When the correct answer is a question

You ask for the balance for Andersen. Three customers are called Andersen. A helpful assistant picks the largest one and answers with confidence.

That is the worst possible behaviour, and it is the default behaviour of a system built to be agreeable. This page is about the alternative: an assistant that asks one short question and then answers correctly.

When must an assistant ask instead of answer?

Three cases, and they cover almost everything.

The entity is not settled. More than one customer, vendor, item or project matches the name in your question. Choosing by size, by recency or by best text match is a guess wearing a confident sentence.

The period is not settled. "Last quarter" can mean the last calendar quarter, the last fiscal quarter, or the last three months. Companies differ, and the same company differs between the finance team and the sales team.

The measure is not settled. "Revenue", "overdue" and "active project" all have more than one reading until someone settles them. Aging by posting date, document date or due date gives three correct and different answers, and Business Central supports all three.

In each case, a guess produces a confident wrong answer, which is the most expensive output an assistant can produce.

Why is a guess worse than a question?

Because of what happens next.

A question costs the reader four seconds and one click. A wrong answer costs nothing at the moment it appears, and then it costs a decision. Between those two moments, nobody knows there is a problem, because a confident wrong answer and a confident right answer look the same.

A wrong answer also damages the next fifty answers. Once a finance team finds one invented reading, they check everything, and an assistant that has to be checked saves nobody any time.

How should the assistant ask?

The form matters as much as the decision to ask.

Ask one question, not three. A list of clarifications reads as a form, and a form is worse than a search.

Offer the answers. Two to six options that the reader selects with one click. Free prose asks the reader to type again, and it hides the shape of the ambiguity.

Show what distinguishes the options. Three customers named Andersen are not usefully listed as three identical names. List the number, the city or the balance, so the choice is possible.

Stop the turn. The assistant should not ask a question and then continue to work on a guess in parallel. The question ends the turn and waits.

An example

You: What is the balance for Andersen?

Assistant: Three customers match Andersen. Which one?

• Andersen A/S (10000), Aarhus
• Andersen & Sønner ApS (10420), Odense
• P. Andersen Holding (11930), Horsholm

Three lines, one click, and the answer after it is correct.

Compare that with the alternative, which is a correct-looking balance for a company you were not asking about, delivered in one turn instead of two.

Does this make the assistant slower?

It makes about one question in ten take a second turn. It removes a class of error completely.

There is also a compounding effect in the other direction. When a reading gets settled in a conversation, a good system writes it down: overdue means this, revenue means that, this company reports in DKK. The next time the same term appears, there is nothing to ask. The number of clarifying questions falls over the first weeks and then stays low.

An assistant that asks on day one and never asks again on day ninety has learned your business. An assistant that never asks has been guessing since day one, and you have not noticed yet.

What to test in a demonstration

Ask a question with a name that is not unique in your data. Ask a question with "last quarter" in it. Ask for a measure your company defines in its own way.

Watch whether the assistant asks or answers. Then ask it how it decided. The second answer is more informative than the first.