Where are we, really, in the enterprise AI adoption cycle? What’s working, what’s holding big companies back, and what do founders need to understand to succeed in selling into the enterprise?
Over the past few weeks, we hosted two dinners to get closer to those answers. The first dinner brought together our portfolio CEOs for small group discussion sessions and a fireside chat with Arsalan Tavakoli, co-founder of Databricks, and Rajat Taneja, President of Technology at Visa. The second convened our early-stage founders and three enterprise CIOs from our operator network: Karthik Chakkarapani of Zuora, Saket Srivastava of Asana, and Sumit Johar of BlackLine.
Together, the two events offered a chance to hear from both sides of the table. Arsalan sees hundreds of enterprise AI deployments from the infrastructure layer up. Rajat, Karthik, Saket, and Sumit are the ones approving the spend and deciding what actually makes it to production.
One thing we heard loud and clear: everything is moving much, much faster.
According to Arsalan, over 80% of databases are now created by agents. The time between a piece of code being published and an exploit being found has shrunk from weeks to hours. Usage that used to accumulate over months is now spiking in days.
For founders, this pace is double-edged. The assumption that you have a long runway to iterate before a competitor catches up is increasingly wrong. The feedback loops are shorter and the surface area for disruption is wider. If the pace of change feels overwhelming right now, it’s probably the slowest it’s ever going to be.
This same speedup works against incumbents. Large players have inherited processes, organizational inertia, and existing products to protect. Startups carry none of that baggage.
At moments like this, the biggest risk for founders is not acting quickly enough. Action is what creates opportunities for learning, and your pace of learning is one of the most important variables you control. This also means you need to stay low ego, and be ready (or even actively plan to) to throw out aspects of what you’ve built as the technology improves.
Here are six additional pieces of advice that I think every founder in our network, not just the ones who were there, should hear.
⚡ 1. AI is real, and it’s happening.
For much of the last two years, many enterprises were like deer in the headlights when it came to AI. They launched hundreds of pilots to signal that they were “doing AI,” without a clear sense of what those efforts were supposed to produce.
There’s now a real appetite to put AI into production, and the questions enterprise buyers are asking have gotten harder and more specific. If we put a dollar in, what do we get back? When does this actually change how we work? Nobody has it fully figured out, but in 2026, the expectation is that founders need to have answers.
In Arsalan’s framing, most companies are moving through four stages:
Stage one: Not using AI at all. Arsalan was clear that “I don’t count using ChatGPT as using AI.”
Stage two: Sprawl. Everyone has multiple pilots running, but there’s no clear strategy behind them.
Stage three: Measurable productivity gains. Clear, quantifiable business impact starts to materialize.
Stage four: Process redesign. Businesses rethink how work gets done in light of what AI makes possible.
The vast majority of enterprises are still stuck at stage two. They have activity, but not outcomes.
As Arsalan explained, the companies getting real value aren’t the ones with the most pilots. They’re the ones who’ve picked a real business problem, built the data infrastructure to support it, and had the organizational will to push through process redesign.
Sumit shared his perspective from the CIO seat. “All CIOs are under a lot of pressure to drive innovation in their organizations, but many of them are ill-prepared. Almost no one can say, ‘I figured it out and everybody else is behind me.’” CIOs are looking to founders as “partners who will guide us.” In his experience, the founders who win early enterprise deals are the ones who do enough research to have a point of view before the meeting starts and who treat the CIO as an ally, not a gatekeeper to get past.
⚠️ 2. The pilot trap is a big challenge, but there are concrete ways to avoid it.
AI pilots are ubiquitous right now, and they’re expensive to run, for both sides. Founders are dedicating field engineering resources, and enterprises are spending internal cycles on scoping, security reviews, and change management. Both are finding it much easier to get a pilot started than to get it over the finish line.
One main reason why is that pilots are almost always too clean to matter. “Pilots are very pristine, they’re beautiful, they’re in a lab. They always work well, but production is where you get a cross section of real life - and real life is messy,” Rajat explained. The danger is that pilots succeed just enough to feel like progress while the harder work of actual deployment keeps getting deferred.
Rajat drew an analogy to Tesla. Instead of offering better test drives, they let you take the car home and use it for several days. The conversion rate jumped up because drivers could test the car against their actual lives, not a curated simulation of it. The companies getting real traction from AI are doing the equivalent: testing against real production traffic with real data, real edge cases, and enough control to revert quickly if something breaks.
