When we build an AI assistant into a business, the first decision isn't the model or the prompt. It's the list of things the assistant is never allowed to generate on its own: prices, stock, order status, balances, dates. Anything a customer might act on.
Those come from tools that read the real system every time, and the assistant's job is to call the tool and report what it says. If the tool fails, the assistant says it doesn't know. It doesn't guess.
It sounds limiting. In practice it's what makes the assistant safe to put in front of customers, and it puts the engineering work where it belongs: clean data and clear tools, not a longer prompt asking the model to be careful.