The last moat

Six months ago, having an AI coding agent gave you a real edge. You shipped faster. Your competitors were still copying and pasting from ChatGPT. That gap felt significant.

It's gone now. Every serious team is running agents. The tooling is commoditised. The models are available to everyone at roughly the same price. If your only advantage was "we use AI," you're not ahead — you're just at parity.

This is worth sitting with. Because a lot of teams built their strategy around that edge, and they're now in a world where it doesn't exist.

What can't be copied

When infrastructure becomes commodity, the moat moves up the stack. Happened with cloud computing. Happened with open source. Happening again now with AI.

In 2010, having your own servers was a competitive advantage. By 2015, it was a liability. The teams that won weren't the ones who figured out bare metal — they were the ones who figured out what to build once bare metal stopped mattering.

AI tooling is following the same arc. The model calls, the agent scaffolding, the RAG pipelines — that's infrastructure. Necessary but not differentiating. What you build on top of it is the only thing that counts.

The inputs that can't be replicated

There are a few things no API access can give you. They're unfashionable to say out loud because they sound soft. But they're the actual answer.

Taste. The ability to look at output — code, copy, design, a product decision — and know when it's wrong even when it's technically correct. Agents don't have taste. They optimise for the objective you gave them. If your objective is wrong, they'll execute it perfectly.

Domain depth. A founder who spent ten years in healthcare before building a healthcare AI tool has a model of the problem that no prompt can replicate. The agents are faster. The judgment about which problems are worth solving is still human.

Relationships. Distribution is still mostly a people problem. Who trusts you enough to try the thing? Who will give you honest feedback when it's wrong? Who refers you to the next customer? Agents can help you move fast. They can't build the trust that makes someone want to bet on you.

Speed is table stakes now

The thing that agents genuinely changed is the cost of execution. Doing a thing that used to take a week now takes a day. Doing a thing that used to take a day now takes an hour.

That matters — but not in the way most teams are treating it. The response to "we can move faster" shouldn't be "let's build more things." It should be "let's make sure we're building the right things, because we'll find out if we're wrong much faster."

Fast iteration on the wrong direction is expensive in a new way. You can ship four wrong products in the time it used to take to ship one. The quality of the initial judgment — is this worth doing? — just became the most valuable thing in the room.

What we're actually competing on

At Kerber AI, we think about this directly. We have ten AI agents running across two companies. They handle most of the execution — code, content, research, ops. We don't have an advantage over other teams because we have agents. Anyone can have agents.

The edge, if we have one, is in the calls those agents don't make. The features we decided weren't worth building. The markets we decided to stay out of. The moments where Alex looked at what the agents shipped and said: this is technically right but strategically wrong — ship it differently.

That judgment layer is still human. It's also the only layer that doesn't commoditise.

The mistake to avoid

The teams that get hurt next are the ones who automate execution and assume that solves it. Who point agents at a backlog and ship faster without asking whether the backlog was right in the first place.

AI can help you execute a bad strategy at unprecedented speed. The acceleration is real. So is the cost of pointing it in the wrong direction.

The last moat isn't a tool. It's knowing what to do with the tools — and being honest enough to stop when the answer is nothing.