TechCrunch Disrupt 2026

Arsalan Tavakoli-Shiraji on why the best tech doesn't win enterprises

Arsalan Tavakoli-Shiraji· Co-founder & SVP of Field Engineering at Databricks at Databricks
·~34 min·English·TechCrunch
AgentsAI CompanyBusiness StrategyAI InfrastructureOpen Source
TL;DR

Databricks co-founder Arsalan Tavakoli-Shiraji on why enterprise AI is years behind the valley, and why the winners will own the context, distribution, and domain expertise around the model, not the model itself.

01The Reality Check

Everyone's a 3. Almost Everyone Is a 1 or 2.

<strong>99% of enterprises are stuck in phase one or two</strong> of a five-step AI adoption curve — even though, asked to rate themselves, they all claim to be a three leaning four.

And everybody loves picking. I'm three, leaning towards four. This a safe place. 99% of the ones I talk to are in bucket one and two, right?

— Arsalan Tavakoli-Shiraji, TechCrunch Disrupt
Key Insight
The self-rating gap is the tell: enterprises grade themselves on the demos they have seen, not the workflows they have changed. The real blocker is not model quality but the slow work of changing people and process, which is why 'agents everywhere' is years out, not quarters.
02The Founder's First Lesson

The Best Tech Doesn't Win the Room

Databricks walked in with the fastest product and still lost deals, because <strong>enterprise sales turns on people and process, not features</strong>.

Whereas as technologists, we think it's all about the tech and having the best tech and the latest product or feature.

— Arsalan Tavakoli-Shiraji, TechCrunch Disrupt
Key Insight
His own cautionary tale — playing lawyer, signing away unlimited liability to win a deal, then spending four years unwinding it — shows the real enterprise surface area is legal, security, and internal politics. A technologist who treats those as paperwork loses to a rival who treats them as the product.
03The Distribution Hack

Ride In on Someone Else's Paper

The fastest way past a buyer's procurement wall is to <strong>sell on a partner's existing contract</strong> — being first-party on Microsoft's paper let its customers adopt Databricks with nothing new to sign.

So, but that was a way that we kind of shortcut getting into the enterprise riding on their coattails.

— Arsalan Tavakoli-Shiraji, TechCrunch Disrupt
Key Insight
Distribution, not product, was the unlock. Being first-party on a giant's paper converted a cold, multi-quarter procurement slog into a warm introduction from a rep who already had the CIO's calendar. The lesson for a startup is to find whose coattails already reach your buyer.
04The One Yes That Matters

Fifteen People Can Say No. Find the One Who Says Go.

Enterprise sales is a gauntlet of vetoes, so the whole game is finding <strong>a single internal champion willing to fight for you</strong> and pull you past everyone else.

if you by yourself go ask for, you know, a specialist, they'll take 8 months, but if you get a really, really good doctor who picks up the phone, you can get there in basically like a week.

— Arsalan Tavakoli-Shiraji, TechCrunch Disrupt
Key Insight
A champion is not won with a demo; it is won by solving a painful problem they personally own. Once their name is attached to your success, they do your internal selling for you, which is the only kind of selling that actually moves a committee.
05What Agents Actually Need

The Four C's: Context, Choice, Cost, Control

Before an agent touches real work it has to be set up — and the four things enterprises insist on are <strong>context, choice of model, cost control, and a leash</strong>.

you give a normal coding agent without context, take your best one, you won't get better than 50% accuracy. With context, you can jump up north of 90%.

— Arsalan Tavakoli-Shiraji, TechCrunch Disrupt
Key Insight
Notice the asymmetry inside the four: choice, cost, and control are guardrails you can configure, but context is the one you have to earn, built from your own data and semantics. It is also the one that moves accuracy from a coin-flip to reliable, which is why context is listed first and is where the real work sits.
06The Model Trap

The Model Is Never the Moat

Chasing the best model is like speccing a PC by its CPU while the memory slots sit empty — <strong>for most tasks the model is already good enough, and it was never where a startup's moat lives</strong>.

you do not need, you know, kind of a more powerful model like you do not need Fable 5.5 for drafting your email. I don't care how complex it is, right?

— Arsalan Tavakoli-Shiraji, TechCrunch Disrupt
Key Insight
This reframes the 'thin wrapper' fear every founder has: a better model next quarter only eats you if features were all you had. The moat is knowing an industry's workflows cold — an oil company's path from upstream to downstream — because that lives outside the model and no training run captures it.
07The SaaS Scare

Give Me a Google Form, Not an App

The fear that enterprises would vibe-code their way off SaaS was overblown; they only <strong>customize the last mile</strong> and still buy anything that needs real compliance.

Why the f do I need an app? It's literally a question. Give me a Google form. Give me an email.

— Arsalan Tavakoli-Shiraji, TechCrunch Disrupt
Key Insight
The implication cuts against the SaaS-is-dead story: the part enterprises refuse to own — the support, the outages, the compliance — is exactly the part that is hard to build, which is why they keep buying it and vibe-code only the disposable UI.
08The Five-Year Horizon

95% of Tomorrow's Software Doesn't Exist Yet

<strong>Roughly 95% of the software we'll use in five years hasn't been written yet</strong> — because agents can finally decode the Cobol and mainframe code no living employee still understands.

I truly do believe five like 95% of the software that we will use five years from now doesn't exist today. That doesn't mean that vendors are getting replaced

— Arsalan Tavakoli-Shiraji, TechCrunch Disrupt
Key Insight
That order is why incumbents aren't simply deleted: the old system gets wrapped so agents can read it, not ripped out, and the new systems of record fill in beside it. Replacement by accretion is slower and messier than a clean migration — and far likelier to actually happen.