When an AI project disappoints, the usual verdict is that the AI wasn't good enough. In most of the cases I've seen, including inside our own company, the model was fine. The setup wasn't.
It gets connected to everything at once. Nobody agrees which questions it should answer. It can read records some people shouldn't see. There is no rule for what happens when it doesn't know, so it guesses. After a few wrong answers the team stops trusting it and goes back to the spreadsheet.
We run Aion on AI every day, and we've made most of these mistakes ourselves. What fixed them was dull and effective: guardrails first, freedom later.
In practice that means starting with a small set of agreed questions, answered from records the company already keeps, with the source shown every time. Each person sees only what their role allows. When the records don't hold the answer, the assistant says so and hands the question to a person.
Once that works and people rely on it, we add more sources, more questions and, one at a time, actions. Nothing gets more freedom until it has earned it.
AI can do a great deal on its own. That doesn't mean it should. The companies that get value from it decide where their people stay in charge before they switch anything on.