Why the mid-market is about to get the advantage startups have had for twenty years

Every company that ate an incumbent's lunch over the past two decades had the same secret weapon, and it wasn't their fancy new app.

It was a person.

Tobi Lütke was writing Ruby when Shopify was still a snowboard shop. The Collison brothers were technical enough to make "seven lines of code" a sales pitch. Amazon put engineers in rooms where merchandisers used to sit alone. Different companies, same pattern: someone with deep technical judgment sat in the room where business decisions got made. Not consulted afterward. Not handed a requirements document. In the room, at the moment of decision, with a respected voice.

That person could hear a revenue problem and see the systems that would solve it. They could smell a doomed architecture two years before it sent the P&L sideways. They translated in both directions, board ambition into technical reality and technical reality into board language, every single day. Startups got these people by paying in equity and adventure. Big tech got them by paying in piles of cash.

Mid-market companies got them essentially never.

Not for lack of wanting. The math didn't work. A person with that profile runs $300K and up, when you can find one at all, and the old consulting model wrapped them in scaffolding: a partner to sell, a manager to manage, six juniors to type. Instead of an engineer in the room you ended up buying a pyramid in the hallway. Eight people, sixteen weeks, a deck at the end. So the $200M manufacturer made do with vendors and body shops, and somewhere in the building a quoting spreadsheet or a home-grown supplier database quietly became load-bearing infrastructure that everyone agreed not to look at directly.

But in 2026, the math has changed.

What’s actually changed (and what hasn’t)

I want to be precise here, because this is exactly where every AI pitch goes off the rails.

The tools are not magic. They aren't miraculous. They are not a "country of geniuses in a data center," whatever the keynote said. They're power tools: genuinely astonishing ones, with strengths and weaknesses like any tool, and an interface that's still primitive. The technology problems will get solved. The models will get better. We'll learn the management techniques to aim them properly. That part is a function of time, and betting against it looks a lot like betting against the internet in 2001.

But, here's what the tools actually deliver today, measured on our own engagements rather than a vendor's slide: 50 to 70 percent faster on boilerplate and scaffolding. Five to ten times faster on pattern-based migrations. Ten to twenty percent faster end to end on real bespoke software projects. On the architecture decisions that determine whether the whole thing lives or dies?

Roughly zero.

That zero is the most important number in this essay.

Why the zero

Deciding what to build is not a generation problem. A model can produce a thousand plausible roadmaps before lunch. It cannot want any of them.

Wanting is the whole game. Intent requires motivation, and motivation requires an environment of consequences: a payroll to make, a market that punishes hesitation, a board meeting on Thursday, skin in an actual game. That loop took roughly four million years of evolution to build, and no amount of compute has reproduced it. The models have become undeniably eloquent but they still don’t feel skin in the game.

So here's the trade the industry keeps pretending isn't happening. Programming, the typing part, is commoditizing faster than ever and will keep commoditizing. What's appreciating is everything the typing was always downstream of: engineering judgment, operational scar tissue, product instinct, the feel for where a market is leaning, the ability to stand between a CFO and a codebase and translate honestly in both directions.

Execution is getting cheap. Knowing what to execute is getting expensive.

Yes, it's a bubble. That's not the interesting part.

AI currently polls badly at dinner parties and tops every board agenda, often for the same reasons. There's froth everywhere: in the funding, in the punditry, in the bull case and the bear case alike. Fine. There was froth in railroads, too. In 1893, roughly a quarter of America's railroads went into receivership. The track stayed in the ground, and Sears built an empire on rails it never had to lay. Same story in 2001: the fiber companies died, and the dark fiber they left behind ended up carrying Netflix.

Bubbles kill business models. They rarely kill capabilities. The capability underneath this bubble is real and durable: intelligence you can deploy like a resource, metered, on demand, at two in the morning. That survives any pop. The winners of the next decade will be the companies that built on the thing that survives, not on the stock prices.

Which is why our bet doesn't depend on which way the hype breaks. If the technology plateaus tomorrow, the math above already works. If it accelerates, execution gets even cheaper and judgment gets even scarcer. Either way, the same asset appreciates: a senior human who can aim the machines at the right target.

What we're doing about it

We're rebuilding our firm around the two things that survive: deployable intelligence, and the people who can aim it.

We call those people Conductors. A Conductor is not an engineer with soft skills bolted on. It's someone who has lived inside businesses, boardroom to product owner's desk to the engineering bullpen, and carries ten to fifteen-plus years of shipped systems in their hands. They're insatiably curious, which is where deep understanding actually comes from. They know the AI tools well enough to use them hard and distrust them appropriately. They make human plus machine worth more than the sum of the parts.

Strip away the titles and the Conductor's real job is this: they hold your intent. They form it with you at the whiteboard. They encode it into specs and architecture decision records. They enforce it in every review. Then they transmit it, deliberately and in writing, to machines that have none of their own. Most of what the industry calls "context engineering" is exactly that: the discipline of transmitting intent to something that can't generate it.

One more conviction, earned the hard way across fifty-plus engagements: you can't build intelligent systems on broken foundations. Most AI projects don't fail because the model was weak (most software projects in general don’t fail because of technology); they fail because the data and foundations underneath were never ready. So we fix the foundation first and build the ambitious thing second - integration before intelligence.

What we believe

Stated plainly, so you can disagree with it:

  1. AI-assisted development will be how most software gets written. Fighting it is malpractice. Worshiping it is malpractice with better marketing.
  2. The tools are power tools, not employees. Use them with respect and suspicion in equal measure.
  3. Execution is getting cheap. Judgment is getting expensive. Hire, staff, and price accordingly.
  4. AI can't want what you want. Intent stays human. Build your plans around that.
  5. You can't build intelligent systems on broken foundations. Fix the data before you buy the dream.
  6. The firm that promises 10 to 20 percent and delivers it beats the firm that promises 10x, on every project that matters.
  7. The person who scopes the work should be the person who does the work. We've run that play since 2011 and AI-tooling finally made it scale.

Skin in the game

For twenty years, the best companies kept an engineer in the room where decisions got made, and for twenty years that seat was priced out of the mid-market. The price just dropped. The judgment didn't get cheaper; the scaffolding around it did.

We're not asking anyone to take this on faith. Faith is what the keynote sells. We're rebuilding a fifteen-year-old firm around this argument instead, which means we carry the one credential no model can: skin in the game. If we're wrong, we'll feel it before you do. We don't think we're wrong.