Imagine readings for three days: 4, 6 and a value that was not recorded. The average of the two observations is 5. Replace the missing reading with zero and average three numbers, and the result becomes about 3.33. The arithmetic is simple; the assumption changes the question being answered.
Tell the assistant whether an empty cell means not measured, not applicable, still pending or a genuine zero. If several meanings occur in one table, record them in a separate status column. A request to “clean this table” is too vague to settle that choice.
Before using the method on the whole dataset, verify a small example yourself. Ask for the number of observed values alongside the average. In this example, the average of 5 describes two recorded days; it does not establish what happened on all three days.
Keep the original table. Make replacements or exclusions in a working copy, document the rule and report how many rows it affects. If a large share of readings is unavailable, say so in the summary. Do not let the AI fill gaps with plausible numbers unless there is an explicit, justified estimation method and the estimates remain distinguishable from observations.