Imagine I’m a well-funded competitor with a technical team that is great at utilizing agentic engineering. First thing I’d want to know: where does your AI advantage actually come from?

Not the features. Those I can rebuild. I want to know what I can’t buy.

Funded teams are replicating production-grade AI systems in weeks, which used to be months. The barrier to building something great is gone. In my view, features were never a moat to begin with. Agentic engineering just made that more obvious.

The layers most companies sit in

When speaking of moats, most companies I speak to are at the Surface layer: chatbots, basic automation, things a competitor can replicate in weeks. Some have reached Operations, where AI runs inside real workflows with impact.

Fewer have built the Compounding layer, where knowledge builds up through years of human-AI interaction and gets harder to copy the longer it runs. This is also the place where the moat starts to form.

Migration at that point doesn’t mean switching tools. It means losing years of learned context. Switching at that point is more than painful, it’s operationally prohibitive. Better yet, it can’t be prompted.

Operations is where most of the useful work sits today, and it is also where people stop. The system runs, the savings are real, and nobody sets up the part that would make it compound. Moving from Operations to Compounding is mostly a decision about what you keep, rather than a bigger engineering effort.

The difference between the layers is not how advanced the technology is. A simple system that has been recording decisions and corrections for two years is further along than a sophisticated one switched on last quarter.

Three questions to find out where you stand

• Can a competitor buy your data? If it comes from a third-party feed, the answer is yes. What compounds is data your own workflows generate.

• Is your domain knowledge in the system? And no, I don’t mean it’s captured in a prompt. It needs to be built into how the AI handles your specific problems through years of trial and error.

• Can you show the P&L line? Not an “hours saved” calculation or a productivity estimate. No number, no moat.

The second one is where most people flatter themselves. A long prompt feels like encoded expertise, and it is still something a competitor could write in an afternoon once they see the output.

The classic moats still apply

The classic moats like switching costs, brand, distribution, network effects still apply. AI just raises the stakes on each one.

Most AI systems aren’t built to improve with use. The ones that are create a different kind of problem for competitors, who have to catch up on everything the system has learned as well as on the features. That’s when switching stops being a decision and starts being a reconstruction project.

Building for that from the start costs very little. You keep the corrections people make, the cases that were escalated and what happened next, the decisions that got overturned. Most teams throw all of it away, and it is the only part a competitor cannot start with.

Brand matters too. When buyers can’t see what’s under the hood, they default to names they already trust.

The window is open

Most companies are still at the Surface. Which means the Compounding layer is mostly uncontested right now. That changes fast when a funded team starts paying attention.

If you want one thing to do this quarter, make it the recording. Decide what you will keep from every interaction before you scale anything, because you cannot go back and collect last year’s corrections.

Originally posted on LinkedIn