Every organization already knows the answer to almost every question its customers ask. Almost none of them can get that answer to the customer at the moment they ask it.

A confident wrong answer is worse than no answer

A general-purpose assistant produces fluent, plausible, wrong answers with exactly the confidence of correct ones, and the customer has no way to tell them apart. It spends trust it did not earn, and the organization finds out downstream. Which means the hard engineering problem is not making an assistant answer — it is making it refuse.

Refusal cannot be achieved by asking politely

Telling an AI to 'only answer from the approved material' works most of the time. Most of the time is not a standard: it means the failure is rare, therefore unmonitored, therefore first discovered by a customer. The restraint has to be built into how the system works — it searches your approved answers first, and when nothing is found there is simply no route to an answer — rather than relying on the AI to follow an instruction.

Handover is where the goodwill is spent

An assistant that fails to a human who then opens with 'how can I help you today' has wasted the customer's time twice, and the second time is the one they remember. The transfer has to carry the whole conversation and what was already attempted.

The questions it cannot answer are the valuable output

Every unanswerable question is either a hole in your documentation or a scope decision you have not made. Collected and reported, that is a to-do list written by your customers — and organizations act on it in a way they never act on a content audit.