AI-Soda Brain

When the question is right but the data source is wrong

When someone asks about inventory while looking at sales data, a useful answer identifies the missing information and shows where the person can go next. Reliable conversational analytics depends on handling that gap without guessing.

Trustworthy answersUpdated November 20264 min read

A simple question with a missing field

Consider a synthetic example. A user opens a sales source and asks, “How much stock of product A is available right now?” The source has orders and customer segments, but no stock balance. Product capacity in a manual, historical shipment volume and an old report all describe different facts. None establishes current available inventory.

The system can help by identifying the data in scope, naming the missing field and guiding the user to an authorized inventory source. A specific explanation gives the user a practical next step.

Distinguish three kinds of knowledge

A system may know which kind of application usually manages inventory or retrieve a document about the inventory process. Business knowledge, historical records and a fresh reading of current stock are distinct, and an answer should tell the user which it has.

A useful response says what was checked, why it cannot answer yet and what would make the question answerable. For example, “This sales view has no inventory field. With the necessary access, select an inventory source and check by product and warehouse.” This illustrates response design. It does not report a completed cross-database query.

Make the next step fit the failure

Offer an action specific to the problem. The user might choose another authorized source, revise the question, open field definitions or submit a data request. A generic “Contact your administrator” button or repeated rephrasing leaves users guessing what to do.

When continuing the query, retain the product, time and organization from the original question and recheck access. Switching sources can change the meaning of a metric. Explain differences such as available stock versus book inventory when the source changes.

Test boundaries with counterexamples

Prepare three synthetic questions. Give the first complete data, leave a key field out of the second and deny the current identity permission for the third. Check that answers distinguish no records, no field and no access, and that the suggested next action matches the reason.

  • For missing data, identify the gap and avoid an unsupported precise figure.
  • For insufficient permissions, protect the restricted information in the answer, its citations and its exports.
  • For multiple possible definitions, clarify the intended meaning before making the comparison.

Make boundaries part of everyday use

Showing the missing element, visible scope and next step in ordinary interactions helps users form accurate expectations. It also shows product teams where additional data is needed instead of attributing every unanswered question to the model.

Build this verification habit from the first question. It supports adoption beyond a single successful demonstration. Use “Ask your first question and check the answer” to assemble your own samples of success and failure.

Back to all articles

Try “permissions,” “data source,” “manufacturing” or “API.”