Software is getting easier to build and harder to defend, and that tension explains a lot of what we're seeing in 2026.
Most AI outlooks turn into a sector map: AI in healthcare, AI in legal, AI in sales, AI in defense. Pick a vertical, add agents, and call it a thesis. That isn't wrong exactly, but it doesn't tell you much about which companies should exist or why any one of them will matter.
We're more interested in what happens when software can do the work, not just organize it.
For most of the SaaS era, a good product helped someone do a hard job. The customer bought access, adopted a workflow, and usually paid by the seat. Now the promise is becoming more direct. I don't want better bookkeeping software; I want the books closed. I don't want another security dashboard; I want the system continuously tested, the obvious issues fixed, and the real exceptions put in front of the right person. The interface still matters, but it isn't the whole product anymore.
Once software starts completing the work, it needs to understand context, hold the right permissions, remember what happened, and know when to involve a person. It has to earn the right to act and remain accountable for the result. Access to a model will be widely available; getting software to work reliably inside a real workflow will be much harder.
This should lead to more software, not less. We'll get more internal tools, more narrow products, and more systems built for work that the last generation couldn't economically serve. A few strong people can already test an idea and get it into production faster than a much larger team could a few years ago.
That's great for users and rough for undifferentiated companies. An obvious feature can be copied, bundled by an incumbent, absorbed by a model provider, or built internally. A polished interface on top of a shared model may be useful, but that alone doesn't make it durable.
The old SaaS playbook will still work in some markets; we just don't think it'll work as automatically as it used to. Build a product, sell seats, expand upmarket, and compound annual recurring revenue is no longer a complete theory of the business when software itself is becoming easier to produce.
A company needs a stronger reason to own its place in the workflow. That may come from distribution, a system of record, hard-won trust, permission to act, or a feedback loop rooted in the work itself. These advantages are hard to bolt on later because they come from how the product is used and what the customer allows it to do. "We applied AI to a known category" explains how a product was made. It doesn't explain why the company will matter.
We think 2026 will make that gap more obvious. A compelling demo and fast early adoption created a lot of attention in the first wave. Now buyers know more of what to ask for, model providers and incumbents are shipping faster, and the novelty window is getting shorter. The application layer isn't disappearing. A company simply has less time to turn a new product into a position that lasts after the feature stops feeling new.
The business model may move with the product. A seat-based tool gets paid for access, while a system that closes books, handles claims, or verifies compliance steps may be able to charge for completed work or reduced risk. Some of the best companies will still look like conventional software businesses. Others will combine software and operations in ways that would have looked too service-heavy in the last cycle.
We're interested in that possibility, but we don't want to turn it into a universal rule. Buying a service business and automating it from the inside may work in certain markets; it's one tactic, not the thesis. The larger opportunity is to open a capability that used to be too expensive, too difficult, or impossible, because that's how software creates a new market instead of merely cutting the cost of an old one.
Cheaper building also makes the difference between good and mediocre judgment easier to see. The strongest technical founders can test more ideas, cover more product surface, and learn faster with a small team. Mediocre teams can produce more software too, but often it's just more noise. The bottleneck moves toward the judgment required to decide what should exist, where it should enter the market, and what has to be true for it to become a business.
This is what we mean by founders who are native to the stack. We don't mean people who can recite model benchmarks or attach an agent to a familiar product plan. We mean founders with practical technical taste: they understand what these systems can do, where they fail, how to make them reliable, and where that capability becomes economically important.
Their early hiring choices are part of the same test. Small teams have more leverage, but the best potential hires have more leverage and more options too, so the first five to ten people need enough range and trust to build while the product, market, and tools are all moving. Team formation is one of the first places a founder's judgment becomes visible.
In practice, this pulls us toward infrastructure and tooling that makes agentic software work in production, along with software that can own a sophisticated workflow and move close to a concrete outcome. On the infrastructure side, we're interested in evaluation, observability, permissioning, memory, and all the less glamorous machinery between a good demo and a system people can rely on.
Either way, we want to know what the product can do now that wasn't possible before, why this team is unusually equipped to build it, where the company earns permission to act, and what gets better with use. We also want to know what remains valuable as models improve and obvious features become cheap.
The founders we want to back are already behaving as if software changed again. They aren't waiting for the market to settle or adding AI to a static product plan. They have a specific view of what's now possible, and they're willing to risk being wrong in a specific way.
By now, "AI is big" is just the weather. The investment question is what becomes scarce.
We think judgment becomes more valuable as the underlying capability becomes more abundant. More of the economy will become buildable in software, but building something will be the easy part. The companies that matter will be the ones that earn the right to do real work and keep that position as the tools improve.