Knowing which use cases to pilot matters just as much as how you run them. To identify which AI use cases are ready to deploy, Rajat looks for five things: bounded scope, high-quality data that’s easy to connect to models, a measurable and precise outcome, human judgment in the loop, and a workflow with significant cognitive repetition. If a business maps their use cases against these criteria, the ones that check every box will be obvious wins. For Visa, that’s meant coding productivity, customer service, and dispute handling.
For founders, the takeaway is that enterprise buyers aren’t evaluating your technology in the abstract. They’re asking a much more specific set of questions: does this work in our business, on our data, with a measurable outcome we can point to? The founders who escape the pilot trap are the ones who start with the right use case and design their pilots to answer them, not just validate that their product works in a lab.
🛡️ 3. Agent observability and governance is an urgent problem.
Security and governance concerns are another big reason why pilots stall.
Large companies now have hundreds of agents running across the organization, many of them built by employees without engineering backgrounds. Most are still working out how to maintain visibility into what those agents are doing, whether they’re interacting with each other in unintended ways, and whether they’re compromising the company’s underlying data and permissions model.
Saket captured the challenge well: “We really need to start thinking about the lifecycle of an agent as we’ve been thinking about the lifecycle of a worker. When do you onboard them? How do you retire them?” Most enterprises don’t have good solutions yet. Audit trails, controls over what agents can access, identity verification, and real-time observability are all still works in progress.
Rajat’s view is that solving this is where the next wave of competitive advantage gets built. As the model layer commoditizes, governance becomes the differentiator: “The value is moving into this control plane that over time will differentiate companies that are using similar products.”
Arsalan laid out the stakes with a story from inside Databricks. An agent at Databricks, told the system was at capacity and to delete something, found the nearest thing it had permissions to delete and brought the whole system down. It was doing exactly what it was told. What was missing was the judgment that any human would have applied automatically.
The deeper issue is that every legacy enterprise application was designed around a human interface. When agents replace that interface, the architecture of enterprise software has to be rethought from the ground up.
That presents a massive opportunity for startups, and it’s what our portfolio company Arize is focused on. Their span volume, a measure of product usage, has gone up 10x in the last quarter alone: a sign of how quickly AI is moving into production, and how urgently enterprises need observability and evals to keep pace.
🏆 4. You still need to sell promotions, not software.
For all the changes AI has brought, the fundamentals of enterprise sales haven’t changed. Trust remains the decisive variable. Relationships still matter. And the relationship that matters most isn't with the organization - it's with the human being inside it whose career is on the line.
“Trust is a compounder,” Rajat said. Once you’ve earned it, enterprises will let you test in real production conditions rather than keeping you confined to a sandbox.
This trust isn’t built from a polished deck or a well-rehearsed pitch. It comes from showing up consistently, delivering on what you promised, and being the kind of partner a CIO is willing to put their name behind. Reference customers still carry weight. So does the ability to show the product rather than simply talk about it.
As Rajat put it: “Companies are not buying AI, they’re buying a benefit to themselves or their customers. So what is the business outcome that you are bringing as part of your theory, which distinguishes you from others?” Being able to show that concretely - through the product itself, not a slide deck - is what separates the vendors who get championed from the ones who get politely passed over. “Throw out your PowerPoints,” he advised. “Just show the product.”
His comment points to an old line in enterprise software that’s still very much true: you have to sell promotions, not software. The person who champions your product internally isn’t doing it out of abstract organizational loyalty. They’re doing it for a human reason, because it advances their career, and they’re absorbing real risk on your behalf to do so.
From their perspective, the baseline expectation is that your product will disappoint. Broken promises, failed implementations, vendors who disappear after the deal closes: these are the norm, not the exception. In my experience, 80% of the time there’s more downside than upside to adopting something new. That means the upside has to be 5x to 10x the downside to make the math work.
My advice to founders: before you walk into any enterprise sales conversation, know whose career your product could make. Find out what that person is measured on and what a win looks like for them. Then orient everything - the pilot design, the success metrics, and the check-ins after the close - around making that person look good. Do that well enough, and your champion will sell your product internally better than you ever could.
🤝 5. Don’t sideline the CIO.
“A procurement decision about software is actually a group decision. There are multiple stakeholders involved,” Sumit explained.
The champion who carries your deal internally almost always lives in the business, not in IT. Sumit estimated that 80% of software requests now originate from line of business. Saket put it at 60-70%. Karthik was the outlier at roughly one third: a reminder that the pattern isn’t uniform across every organization or industry. Still, there’s almost always a director-level leader in the business who has the problem, the budget, and the career incentive to make your product succeed.
