The AI Deal Killer Nobody Talks About
Databricks co-founder Arsalan Tavakoli-Shiraji says enterprises aren't afraid of AI — they're afraid of what happens when it breaks at scale.
Written by OutOfToken AI
June 6, 2026 · 4 min read · Synthesized from reporting by TechCrunch Startups · How this works
Enterprise AI has crossed a threshold. The boardroom debates about whether to adopt artificial intelligence are largely over — replaced by something far more consequential: rigorous interrogation of whether AI systems can be trusted to run inside complex, regulated, high-stakes organizations without creating new categories of risk. At TechCrunch Disrupt 2026, Databricks co-founder Arsalan Tavakoli-Shiraji put a sharp point on the dynamic that is quietly killing deals across the industry — not bad models, not poor demos, but operational fragility.
Beyond the Pilot Graveyard
For the better part of three years, enterprise AI adoption followed a predictable arc: a promising proof of concept, enthusiastic internal champions, a pilot that impressed in controlled conditions, and then a slow, painful stall somewhere between staging and production. Tavakoli-Shiraji's argument at Disrupt is that this pattern persists not because the technology is immature, but because vendors consistently underestimate what enterprises are actually buying. They are not buying a model. They are buying a system they can operate, audit, explain to regulators, and roll back when something goes wrong. When the sales pitch doesn't map to that reality, the deal dies — sometimes months after it closes.
Governance Is the New Evaluation Criteria
What's changed heading into 2026 is where enterprise scrutiny now lands. In earlier cycles, procurement teams asked whether the AI worked — whether it could summarize documents, generate code, or accelerate a workflow with measurable accuracy. Those questions haven't disappeared, but they've been joined by a second layer that many AI startups remain structurally unprepared to answer: Who owns the outputs? How is data residency handled across jurisdictions? What's the incident response protocol when a model hallucinates inside a customer-facing product? These are not edge-case concerns. For financial services, healthcare, and defense-adjacent sectors, they are gate-level requirements. Tavakoli-Shiraji's position is that the companies winning enterprise contracts right now are the ones that arrived at those conversations with answers, not roadmaps.
""Enterprise organizations are not rejecting AI. They are rejecting operational instability." — The distinction that separates scaling AI companies from those stalling after early momentum."
What Databricks Gets Right — and What the Market Still Misses
Databricks occupies an instructive position in this conversation. Founded in 2013 as an outgrowth of the Apache Spark project out of UC Berkeley, the company built its early identity around making large-scale data processing accessible to engineering teams. That foundation — deeply embedded in data pipelines, lakehouse architecture, and enterprise-grade governance tooling — positioned it unusually well for the AI deployment era, where the difference between a working model and a deployed model often comes down to data quality, access control, and observability infrastructure. Tavakoli-Shiraji's message at Disrupt carries implicit weight because Databricks has navigated the exact transition it's describing. Startups without that infrastructure layer face a harder sell: they must either build it, partner for it, or convince enterprise buyers to accept its absence. Increasingly, that last option is off the table.
The enterprises writing the largest AI checks in 2026 are not the most adventurous — they are the most prepared. For founders still optimizing their pitches around capability benchmarks and model performance stats, Tavakoli-Shiraji's Disrupt appearance is a useful correction. The companies that will define enterprise AI over the next five years won't necessarily have the best models. They'll have the most defensible operational story — and the proof that they can scale inside organizations that cannot afford to get it wrong.
Editorial Note
TechCrunch is a reputable technology publication with established credibility for covering industry events and executive perspectives. The claim about enterprise AI shifting from evaluation to deployment safety aligns with documented industry trends in 2024-2025, where enterprises increasingly focus on governance, compliance, and risk management. However, the specific event and quotes cannot be verified without accessing the full article or event coverage.
Claim Tracker
AI-assessed
Described as occurring 'for the better part of three years' but no empirical data provided to support this specific pattern claim
The article references a future event (2026) which appears to be fictional or the article contains a date error
Vague claim without supporting data about current state of enterprise decision-making
Attributed to one speaker but presented as industry-wide pattern without broader validation
Ask AI about this story
// discussion
sign in to join the discussion