Separate numbers from narrative

A weekly report is a good candidate for automation because its structure repeats. The dangerous shortcut is asking a language model to compute totals from a loose pile of records. Define each metric first: what counts as a customer, a sale, an open ticket or a pageview? Record the time zone and period boundaries. Calculate the figures with a spreadsheet formula, database query or reporting tool you can inspect. Save the previous week's result for comparison.

Before any AI summary is written, validate the dataset. Check row counts, missing dates, duplicate transactions and impossible values. Compare the total with an independent control figure when possible. If the input is incomplete, the report should say so prominently instead of producing a smooth narrative. A model should receive the approved numbers and concise definitions, not be left to guess both.

Ask for analysis with boundaries

Give the assistant a table of verified metrics and ask it to describe changes, anomalies and questions worth investigating. A useful instruction is: “Do not calculate new figures or infer a cause from correlation. Mention the denominator for percentages. If a metric is missing, state that it is unavailable.” Review every number in the prose against the table. An increase in clicks may reflect an additional campaign, a tracking change or more visitors; the report alone cannot identify the cause.

Use the same headings each week: key changes, likely explanations to test, unresolved data issues and actions. Make a person approve the final report before distribution. For sensitive business data, use an approved AI service and minimize the fields sent to it.

Learn from the exceptions

Keep a log of corrections. If a model repeatedly calls a normal seasonal change a “surge,” adjust the prompt and add context. If a metric often arrives late, move the deadline or mark it provisional. The benefit of automation is dependable preparation and clearer attention to exceptions. It is not a license to send unreviewed statistics to a team every Monday.