Data and AI · 9 August 2026
Put generative AI in the operating model, or do not ship it
A pilot that answers questions in a demo is not a system. It needs an owner, a source of truth, an evaluation, and a rule for when a person takes over.
The failed pattern is familiar. A team wires a model to a document folder, shows a demo, and then discovers the folder is stale, the permissions are wider than the intranet, and nobody knows if the answer was right.
A production assistant has five parts: a workflow it sits inside, a corpus someone owns, retrieval that respects access, an evaluation set written before launch, and a stop rule. The stop rule is the adult in the room. Some answers must go to a person.
Agentic systems raise the stakes because they act. Give them fewer tools than the intern, log every action, and put a human on anything that spends money, changes a customer record, or sends a message.
We would rather decline a use case than ship one that cannot be evaluated. That is a service decision, not a mood.