A numerical summary generated by AI may appear neat and contain incorrect counts. Before using it, it is necessary to verify the relationship with the starting data and with the question formulated. Accuracy requires relevant controls, especially when the result supports a decision or is communicated to others.
I propose to keep data and calculation criteria available. A figure should be able to be reconstructed through a verifiable procedure. Information must be treated with appropriate tools and permissions, avoiding sending confidential content only to make the summary faster.
In a hypothetical example, a table contains undone rows that do not have to enter the total. The result can be wrong if the rule is not applied. A comparison with a controlled calculation allows to verify both the number and the perimeter used.
NIST proposes a voluntary framework for AI risk management. On the operational level, I would distinguish reading support from the calculation used as a reference. The summary becomes useful when the relevant results are verified and the reader can understand what data and criteria support them.