I recently listened to a Founders podcast episode on Kelly Johnson, the legendary Lockheed engineer behind the U-2 and the SR-71 Blackbird. Designed in the 1960s, the SR-71 still holds the record as the fastest air-breathing manned aircraft ever built. Over 3,000 missiles were fired at one in service. None ever hit. Its only defense was acceleration.
The episode is built around Johnson's autobiography and the operating system he created for Skunk Works: the rules, the 14 Points, the insistence on small teams and minimal paperwork. Most of it comes across as timeless management advice. But one line landed relative to what we argued in Part II of our Vertical AI series:
A breakthrough program is an organization before it is a design.
That is the argument of "Building the Moat" in nine words, from a man who died before the world wide web existed.
Why this line matters
In Part II, I argued that durable advantage in Vertical AI is not a feature, a model, or a UI. It's a set of organizational properties: the tacit domain knowledge in customer conversations, the cornered data you're uniquely positioned to capture, the best practices you encode, the system-of-action position you earn one deployment at a time. You don't build these the way you build a product. An organization accumulates them because it's structured to.
Johnson understood this sixty years ago in a different domain. His genius wasn't only aerodynamic; he never separated technical design from institutional design. His most important invention may have been Skunk Works itself: small empowered teams, the builder holding real authority, ruthless removal of anything that consumed time without advancing the work. Speed mattered, but the org was the moat. The planes were the output.
If you're building Vertical AI, that's the reframe worth sitting with. You're not assembling a product that happens to need a company around it. You're designing an organization built to capture context faster and deeper than any outsider can. Get it right and the product compounds. Get it wrong and no model will save you.
The parts of Kelly's advice that transfer cleanly
- Speed as a design requirement, not a side effect. Johnson's team delivered the XP-80 prototype in 143 days against a 180-day schedule. Our Part II line, that the window is measured in quarters rather than years, is the same claim in a different century.
- A few exceptional people over many average ones. The SR-71, the most advanced aircraft in the world, was built by 135 engineers while others used thousands. Johnson restricted headcount "in an almost vicious manner."
- Bureaucracy as an engineering variable. He treated reports, approvals, and meetings as drag, and pared away anything that used up time without advancing the project.
What needs inverting
- Secrecy. For Johnson, secrecy was both a moat and an accelerant. It kept outsiders from interfering. Your moat is the opposite. It comes from being maximally embedded and visible inside your customer's operations. The context you need lives in their workflows, not behind a wall.
- "Keep it simple, stupid." Johnson's KISS was a survival method for complex programs. But the thing that's simple in Vertical AI is the process, never the domain. Your moat is partly irreducible domain complexity: the taxonomy, the error-mode asymmetries, the liability lines. Strip the org and process drag and definitely use the KISS principle on your marketing messaging. But never strip the domain depth. That's where the moat lives.
And one tension worth naming directly, because it sits right at the center of both Johnson's model and ours:
Johnson's speed compressed a known build. He knew what an airplane was. Our moat thesis says the valuable asset, the accumulated "why" behind your customers' workflows, accumulates on a curve you cannot compress.
So speed doesn't sprint you to the moat. Speed gets you embedded early, because the clock on the irreplaceable asset only starts once you're inside the workflow. You're not racing to build the moat. You're racing to start the compounding.
The operator playbook
Here's how I'd translate Johnson's advice into concrete moves for a Vertical AI founder or operating team:
- Design the org chart around context capture, not function. Johnson put the design engineer, the mechanic, and the manufacturing process in one shop with no departments between them. The Vertical AI equivalent: the people who talk to customers, encode domain logic, and ship product can't be three teams separated by handoffs. Every handoff is where the why behind a workflow leaks out before it gets encoded.
- Give the builder real authority, and put them in the workflow. Johnson insisted designers fly-test their own aircraft: "I figured I needed to have the hell scared out of me once a year." You design differently when your own judgment is on the line in the real environment. Get your engineers and PMs into customer operations: on-site, in the support queue, watching a deployment go sideways live. The error-mode awareness from Part II, knowing which mistakes are catastrophic versus tolerable, isn't learnable from a spec. It's learnable from watching a customer's day go wrong.
- Treat process drag as a measurable cost. Johnson found a subcontractor with 40 people inspecting and reporting on work that 35 were doing. Audit your own ratio of people producing encoded knowledge and shipped product versus people coordinating about it. As AI drives build cost toward zero, the binding constraint is no longer engineering capacity. It's organizational friction. The winners will be the ones who noticed the constraint moved.
- Let AI resolve the small-team paradox. Johnson wanted tiny walled-off teams. Part II argues that when build cost approaches zero you can say yes to nearly every customization, turning the roadmap into a customer-service function. Both are right. Keep the core team small and high-judgment, Johnson-style, and use AI-assisted engineering as the multiplier that lets it service a breadth of workflows that used to require an army. The small team designs the system. AI handles the instances.
- Kill your own brilliant ideas on schedule. Johnson scrapped a hydrogen-powered plane deep into design once he saw it was impractical: "The measure of an intelligent person is the ability to change his mind." The idea you'll be most reluctant to kill is the impressive general-purpose capability that doesn't serve your vertical's last mile. Run a standing review: which investments deepen the moat (domain depth, cornered data, system-of-action), and which are demos that impress a board and change nothing about sales or retention?
- Make integrity a performance multiplier. Say no to the wrong customers. Johnson returned government money or declined work when he didn't believe in the result. The parallel: a logo outside your vertical, or whose workflow you can't go deep on, doesn't just fail to build the moat. It dilutes the cornered-data asset with noise and pulls the org off the compounding curve. Disciplined customer selection is moat construction.
- Punish concealed mistakes, never honest ones. Johnson allowed mistakes but demanded they be reported immediately, with a fix. For AI products in high-liability verticals, this isn't culture, it's survival. The error you don't hear about is the one that becomes an existential incident. Reward surfacing a model failure fast; make burying it the only unforgivable act.
The closing thought
In Part II, I wrote that positioning alone doesn't win. Execution against these priorities is what separates the companies that define their industries from the footnotes in someone else's case study. Johnson would have put it more bluntly. He spent decades trying to get other companies to adopt the Skunk Works model and watched almost none do it. Not because the formula was secret, he published it, but because most companies "will not pay the price in revised methods and procedures."
The intelligence layer is commoditizing. The execution layer is the prize. And the organization you design to capture it is the part competitors can't copy from the outside.
Go build it. Build the org first.
Related reading: Building the Moat, Vertical AI Part II. The Kelly Johnson episode referenced here is from David Senra's Founders podcast (#419), drawn from Johnson's autobiography, More Than My Share of It All.