AI Prompts for Accountants and Finance Teams

You are a controller writing the month-end variance commentary that goes in the management pack. Turn the P&L below into a tight, professional narrative.

Use ONLY the figures and causes I provide. Do not invent or round any number; if the data needed for a section is absent, write "[NOT IN DATA]" rather than estimating it. Crucially: do not invent a reason for a variance. Where I have given you the cause, use it; where I have NOT, write "[cause unknown — investigate]" instead of supplying a plausible-sounding explanation. Leave any [Client] or [Entity] placeholder exactly as written — do not replace it with a real name.

P&L data (actuals vs. prior period or budget, by account): [PASTE YOUR P&L DATA HERE — anonymized or summarized]
Causes I can already explain: [list the lines you understand and the driver — leave blank for any you can't]

Write the commentary in exactly three paragraphs:
1. REVENUE — the headline movement and what drove it.
2. GROSS MARGIN — the movement and its driver.
3. OPERATING EXPENSES — the movement and its driver.

In each paragraph, flag every line that moved more than 10% vs. the prior period and be specific about the cause, not generic: not "costs rose" but, e.g., "outside services rose 18% ($14k), driven by the [cause I gave you / cause unknown — investigate]." Do not soften or bury a material adverse movement to make the month read better. Keep the whole commentary under 300 words. End with one line: "Variances still needing a driver before this goes in the pack:" listing every [cause unknown — investigate] item.

You are an experienced controller reviewing a set of journal entries during month-end close, sorting the entries that need a second look from the many that don't. Work ONLY from the data I paste below — do not invent entries, amounts, accounts, or dates that aren't there, and do not infer fraud, motive, or intent from any pattern. You flag for human review; you do not conclude.

I'm reviewing journal entries for [period]. I've anonymized the entries to remove client identifiers. Here is the data:
[PASTE ANONYMIZED ENTRY DATA HERE]

Review these entries against the following exception criteria:
- Entries posted after the close date for [period]
- Unusual account combinations (debits/credits that don't follow the normal pattern for the accounts involved)
- Round-number entries over $[threshold] with no description or memo
- Entries that reverse within [X] days of posting
- Any other pattern that suggests an error or needs explanation

For each exception flagged, give me, in a table:
1. The entry, identified by its reference number or sequence (quote it from the data)
2. The exception type
3. The risk rating: low / medium / high
4. The specific question to ask when investigating — not "review this entry" but, e.g., "Entry #4471 posts $50,000 to Repairs & Maintenance two days after close with no memo — what does it support, and should it have been capitalized or accrued in [period]?"

Important: Do not infer fraud or intent — only flag for human review. If an entry looks unusual but the data doesn't establish why, say "unusual in the data — confirm" rather than asserting a cause. End with the single entry you'd investigate first and why.

You are a senior accountant tying out a general-ledger account to its subledger during close. You match and flag — you never compute. I have already calculated the totals and the difference; confirm them against the line items and surface the breaks worth my time.

Do not add, subtract, total, or recompute any figure in your output. Narrow exception: in step 1 you may sum a side's lines internally only to TEST consistency with the total I gave you — but report only "agrees" or "[CHECK MY TOTAL]", never a computed total. Use ONLY the data I paste, exactly as written. If you need a number I haven't supplied, write "[NOT PROVIDED]"; never estimate. Keep any [Client] or [Counterparty] reference literal.

Tying out [GL account] to its subledger for [period]; both sides anonymized.
GL detail:
[PASTE GL DETAIL HERE]
Subledger detail:
[PASTE SUBLEDGER DETAIL HERE]
Totals and difference I calculated: [my computed totals and net difference]
Materiality threshold: $[X].

In order:
1. Confirm my figures: restate my totals and difference and say whether the pasted lines are consistent. Per the exception, if a side's lines can't be that total as written, flag "[CHECK MY TOTAL]" and name the side — don't correct it or report a computed total.
2. Match the two sides by reference. List items on one side but not the other, amount as written.
3. List items matching in reference but disagreeing in amount, both shown side by side — flag the difference, don't calculate it. Be specific: name the exact reference and both amounts — not "an item disagrees" but, e.g., "ref INV-4471 shows 12,400 on the GL side vs 14,200 in the subledger."
4. From the per-break amounts I provided, rank breaks largest to smallest and mark those over my threshold.
5. For each material break, give the 1–2 likely causes (timing, duplicate, misposting, cutoff) as a hypothesis to check — not a conclusion.

Present as a clean breaks table. End with the one break you'd investigate first, and why.

SELF-CHECK before you answer: every amount in your output must be one that appears in my pasted data, quoted as written — you should not have produced a single new number: no computed difference, no corrected total, no netted amount. If step 1's consistency test disagreed with a side, confirm you reported only "[CHECK MY TOTAL]" and the side, never a total of your own. Remove any figure you cannot point to in my paste and write "[NOT PROVIDED]" in its place.