We've Seen Before

This echoes something that happened with cloud computing in the mid-2000s. Cloud evangelists back then often blamed enterprises for clinging to old ways, when the real issue was that cloud's security model, compliance setup, and integration approach didn't fit how enterprises actually ran their systems. Companies that succeeded with cloud were those that worked with vendors who understood existing operational constraints rather than dismissing them.

Data infrastructure followed a similar arc. The vendors that won were those who built systems that worked alongside existing enterprise data management practices—not those that demanded enterprises throw everything out and start fresh. Databricks, in fact, came out of this exact lesson. The company's core product, called a lakehouse architecture, was designed to integrate with how enterprises already manage data, governance, and security.

What This Means in Practice

Databricks built its business around a simple premise: enterprise AI needs infrastructure that fits within how enterprises already operate. That sounds straightforward, but it's different from much AI infrastructure designed in Silicon Valley startups, which often assumes you can rip out old systems and replace them with new ones. Enterprise environments are constrained by security rules, compliance requirements, and risk management practices that may look bureaucratic from the outside—but serve essential functions.

Enterprise success with AI tends to come from implementations that work within those constraints, not around them. When AI projects stumble in large organizations, it's often because the AI system clashed with existing processes, governance rules, or compliance requirements, not because the AI wasn't smart enough.

Looking at what Tavakoli-Shiraji is likely to address, the session positioning suggests he'll examine how AI practitioners can better understand and respect enterprise operational reality rather than override it.

The Bigger Picture

The timing of this talk matters. The enterprise software industry is recognizing that AI adoption requires different thinking than consumer AI applications. Consumer AI can prioritize speed and constant iteration. Enterprise AI has to balance innovation with risk management and regulatory compliance from day one.

Consumer AI applications can make mistakes, learn quickly, and ship updates. Enterprise applications often can't afford that. A bank's AI system for loan approval or a healthcare company's AI for patient care both have to work reliably on the first deployment, not iterate their way to reliability.

The session also comes during the annual budget planning cycle, when enterprise IT leaders are deciding where to invest in 2027. That makes it a strategic moment for these conversations.

For Practitioners and Builders

For people working on enterprise AI projects—whether you're in a startup, a vendor, or inside a large company's IT department—this framing offers something useful. It suggests that many enterprise AI challenges aren't failures of technology or organizational willingness. They're failures of assumption. If you approach enterprise AI as if enterprises work like startups, or like consumer apps, you'll run into problems that aren't really problems at all—they're just the actual shape of enterprise work.

The conversation also reflects a maturation of the AI industry. Early-stage technology tends to attract people who want to sweep away old ways of doing things. As that technology moves into production at real scale, it turns out the old ways often existed for reasons. Understanding those reasons, and building around them, is where real progress happens.