Essay · AI · Company Building
Why companies that sell improvement outlast companies that sell machines · Juan Vegarra
Here is the conclusion first, because it changes how you read everything after it. The machine gets you into the room. The loop is what the customer is actually buying, whether they know it or not. A product that does not improve with use is a feature waiting for its commoditization date. In every industry I have built in, the company that won was not the one with the best machine on the day of the sale. It was the one whose machine was better a year later without anyone having touched it.
I wrote elsewhere about why a dataset is not a moat and what a real flywheel requires. This piece is about the other half of that argument, the half most teams skip. If the loop is the moat, then the loop is also the product. And once you say that sentence out loud, it rearranges the whole company: what you sell, how you price it, who counts as the production line, and what your board should be asking you in the quarter you just closed.
Start with the uncomfortable physics of selling machines. Whatever your instrument does today, a competent competitor will do most of it within a product cycle or two. Resolution gaps close. Speed gaps close. Feature lists converge because feature lists are visible, and visible things get copied. The spec sheet is a photograph of a race at one instant, and buyers who choose on the photograph are choosing the past.
What does not converge is the rate of improvement. A competitor can copy where you are. They cannot copy how fast you are getting better, because that speed lives in something they cannot see from the outside: the loop that runs from use, to data, to learning, back into the product. Two companies can ship identical spec sheets on the same day, and if one of them learns from every unit in the field and the other learns only from its own lab, they are not the same company. They are not even in the same business. One sells a position. The other sells a derivative.
The spec sheet is where you are. The loop is how fast you are moving. Competitors copy positions. They cannot copy velocity.
The lesson did not come to me from AI. It came from drill core. In mining, exploration drilling is the most expensive data acquisition most people have never thought about, thousands of dollars a meter to pull a cylinder of rock out of the ground. I watched junior companies on similar ground diverge over a decade, and the difference was never the drill. Some teams treated core as rock samples: assay it, report it, store it in a shed. Others treated every hole as an update to a living model of the deposit, where each result re-ranked where the next hole should go. Same instrument. Same spend per meter. One company's knowledge compounded and its targeting got cheaper and sharper every season. The other paid full price for every lesson, forever.
I saw the same physics at Microsoft, in a channel program that grew to roughly a billion dollars. The asset was not the program guide. It was the feedback cycle: tens of thousands of partner engagements teaching us what worked, folded back into how the program was designed, priced, and taught, on a cadence competitors could not match by hiring more people. And the pattern is everywhere once you see it. Engine makers stopped selling turbines and started selling thrust hours precisely because the telemetry loop, not the metal, was where the margin lived. The industries change. The physics does not.
Now take the sentence seriously inside the building. If the loop is the product, the team that runs the loop is not a support function. It is the production line. Most hardware companies get this exactly backwards. The data and learning work sits in a corner of R&D, funded as overhead, staffed with whoever can be spared, while the company's identity and budget concentrate on the next physical unit. That allocation quietly declares that the company sells boxes, whatever the pitch deck says about platforms.
A company that means it organizes differently. Someone senior owns the loop end to end, from what gets captured in the field, to the rights that make the data usable, to the cadence at which learning returns to the product. That last word matters. Returns. Data that flows in and never flows back out as a better product is not a loop. It is a landfill with good intentions. The test of the org chart is simple: point to the person whose job is the rate of improvement, and point to the budget line that pays for it. If the answer to either is nobody in particular, the loop is a slide, not a product.
Selling a machine once, at a margin, is pricing the photograph. If the product genuinely improves with use, a one-time sale gives away every future version of the product for the price of the current one. The pricing that matches the physics is recurring: the customer pays for access to a product whose value rises over the term, and the vendor is paid for the improvement it keeps delivering. This is not financial engineering. It is honesty about what is being exchanged.
And it cuts the other way, which is what keeps everyone honest. Recurring pricing means the renewal is the loop's report card. A customer who re-signs is testifying that the product they hold today is worth more than the one they bought, and a customer who walks is telling you the wheel stopped turning. Companies that sell boxes learn the truth at the next product cycle, years late. Companies that sell the loop learn it every renewal quarter. Faster truth is worth the discomfort.
The demo shows today's product. The contract should sell next year's. That distinction sounds subtle and changes everything about how serious buyers decide. A procurement process built on the spec sheet will happily choose the competitor whose photograph looks marginally better this quarter. The counter is not to argue the photograph. It is to change the question: ask what this product will be in three years, and ask what mechanism guarantees the answer.
The vendors selling positions have no answer beyond a roadmap, which is a promise about what their engineers might do. A vendor selling a loop has a mechanism: here is what the product learns from use, here is the cadence at which that learning ships, here is what last year's customers received without paying for it or asking. One of those answers is hope with a timeline. The other is physics with a track record. Buyers can tell the difference when someone shows them where to look, and teaching the buyer to ask that question is the highest-leverage sales work a company like this can do.
None of this is free, and the teams that have built loops will recognize the costs. The loop must be designed in at capture, with the rights to use what is captured, because retrofitting either is somewhere between expensive and impossible. In regulated fields, medicine above all, the loop runs through validation, which means improvement ships on the timescale of evidence, not the timescale of software. That is slower. It is also why, once turning, a regulated loop is so much harder to chase: the pursuer inherits the same clock without the head start.
And there is a failure mode on the other side worth naming. A loop is not an excuse to ship a weak product and promise compounding later. The product must be worth buying cold, on day one, before the wheel has turned once. The loop decides who wins the decade. The day-one product decides whether you are around for it.
If you sit on the board of a company that claims this story, four questions will tell you in an hour whether the loop is real. What did the product learn last quarter without an engineer touching it, and where is that measured. Who owns the data at the point of capture, and would that survive a customer's lawyer reading the agreement. Which budget line pays for the loop, and what happens to it in a cost cut. And if the company stopped shipping new features for a year, would customers still renew, and why.
Strong answers name numbers, owners, and mechanisms. Weak answers name intentions. And the last question is the one that separates them fastest, because a company whose renewals depend entirely on the next feature is a company selling photographs, however sophisticated the camera.
Every industry I have entered had experts who could describe the current order perfectly: what the machines do, what they cost, who buys them, and why. The order they defended was built on products that hold still. The reorder underway now rewards products that do not. The instrument opens the door. The loop keeps it open, and widens it, and eventually locks it behind you.
The feature is what you demo. The loop is what they renew. Build the second one, and the first takes care of itself.
Juan Vegarra is the author of An Outsider's Playbook (forthcoming). The views here are his own. Companion essay: The Data Moat Nobody Has · More from the Notebook · Continue the conversation
READ NEXT