A polished summary cannot repair an ambiguous source table. Before using AI, make the input understandable to a person who did not create it. Define the reporting period, units, scope and meaning of each field.
Make the dataset interpretable
Include a small data dictionary. For example, explain whether revenue is gross or net, whether cancelled orders are included, and whether a blank means zero, unknown or not applicable. Preserve the original values and document transformations separately.
Use stable row identifiers so a reviewer can trace a generated statement back to its evidence. Avoid relying on spreadsheet row positions if rows may be sorted or filtered. Reconcile totals with the approved source system before asking for interpretation.
Check time and comparability
Compare like periods and scopes. A ten-day partial month is not directly comparable with a complete prior month. A newly included region can create apparent growth that reflects coverage changes rather than underlying performance.
For an illustrative table, suppose April contains 1,200 orders across four regions and May contains 1,500 across five. The raw increase is 25%, but that does not establish 25% growth in the original four regions. Request both the comparable-region calculation and the expanded total.
Limit what you share
Remove fields that are not needed for the task. A summary of sales by category generally does not need individual customer addresses. Use the tools and data-handling processes approved by your organization. Do not assume that a personal account has the same controls as a contracted business environment.
Sanitization should preserve enough structure for the test while removing information you are not authorized to transmit. Replacing names with arbitrary identifiers may help an experiment, but it is not a universal guarantee of anonymity when other fields can identify a person.
Specify how to handle uncertainty
Ask the system to flag missing inputs rather than estimate them silently. Require source row references for material claims and separate computed facts from possible explanations. An observed fall in conversion does not, by itself, prove that a marketing campaign caused it.
| Before submission | Reviewer question |
|---|---|
| Scope | What population and period are covered? |
| Units | Are currencies and scales consistent? |
| Missing values | Are unknowns distinguished from zero? |
| Permission | Can this information be used in this tool? |
| Traceability | Can claims be traced to stable source references? |
Keep the original data, the prepared input and the generated output together in your approved record system. That makes it possible to investigate an error and repeat the task after correcting the input.