AI-Pilled is not a binary: AI adoption is a game of autonomy

First of all, props to Ann Miura for writing this article, which I found quite interesting, and for establishing this hierarchy of “levels” that have led me to writing my own version that I’d like to put in the context of my previous article on moats.

Here’s Miura’s hierarchy, summarised:

L0 — AI as theater. Announcements ≠ adoption.
L1 — Personal productivity. “80% of employees use AI weekly!” which is probably true and also meaningless.
L2 — Team workflow. A collection of AI-enhanced silos rather than an AI-native org.
L3 — Organizational infrastructure. Agents act across systems. Non-engineers don’t just consume shared skills — they author them.
L4 — Compounding operating system. The system gets better because it learned, not because one heroic person manually improved it.
L5 — Virtually self-driving organization. Doesn’t exist yet.

I think both extremes, L0 and L5 are pretty self-explanatory and I don’t have much of an issue with how they’re described. Here’s what I find interesting about the complicated 1 to 4.

This is the “hard test”, as Miura puts it: “If your best AI user left tomorrow, would their workflow remain in the company?”

This test begs the question: in what business context would your team’s best performers leaving not negatively impact your company’s work? I can’t imagine any possible answer to this question that doesn’t imply your actual team has very little value.

AI is great at amplifying what a team can do, but if what every individual in the team is doing can be easily replaced, then your team’s skillset and talent is not a moat.

Like I wrote in my article «The Only Question that Matters after “Does It Work?”», a team that has worked together for a while and dominates experience curves, procedural knowledge, institutional memory, and deep customer insight will be (in my estimation) more valuable in the age of AI, not less. It’s questionable whether this kind of knowledge can be fully passed on to AI at all, since a lot of it is relational, but even admitting that it can, it’s highly questionable to me whether you should want to.

The gist of levels 2 to 4 is essentially twofold: visibility and delegation. Visibility, I think, is inarguably a good thing. You want whatever AI systems you implement in your company to have access to as much information as possible and be as cross-departmental as possible.

Delegation, I think, once again begs the question: if the system is “noticing”, “deciding”, “acting”, etc. and the system is entirely run by an LLM, what do you bring to the table?

The article answers this in a single line, that isn’t elaborated on and feels like a second thought: “strategy, taste, risk, values, and exceptions”.

There are, of course, other kinds of moats to be had, but these aren’t typically easy for companies to achieve. If you have government-granted protections, then none of this should concern you. If you’re building against a relatively rigid system, then being quicker and more adaptable is good, but it won’t stop your peers. Scale is not something you just chance upon, it’s compounded over time.

So my question isn’t really whether this taxonomy of AI levels is correct. My question is where competitive advantages truly lie.

Like I’ve said before, I am very weary of top-down mandates and tying compensation/promotion to AI proficiency, rather than business results, feels very strange to me.

The ultimate question is this: if all you’re bringing to the table that can’t be replicated by someone else using the exact same models and harnesses you have is “strategy”, “taste” and “exceptions”, can you build a moat? I see a better case for strategy than the others, but if everyone’s executing at roughly the same level, and in roughly the same ways, strategies will be copied pretty quickly.

This is how Miura closes the article and I like the choice of both words, though I feel they somewhat come into conflict with what came before.

I love the notion that “AI-pilled” comes in different levels. I agree with it. I absolutely concur that companies will reorganise and won’t be “AI-assisted versions of old companies”.

What I struggle to understand is how you build a competitive advantage if all you have to throw at problems is the exact same thing (I’m purposefully avoiding the word «tool») your competitors are using. “The system gets better because it learned” is a load-bearing sentence, but it fails to address what I think is a key issue: if the one doing the learning is Claude, then all your competitors are learning as well (regardless of what lab policies say). The solution seems to be to use it more. I am sceptical this can constitute any kind of moat, even without getting into the weeds of what that means in terms of deskilling your actual team.

Here’s what I wrote earlier in the article I quoted:

«You don’t have a moat. Any moat. It’s your job to pick up a shovel and dig one. But why would you? What protects you in the early days?

The answer, which you may love or hate, is uncertainty.

The very thing that makes startups terrifying is the fact that nothing is proven, that the market is unclear, that the technology might not work. This is also what keeps bigger, better-resourced competitors from bothering to enter. Established companies don’t chase uncertain markets. They wait. And in that window of waiting, smart founders build their moat.»

A lot of the advantages of levels 3 and above of being AI-pilled seem to be about reducing uncertainty and finding optimal systems that the exact same LLM your competitor is using can run. I don’t know if that’s desirable.

As a businessman, I find that optimising for non-business goals is hardly ever ideal. I am curious, however. And engaged! If business operations and margins do in fact benefit from tokenmaxxing, I’ll be very interested to learn how.