← All writing · 4 min read

Practical AI · By Yosef Gerson

When an AI should ask instead of answer.

Ask an AI assistant to plan a quiet afternoon, and it may produce a beautiful itinerary. But quiet for whom: someone who wants an empty trail, or someone who wants a comfortable chair?

The answer can be polished before the question is understood. That is why one of the most useful things an assistant can do is pause.

The missing detail that changes the answer

Some gaps barely matter. If you ask for a short explanation of fractions, the assistant can choose a sensible example and begin. Asking about your preferred font would only delay the useful part.

Other gaps determine the recommendation. “Find a good restaurant” means little until a location is known. “Help me choose a camera” may need a budget, but also a purpose: photographing a bird across a field is a different problem from photographing dinner.

The distinction is not whether information is missing. Information is always missing. It is whether a plausible answer to the missing question would change what you should do.

Try two plausible versions

Suppose you ask for a half-day walk on a trip. The assistant does not know how far you want to go. Give the missing detail two reasonable values: two miles and eight miles.

If both versions lead to the same short riverside route, the detail may not need to block a first suggestion. If one leads to a gentle loop and the other to a strenuous climb, the assistant should ask before presenting either as the right choice.

This is a useful test because it turns vague uncertainty into a concrete comparison. You are not asking whether the assistant feels confident. You are checking whether its recommendation survives a reasonable change in the facts.

A good question is small

“Tell me everything about your preferences” hands the entire job back to you. A better question narrows a real decision: “Would you rather walk for about an hour, or spend most of the afternoon outdoors?”

A good clarifying question has three qualities: it is easy to answer, it separates genuinely different options, and it explains why the answer matters.

When the cost of a wrong assumption is low, an assistant can also offer a provisional answer: “Assuming you want an easy hour-long walk, I would start here.” The assumption is visible, so you can correct it without starting over.

Give it permission not to guess

You can make this habit part of the request:

Before recommending anything, identify the missing detail most likely to change your answer. If it matters, ask me one focused question. Otherwise, state your assumption and continue.

This will not make an assistant infallible. It can still overlook the important gap or choose the wrong assumption. But it gives you a clearer place to inspect the answer.

A useful answer is not always the fastest answer. Sometimes it starts with the one question that prevents the rest from going wrong.