Enterprises aren’t saying no to AI. They’re saying no to operational chaos. That’s the read from Arsalan Tavakoli-Shiraji, co-founder and SVP of field engineering at Databricks, who’s set to break down the shift at TechCrunch Disrupt 2026, according to TechCrunch AI. His session has a blunt title: “The Enterprise Isn’t Broken. Your Assumptions About It Are.”
What stands out here is the timing. For years, AI startups lived in an experimentation market. A sharp demo, a strong model, a bold vision, and you’d land pilots and investor checks. That era is closing. Enterprises have stopped asking whether AI is exciting. They’re asking whether it’s safe to roll out across the whole company.
The pilot was never the hard part
Tavakoli-Shiraji’s core point lands hard: most enterprise AI deals don’t die because the model underperformed. They die because the buyer lost confidence in what deployment would actually require.
The market is full of pilots that worked and still went nowhere. Not a tech failure. An absorption failure. The organization couldn’t take on the operational cost of adopting the thing.
This is significant because a lot of founders are still optimizing for the wrong moment. They build for the first wow. Enterprises buy for what happens on day 90.
It’s a trust problem, not a benchmark problem
TechCrunch AI reports that the startups gaining real traction inside big companies share a pattern. They reduce uncertainty. They plug into existing systems cleanly. They create less workflow friction. They’re easier to govern, easier to explain to a security team, and easier to trust over time.
Less flashy than a benchmark win. But that’s the line now between attention and durable revenue.
The questions enterprise buyers are leading with have changed:
- What happens after deployment?
- How much operational change does this force on our teams?
- How does it affect governance and compliance?
- Can people actually adopt it at scale?
- What happens when the model fails?
Those used to be afterthoughts. Now they’re the buying decision itself.
Why his read carries weight
Tavakoli-Shiraji isn’t a pure technologist or a pure strategist. He was an associate principal at McKinsey advising enterprises on cloud and IT transformation. He also holds a PhD in computer science from UC Berkeley in networking and distributed systems. So he’s seen both the procurement room and the architecture diagram.
That dual lens is the whole argument. Enterprise AI success now depends on how technical systems collide with messy organizational reality: procurement cycles, governance rules, infrastructure limits, operational risk. The winners over the next few years may not have the smartest models. They’ll be the ones who understand how a big company actually absorbs change.
What this means for the next 1 to 3 years
The center of gravity in enterprise AI is moving from novelty to operations. Expect buying committees to get more disciplined, not less. Procurement and security teams are gaining veto power, and “it demos great” stops being a closing argument.
If you’re building or buying AI for the enterprise, here’s where to put your energy:
- Sell the day-after, not the demo. Show what running this in production looks like at month three, including failure modes.
- Treat governance as a feature. Explainability, audit trails, and access controls are now part of the pitch, not paperwork bolted on later.
- Minimize the change you ask for. The more workflows you force people to relearn, the more reasons the deal stalls.
- Map the org, not just the tech. Know who signs, who blocks, and what “safe to deploy” means to each of them.
- Plan for the model breaking. Buyers want the fallback story before they sign, not after the incident.
The builders who internalize this stop chasing the wow and start engineering trust. That’s the unglamorous work that turns a pilot into a contract. More detail on Tavakoli-Shiraji’s full breakdown is available at the original source.