Choose a question worth testing
Pick a question that business users check repeatedly and for which a reference result already exists. One example is “What was the value of confirmed orders in each region last month?” In the first round, complete the full path from retrieving data to explaining and checking it. Reserve forecasting, automated decisions and writes across systems for later rounds.
Write a task card with the intended user, reporting period, comparison dimension and decision the answer will inform. If a term such as “sales” has several meanings, specify whether it refers to orders booked, invoices issued or payments collected.
Gather the minimum material you need
| Item | Ready when |
|---|---|
| Business question | The scope, reference answer and person responsible for checking it are identified. |
| Data scope | The list of tables, fields or documents has been approved. |
| Access identity | A dedicated read-only identity has a defined scope of visibility. |
| Business definitions | The period, time zone, currency, statuses and deduplication rules are documented. |
| Runtime environment | The access path, network boundary and version can be confirmed. |
Run and record one test
- Generate a baseline from the same data and record the snapshot and reporting cutoff.
- Ask the first question and check the result, unit, period and source actually used.
- Change one dimension in a follow-up, such as “Break that down by region,” and check that the original conditions still apply.
- Ask a question that requires missing or unauthorized data. Observe how the system explains the limitation.
- Save the input, output, sources and conclusion together. Mark the test as passed, partially passed or failed.
When to expand the scope
One correct answer validates that particular path. Before involving more departments, test different permissions, empty datasets, expired documents, incorrect metric definitions and failed exports. Change one factor at a time to distinguish data problems from errors in interpretation or execution.
If business data cannot yet be shared, use the site's synthetic demo to learn the workflow, then test real data in an authorized environment. Keep actual account details and business records out of the public preview.
