Built for the 85%: Where the Durable AI Companies Will Actually Come From

On what a Box CEO, a veteran VC, and a chart of 578 AI rounds all happen to agree on.

Over the last two weeks, two independent pieces of analysis landed on essentially the same observation — and together they describe the shape of AI software companies that will still matter in five years.

The first was a 20VC conversation with Aaron Levie, the CEO of Box, making the case that almost everyone in tech is thinking about AI wrong. The second was a study from Mighty Capital — 578 AI companies that raised rounds of $50M or more between September 2024 and March 2026 — classified by what actually defends each one. Igor Ryabenkiy's commentary on the study landed the sharpest line: "data is the new oil" has quietly started to leak.

Reading the two together changes how you should think about where AI value is accruing — and, if you're building in this space, where to position.

1. Tech is 8% of the economy, not 100%

Levie's most forceful point is one the industry keeps refusing to absorb: tech is 8 to 15 percent of GDP. Everything Silicon Valley discusses — which frontier lab is winning, which coding tool is hot, which agent framework has momentum — concerns a single-digit slice of the actual economy.

The rest of the economy — the other 85 percent — is John Deere and Eli Lilly and Bank of America, but also tens of thousands of mid-sized wealth firms, law practices, property operators, clinical research groups, industrial distributors, commercial real estate portfolios. These companies have always had structurally worse technology than tech has had. Not because they're less ambitious. Because the software industry wasn't built for them.

What happens when the other 85 percent finally gets access to the engineering leverage tech has always had?

That's the interesting question. And Silicon Valley keeps not asking it, because Silicon Valley is talking to itself.

2. Data was the new oil. The oil is leaking.

The Mighty Capital study classified all 578 companies by what actually defends the business. The numbers are striking.

Regulatory moats: 32.9% of companies, 14.5% of capital deployed. Workflow embeddedness: 28.4% of companies, 24.1% of capital. Physical infrastructure: 24.9% of companies, 28.4% of capital. Those three categories hold 86% of the total. The interesting absence is what isn't in them.

Data moats — the competitive advantage that dominated AI strategy talk for the last three years — captured 1.6% of capital. Twenty-four companies out of 578. The lowest median valuation in the dataset.

The conclusion is uncomfortable for a lot of 2023 pitch decks: data, in itself, is no longer a moat. Data still matters enormously. But data sitting passively somewhere — a proprietary dataset, a labeled corpus, an exclusive feed — defends nothing on its own. What defends a business is data that's embedded in a product, improving through use, and wired into decisions people are actually making.

That's a different thing. And the market has started pricing accordingly.

3. The workflow trap

Workflow embeddedness is where the capital is going — almost a quarter of all $50M+ AI rounds. That's the bucket most ambitious AI startups are building into. And that's where Ryabenkiy's sharpest warning lands.

Because workflow embeddedness, in 2026, is cheaper to replicate than it's ever been. A competent competitor can prototype a clone of most AI workflow products in two weeks. The defensibility isn't in being embedded. The defensibility is in the cost of being un-embedded.

This distinction matters. A lot of what looks like a moat on a deck — "we're deeply integrated into the customer's operations" — turns out, on closer examination, to be a moat of the client's own convenience. Convenience doesn't survive the arrival of a better tool. What survives is exit cost: the structured memory the customer has built up inside the system, the audit trails they depend on for compliance, the operational logic that would have to be reconstructed from scratch to leave.

When people talk about compounding AI products, this is what they actually mean. Not "our AI gets smarter." The system accumulates things the customer cannot afford to walk away from.

4. The role nobody is staffed for

Levie's most concrete prediction is a number most of the industry isn't discussing: 500,000 to a million new jobs in a specific category over the next five years. He calls it the agent operator.

It's the role that emerges when a company seriously commits to deploying AI agents. Half IT, half business analyst, deeply technical. Someone who understands how agents actually work — context windows, tool interfaces, model constraints, evaluation loops — and can walk into any functional team and redesign its operations for agents rather than for humans. They choose which workflows get automated, which stay manual, how the humans review agent output, what happens when a model drops and breaks everything.

Enterprises are going to hire this role in massive numbers. Fortune 500s will have departments of them by 2030.

Mid-market companies will not. A fifty-person wealth firm cannot justify one headcount who spends a year getting productive before anything ships. The role is too expensive, too specialized, and the talent too scarce. And yet the work still has to be done — the data is still fragmented, the agents still find the wrong document as often as the right one, the models still need continuous care and feeding.

This is the gap the entire mid-market is about to face. Every company outside of tech will need something it cannot staff.

5. Where the durable companies actually live

Overlay the three pieces of analysis — Levie's 85%, Ryabenkiy's moats, the agent-operator gap — and the intersection is narrow and specific.

The durable AI companies of the next decade will be the ones that serve the 85% of the economy tech has historically neglected, operate in regulated or accountability-bearing contexts where liability cannot be delegated to a model, embed deeply enough into customer operations that the system captures things the customer cannot afford to lose, and fill the agent-operator gap as a service because their clients cannot hire their way out of it.

That's a different kind of company than what most AI startups are set up to be. It's not a frontier lab — those compete for the Scale bucket where capital is being concentrated. It's not a pure SaaS tool — those get cloned in two weeks. It's not a consultancy — those don't scale beyond headcount.

It's something more hybrid: a technology platform with consulting-led delivery, operating in regulated mid-market verticals, building the operational infrastructure its clients cannot build themselves, and compounding switching cost through structured institutional knowledge that accumulates across every engagement.

That configuration is exactly what Levie's observation about accountability chains describes — liability can't be delegated to a model, so the business model survives. It's exactly what the Mighty Capital data rewards. And it's exactly the function mid-market companies need filled but cannot fill on their own.

6. What we're doing about it

This is what Reformance is. We discover how a client's business actually operates, build an operating system that transforms and runs on that understanding, and stay with the system as it evolves. Every engagement deepens a platform that compounds across clients. Every client deployment becomes structurally harder to walk away from as the institutional knowledge inside it accumulates.

The verticals we work in — wealth oversight, legal and IP services, short-term rental operations, commercial real estate — sit deliberately in the regulated and accountability-bearing part of the map. The engagement model is the agent-operator function delivered as a service. The platform is what makes that service compound rather than restart with every client.

We didn't arrive at this positioning because we read two pieces of analysis in April 2026. We arrived at it by spending eighteen months actually doing the work. But it's useful when two independent sources — one a public-company CEO, one a VC reading a dataset of 578 funded companies — describe the shape of the category we're already in.

The takeaway

If you're a mid-market operator reading this, the question isn't whether AI will reshape how you run your business. It will. The question is what kind of company you want helping you do it.

The AI software vendor optimizing for seat counts isn't that company. The consultancy that will hand your COO a beautiful strategy deck isn't that company. The platform company that assumes you have an internal agent-operator team — you don't — isn't that company either.

What the 85% of the economy needs is something most of Silicon Valley isn't building: operational software designed for the economy you actually live in, built by people who understand the work your business does, staying with you long enough to make the system keep getting better.

Sources: Aaron Levie on 20VC with Harry Stebbings, April 2026. Igor Ryabenkiy, public commentary on the Mighty Capital AI Moats study — 578 AI companies, $50M+ rounds, September 2024 through March 2026. Both are worth reading in full.