Building a Dashboard with AI? Make Every Number Explainable

Let a business question determine what appears first

Ask the reader what they most want to know when they open the page: which tasks need attention, for example, or which group of content deserves further observation. Select metrics and details that answer that question, and leave information that is not yet needed for later pages. The task should determine how many summary cards belong on the first screen.

For each proposed number, define the thing being counted, its unit, the time range and the deduplication rule. Identify its data source. For visits, orders or processing results, establish what each record actually represents. Give these definitions to the agent building the interface and keep them somewhere the reader can find them.

Check charts and details together when filters change

Use a set of records you can explain to check that cards, charts and detail views share the same time range and category rules. Change one filter and follow the related figures through the page. If a total disagrees with a list, return to the metric definition and data source to establish whether the problem is in the display or the scope of the calculation.

Also test situations with no records, data still loading and a failed request. Check whether the page makes its current state clear. An actual zero, missing data and a retrieval failure each need an appropriate explanation, and the update time should be easy to find. Readers can then decide whether a blank result calls for action.

Help readers trace a number back to its records

Ask a reader to select a change that deserves attention, explain what they think happened and try to find the corresponding records. If their interpretation differs from what the data means, adjust the label, explanation or relationship between views. This gives you a practical way to assess whether a chart helps people understand their work, and a clear goal for interface checks.

Check results with a fixed input

Prepare orders worth 100 and 50 yuan, then record a 20-yuan refund against the second order. Gross sales should be 150 yuan; a net-of-refunds metric should be 130 yuan and be labelled accordingly. Filtering to the second order’s date should update the chart, total and detail together. This is an example dataset proposed here, not an existing repository test result.