Real moats pile up while you operate

Ask a founder what protects the company and you get a list of things they have. We know this market cold. We sit on data nobody else has. We’ve hired the experts, tuned the model, built the relationships. Every item is something you can point to right now, already in hand. Hold onto that list, because the evidence coming out of the AI era is rewarding a different kind of advantage, and it isn’t on it.

Start with what changed in the world. By the middle of 2026, about 78 percent of US firms were actively running AI in production [verified]. Owning the tool stopped being an edge once almost everyone owned it. So McKinsey, the consulting firm, went looking for what still separates winners, and sorted the answers into three tiers. Basic automation is table-stakes, the baseline everyone has. Better instrumentation, the ability to see what your systems are doing, is the emerging edge. And the durable moat, the one at the top, is what they call learning velocity: how fast an organization can deploy, measure, and improve compared to its rivals [reported]. Companies that rebuild their operations around AI rather than bolting it on report operating-profit gains of 10 to 30 percent, averaging near 20 [reported].

A second voice, from a different corner, landed on the same spot. Anish Acharya, a partner at the venture firm a16z, told Lenny’s Newsletter that startups no longer design their moat up front. They discover it by running. And the asset that compounds, he said, is a company’s proprietary logs of the model’s reasoning traces, the records of how the AI worked through each task, piling up over time [reported].

Both land on the same point. The durable AI-era moat is something you accumulate by operating. Domain expertise, the asset every strategist still names first, is the one thing on that list you can go out and buy.

Line up the candidates the field keeps naming. Learning velocity. Proprietary feedback loops. Reasoning-trace logs. Domain expertise. Three of those four share a property the fourth doesn’t. They pile up through use. Velocity compounds because you iterated. Feedback loops deepen because usage kept flowing back into them. Trace logs stack up, in Acharya’s own word, over time. You don’t have any of them on day one. You have them because you ran the loop for a year, and a competitor who starts today cannot write a check for the year you already spent.

Domain expertise is the odd one out. It’s a stock. You can hire the experts, license the data, buy your way into the market’s knowledge. At 78 percent adoption most of your competitors already have comparable access. Expertise sits there. It doesn’t grow just because you kept operating.

Now the belief this cuts against, stated fairly, because plenty of smart people hold it. AI pushed the cost of building close to nothing, so the edge moved off writing code and up into knowing what to build, and knowing what to build is your domain understanding. Deep expertise, the argument goes, is the thing agents can’t replicate and competitors can’t fake. It’s a good argument. It’s why “go deep on your domain” has been the standard advice for two years.

The strongest version of it says expertise and velocity aren’t rivals at all. Expertise is the ground; velocity is what you do on it. You can’t iterate fast on a market you don’t understand, because the loops that compound run straight through domain judgment. So the moat is still domain-anchored, and accumulation is just how you cash it in.

That holds until you look at the world McKinsey is actually describing. Its whole premise is a field where domain access is matched: nearly everyone deployed, everyone holding similar models and similar knowledge of the same market. Set two firms down in that world. Both understand the domain. One has run the loop a year longer and carries the traces and the faster cycle to show for it. That firm pulls ahead, and what it pulls ahead on is the pile it built. The expertise both of them share can’t be the thing that separates them. Expertise is the entry fee. The prize is what you build after you’ve paid it.

This is also where the advice to “measure everything” finally fits, without being the whole story. Instrumentation is how a static stock turns into a moving one. The moment you instrument your domain expertise, tying what your experts know to what actually shipped and sold, you stop banking the expertise itself and start banking what accumulates from running it: the measured loop and the record that grows each cycle. The expertise was the feedstock. The pile it helped you build is the moat.

So here’s a test you can run in your next planning meeting. Take the list of things that supposedly protect you and go down it one item at a time, asking a single question of each: did we have this on day one, or did we build it by running? The first kind gets cheaper as the tools spread and everyone catches up. The second kind is what a year of operating buys that money can’t.

There’s a harder question underneath it. If the durable moat is denominated in accumulation, then “domain expertise is the moat” describes a world that’s already closing: the one where knowing the domain was itself the scarce thing to own. At 78 percent adoption, that world is mostly gone. So change the question you carry out of the room. Understanding your market better than the company across town was the old test, and it stopped deciding the outcome the moment almost everyone understood their market too. The useful question now is simpler. What are we accumulating by operating that a competitor starting today can’t buy back?