Right on Average

Every technology that lowers the cost of producing something changes what is scarce. AI has lowered the cost of producing software and the cost of producing ideas, at the same time. I think this creates two dangers inside companies, and what makes them hard to manage is that they point in opposite directions.

The first is dispersion.

Anyone with enough access can now build and ship a working tool in a weekend. A script that pulls customer data into a spreadsheet, a small internal app, an automation that writes into a production system. It used to require an engineer and a review. Today it requires curiosity and an account. I see it everywhere, and most of the time the intent is good: someone had a problem and solved it.

The difficulty is what happens next. The tool works, so it gets shared. Others start to depend on it. It touches data it was never meant to touch, runs on credentials nobody rotates, and breaks the day its author changes roles. Little by little, a parallel ecosystem forms beside the official one, with no owner, no map and no maintenance plan. When it fails, the cost lands on the technical team, which inherits systems it never designed and cannot easily see.

The instinctive response is to ban. I think that is a mistake. Banning does not stop people from building; it only stops them from telling you. The energy behind these tools is real. The problem is not that people build. It is that they build outside any system someone owns.

The second danger is quieter, and I find it more interesting. It is uniformity.

A language model is, at its core, a machine for predicting the most probable continuation. That is not a flaw; it is what it was built for. But it has a consequence that is easy to miss. Ask a model for a strategy, a product idea or a positioning, and it returns something close to the center of everything it has seen. A well-formed, reasonable, defensible answer. The answer most people would converge on.

For one person, this is useful. Across an industry, it is a problem. If every team drafts its plans through the same few models, fed with the same public knowledge, they will tend to arrive at the same plans. Each will be correct. Very few will be singular.

Yet competitive advantage has never come from being correct. It comes from distance to the mean: seeing something others don't, believing something others find unreasonable, and turning out to be right. The ideas that define companies usually looked odd at the time. A tool optimized for the probable is, by construction, a poor generator of the improbable.

AI makes everyone more productive and everyone more alike. The real risk is not that it is wrong. It is that it is right on average.

The two dangers look contradictory: too much divergence in how we build, too little in how we think. They come from the same place. We have adopted a powerful tool faster than we have decided where it belongs. In both cases, the answer is not less AI. It is more ownership and more judgment.

I have long argued that a company should build the software it runs on. I still believe it, and AI makes it more achievable than ever. But building only creates value inside a system someone owns: shared foundations, known data, clear responsibility for what runs in production. Without that, building becomes sprawl. AI doesn't replace engineering discipline. It makes it mandatory.

In practice, I think leadership should do four things.

First, give every system an owner. Not necessarily the person who wrote it, but someone accountable for its data, its security and its future. If nobody is willing to own a tool, it should not be in production.

Second, set guardrails instead of bans. Offer a sanctioned place to build quickly: approved environments, controlled access to data, a simple path from prototype to supported tool. Make the right way the easy way.

Third, decide explicitly which decisions are allowed to be average. Most of what a company does should be done the standard way; delegate it without guilt. Then name the few decisions where being ordinary is a loss: what you build, how you position yourself, what you refuse to do.

Fourth, protect time and places for unassisted thinking. Meetings where the first draft is written by a person. Questions argued before anyone asks a model. The first draft is where distance to the mean is created, and once the probable answer is on the table, it is hard to think past it.

The companies that do well in the next decade will not be the ones that use AI the most, or the least. They will be the ones that know where to let it make them average, and where to refuse.