NewAI Adoption Index

Writing / Strategy, Adoption to value

You can't train your way to infrastructure

Training makes individuals better. A company that runs on AI needs something built. Here is what has to be built, and which parts you should never have to think about.

"Teach our people and they will figure it out" is the most common AI plan there is. It fails quietly, and not because your people are not capable.

A category error

Training makes individuals better. That produces exactly the private, non-compounding capability that got you stuck.

A shared record of how the business works. One settled source of truth. Work that survives the person who set it up. Reliable, repeatable processes. These are things you build. They are not skills you instil.

Asking individuals to produce them in their spare time is asking a hundred people improvising in parallel to arrive at architecture. Architecture comes from someone designing the system on purpose.

The questions are different

Building it means answering questions no individual user ever faces:

  • What is the source of truth for each kind of answer, and who owns it?
  • How do people and agents share one body of knowledge safely?
  • What rules does everyone, and everything, operate under?
  • How does capability transfer when people move on?
  • How reliable does this need to be before you would bet the business on it?

These are design questions. They get answered before the system can compound at all.

Two layers, and only one is your problem

AI is becoming the layer the whole company runs on, and it splits in two.

The model layer is the raw intelligence. It comes from Anthropic, OpenAI and Google, and it improves every month whether you act or not.

The operating layer is the knowledge, orchestration, governance and measurement that turn that intelligence into your company's capability. No model vendor ships you that. It is the part you either own or rent.

Keeping the two separate has a practical benefit: you are not locked in. When a better or cheaper model appears, you switch, and everything the company has learned stays where it is.

What you should not have to think about

A well-built operating layer takes a set of worries off the table:

  • Security. Who can see and do what, enforced everywhere, for people and agents.
  • Memory. What the company knows is kept, current and findable.
  • Continuous evaluation. The system is tested constantly, so you learn about a drop in quality from a report and not from a customer.
  • Cost. Routine steps can run on smaller models tuned to your work, which cost less and make fewer mistakes on the tasks they are tuned for.

Treat it like the systems you already take seriously

You engineered your financial systems, your data platform and your security deliberately, with expertise, because the business depends on them. The layer everything else will soon run on deserves the same.

Someone has to own building the system, not just training the people.

Next in this series: Don't shut down what your people built. Connect it.

Put AI to work across your company.