Budget variance commentary checklist: check an AI draft

Use this check when an AI draft sounds convincing but you need to establish whether each claim is supported. You need the draft, the approved reporting table and the driver notes. Work sentence by sentence; asking the same assistant whether its own answer is correct is not an independent check.

Separate numbers, explanations and judgments

Consider this deliberately flawed sentence: “Software costs rose 12% because of higher supplier prices, while travel savings offset the increase.” The underlying fictional table shows a software budget of 100 and actual of 112, and travel budget of 40 and actual of 30, all in USD thousands. The driver note attributes 8 of the software variance to additional seats; 4 is unexplained. The travel variance comes from a deferred trip.

Use a repeatable review record

Sentence / claim:
Claim type: number | comparison | cause | forecast | judgment
Source row and evidence ID:
Period, unit, currency and reporting perimeter:
Independent calculation or source passage:
Decision: accept | rewrite | request evidence
Reviewer and date:

Check percentage points separately from percentage change. Check whether a positive variance means good or bad for the measure in question. Do not sum row percentages to create a total. If the budget is zero or the reporting perimeter changed, flag the comparison for a finance reviewer rather than forcing a standard sentence.

Completion condition

Every material claim has a source and a reviewer decision. An unresolved explanation stays unresolved in the final draft. The person authorized to approve the reporting still signs off. This task requires no new software; a spreadsheet and a marked-up document are enough. Use it inside the financial-reporting workflow or independently on a human-written draft.