You already know what this looks like.
Someone builds the thing the whole team relies on, then takes a new job. Three months later nobody can reproduce it. Two teams solve the same problem in two different ways, neither aware of the other. You ask a question and get one answer from one person and a different answer from another, with no way to tell which is right. A new hire spends months climbing back to where their predecessor stood.
These are not bad luck. They are what bottom-up adoption produces, and they run along five fault lines.
The five fault lines
- Knowledge fragments. Every person runs on a private understanding of the business. The company has none of its own.
- Truth fragments. There is no arbiter, so the same question gets different answers.
- Tools fragment. Different apps, different accounts, nothing in common underneath.
- Ways of working fragment. Everyone has their own method, so nothing hands off cleanly.
- Governance fragments, or never exists. Each person decides their own rules.
Put together, that is not a company running on AI. It is a hundred private systems wearing one logo. Each works by itself. None connect.
It never shows up on a dashboard
This cost is real, and it is invisible in reporting. It surfaces as friction, lost knowledge, rework and inconsistency. It is a tax you are already paying and have never named.
You can teach a hundred people to use AI brilliantly and still have a company that knows nothing.
Agents multiply it
Here is what raises the stakes. AI agents are starting to do real work, not just assist with it. That means every part of how your company operates now has to serve a new hire and a new agent.
In a company with a shared foundation, an agent plugs into the same knowledge, the same ways of working and the same rules as your people. One set of methods and one version of the truth, read and extended by people and software alike.
In a company without one, agents do not fix the fragmentation. They multiply it. Every island gets its own private agent, acting on its own private context, under rules nobody else agreed to.
The permission problem
One example makes this concrete. A system that knows everything about the company must also know who is allowed to know what. It should not tell you a colleague is being let go on Wednesday if you are not cleared to know.
That cannot be bolted on afterwards. Permissions have to be part of the foundation, for people and for agents. It is one of the reasons this is an engineering job and not a matter of everyone learning to prompt better.
Next in this series: You can't train your way to infrastructure.