Measurement is what turns an AI asset into a moat

You’ve accepted the first half of the AI story. The tools are cheap now, and everyone has them, so “we use AI” wins you nothing. The harder question is the second half. With the tools common, what actually keeps a competitor from catching you?

The consulting world has an answer, and it arrives as a menu. McKinsey and firms like Teraflow, which advise companies on AI strategy, now name a short list of durable advantages for the AI era. Go deep on domain expertise, so you understand your market better than a generalist holding the same model. Build a proprietary data feedback loop, where your own usage data trains models a competitor can’t simply buy. Or hold trust in a regulated field like finance or healthcare, where being the name people already rely on is its own barrier. Pick one, invest, defend it.

The menu is real. These are genuinely different assets, and they compound in different ways. But it stays silent about a precondition that decides whether any of them count, and one number drags that precondition into the light.

Roughly 73 percent of enterprise AI pilots never reach production [verified]. They never graduate from a trial into everyday use. Not because the technology failed. McKinsey’s research puts the cause elsewhere: no one owned the outcome, and no one had framed the returns in a way anyone could actually check. Read that slowly. Those companies had assets. Some had deep expertise. Some had proprietary data or hard-won trust. What they lacked was any way to tell which of their pilots was working, so none of it compounded. They owned the territory and had no way to measure it.

A moat in the AI era rests on one capability the menu leaves off: the measurement discipline that turns any asset from latent into bankable.

By measurement I mean instrumentation. The dashboards and evaluation harnesses that tie what a system did to what it produced. It sounds like plumbing. Watch what happens when you run each item on the menu through it.

A proprietary data feedback loop only compounds if you can measure which slice of usage actually improved the output. Without that, a “data moat” is just storage with a better name, and you’re paying to keep data you can’t tell is worth keeping.

Trust in a regulated industry works the same way. Regulators and enterprise buyers don’t hand trust to the company with the best story. They hand it to the system whose behavior can be audited and shown. Remove the ability to demonstrate what the system did, and the trust moat is a marketing line.

Domain expertise is the one the field has started to admit out loud. The phrase going around is that unmeasured domain expertise is indistinguishable from luck. Be honest that this exact claim still carries an unverified tag [unverified]. You don’t need it proven to see the shape, though. Put two competitors side by side with equally deep expertise, both using AI. Their expertise is identical, so it can’t be the thing that separates them. Whether they can measure that the expertise changed the outcome is what’s left.

So the three items that looked like separate moats share one dependency: measurement. That single fact reorders the menu. The other three are items on the menu. Instrumentation is what lets you read the menu at all. Every other advantage stays latent, present but impossible to bank, until you can measure it’s paying off.

The buyers already moved. Direct financial impact now accounts for about 22 percent of the primary metrics enterprises use to judge AI, nearly double what it was a couple of years ago, according to 2026 industry measurement reports [reported]. “We adopted it” stopped being an acceptable answer, and “here’s what it earned” became the bar. And this is what tells you measurement is genuinely scarce and not just underrated: the field hasn’t agreed on how to do it. There’s no canonical, trusted method for proving an AI system paid off [reported]. That’s an unsolved problem, and unsolved problems are where scarce advantages come from.

The honest objection is that I’m calling plumbing a moat. Measurement is operations, a support function you run after the real work. The assets are the substance, and telemetry just reports on them.

That holds right up until the input goes common, and the menu’s whole premise is that it has. The reason “we use AI” stopped being a moat is that everyone got the tool. Run the same logic one layer up. When two competitors both bring deep expertise, or both sit on proprietary data, the asset has stopped being the scarce thing between them. Whatever is still scarce is the moat. And the 73 percent says the scarce thing, across all three assets at once, is the ability to prove which of them is working.

The moat migrated once already, and everyone agreed it did. It moved off writing code, which AI made cheap, and up to knowing what to build. The claim here is that it moved again, and quietly, from knowing what to build to being able to prove what worked. The strategy decks haven’t caught up. They still file measurement under operations and put the assets on the marquee.

One thing would flip it back, and it’s worth watching for. Measurement is only the moat while it stays hard. If models keep improving, proving-what-worked could go the way the tools did and ship as a default report inside every platform. The day you can ask “which of our pilots actually worked” and get a trustworthy answer without building the instrument yourself, the ruler becomes as common as the tool, and the moat slides back to whichever asset stayed genuinely rare.

So the real bet, if you’re deciding where to put your money, is a question about time. Does proving-what-worked stay hard longer than your expertise, data, or trust stays scarce? Right now the evidence says it does, by a wide margin. Watch the tooling anyway, because the answer isn’t fixed. The moat is the ruler today. It won’t stay there on its own.