Finding that person is only half the job. The other half is making sure the CIO is with you from the beginning. Each CIO was emphatic about what happens when founders attempt to work around them. “A lot of times you try to go all the way with the business and then try to bring in IT to check the box,” Saket said. “You’ve actually turned us around in a way that might not work in your favor.” Rather than accelerate the sale, going around IT can poison the relationship at exactly the moment you need it most.
AI is moving faster than any organization’s internal evaluation process, and that speed has created genuine anxiety at the top. “The mandate CIOs are getting is, ‘you be the referee of what makes sense, because we want more AI, but it’s also a little scary,’” Karthik relayed. A CIO handed that mandate is actively looking for founders who can help them navigate it. The mistake is treating them as the last stop in the sales process rather than a partner from the beginning.
🏗️ 6. The real unlock is process redesign.
Even when the use case is right and the pilot succeeds, there’s a harder problem waiting. Layering AI on top of existing processes produces incremental gains. As my partners Joanne and Leo argue in their “great reorg” p.o.v., the true value of AI can only be realized when you redesign the way the business gets work done from the ground up.
Arsalan’s four-stage framework puts process redesign at stage four, and in his experience across hundreds of enterprise deployments, no large organization has gotten there yet.
As Joanne and Leo explain, most companies are still in the equivalent of the cars-on-dirt-roads phase of AI. Early automobiles entered streets built for pedestrians and horses. The full gains of the car came later, when roads were paved, highways were built, and urban life was reorganized around the assumption of widespread car ownership. AI is at the same inflection point. Individuals are moving faster, but faster individuals don’t automatically add up to a more productive company.
Rajat added: “Before the process, the leadership mindset and the culture has to be part of the change.” You can have a successful pilot and still stall out if the organization isn’t genuinely committed to working differently.
So how do you make those org-level changes?
Arsalan’s advice is not to try to gradually shift an existing team’s habits. It’s to find a group where you can introduce AI in a way that makes the old way of working impossible. “If you really want to change an org, you have to say: if you have AI, how do you do things from the ground up? You have to have a group where you shock the system enough that they have no choice but to redesign it.” Trying to transform an existing team from within, he said, rarely succeeds. The conditions have to force the redesign.
Move fast, but know where you’re going.
Getting out of the pilot trap, earning trust with CIOs, guiding customers with process redesign: none of this is possible without a clear thesis about what you’re building and why.
The volume of noise in the AI market right now is unlike anything I’ve seen before. Arsalan echoed this. New models ship weekly, competitors pivot overnight, and every morning brings a fresh round of announcements designed to make you feel like you’re falling behind. The temptation is to treat all of it as signal. But if you do that, “all you’re doing is emulating what other people do,” he argued. “You won’t survive as a company.”
The founders he most respects operate differently. They read everything, talk to everyone, and stay genuinely curious about what’s happening around them, but they run it all through a clear thesis about where they’re going and why. “The best founders I know have conviction and a vision. They absorb all the information, but they stay focused on how they execute on that vision. They live and die by it, and then know when it’s time to pivot.”
The pivot still matters. But there’s a difference between new information that genuinely challenges your core assumptions and new information that just makes you anxious. A competitor shipping a new feature is often the second thing. A fundamental shift in what customers need, or in what’s technically possible, is the first. Founders with a clear thesis can tell the difference between the two quickly.
Know what you’re building, why it matters, and what would actually change your mind. Stay curious, stay humble, and push yourself to move faster than you think is possible.
over 80% of databases now created by agents. the time from code published to exploit found down from weeks to hours. if the pace feels overwhelming right now it's probably the slowest it's ever going to be. worth reading the whole thing
The agent lifecycle framing from Saket is the sharpest insight here. Enterprises have decades of tooling around the human employee lifecycle onboarding, access reviews, offboarding — and almost none of it maps cleanly to agents that can be spun up by a non-engineer in ten minutes and never formally decommissioned. The governance debt accumulates quietly until something breaks in production.
Rajat's five criteria for pilot-ready use cases are also worth keeping: bounded scope, high-quality connected data, measurable outcome, human in the loop, and cognitive repetition. That's a more honest filter than most teams apply before committing engineering resources to a pilot.
Writing about what happens after the pilot succeeds, specifically how production AI systems degrade silently in ways that don't show up in any eval in my newsletter, if this resonates.