The most valuable AI memory may not be the documentation.
It may be the exceptions.
Enterprise AI value often lives in override history, decision traces, edge cases, and informal judgment — not in the clean policy documents everyone already has.
I have seen organizations treat documentation as the source of truth while the real system lived somewhere else. The written policy said one thing. The actual decision depended on a customer promise, a trusted operator, an ugly incident, or an exception nobody wanted to formalize.
That is why the AI value-capture question is sharper than "data is a moat." A claims assistant does not only learn the rulebook. It can learn when the company bends the rule, which exceptions become precedent, and whose judgment the organization actually trusts. That memory is valuable precisely because it is sensitive — which is the same reason it is dangerous. The decision residue that would make the AI smart is also the record that leaks how the business really works, encodes its biases, and exposes every exception it never wanted audited. It needs ownership, retention rules, access boundaries, and training controls before it becomes a feature.
If decision traces become the AI moat, they also become the AI liability. The same data is both.
The question is not only what the model remembers. It is who owns the memory of how the organization really decides.