AI for Finance: ChatGPT Prompts for Finance Teams

Finance teams use AI to draft and explain: commentary on budget-versus-actual results, spreadsheet formulas, short summaries of long reports, process checklists, lists of mismatches between two records, and plain-language emails about the numbers. You paste the numbers and the reasons, the AI writes, and you check every figure against the source before anything leaves your desk. AI doesn't replace your spreadsheet, your controls or qualified advice on tax, investments or accounting rules.

• Let the spreadsheet calculate and the AI write. Use AI for words and formulas, not as your calculator.

• Ask it to show its math. Then you can check every figure in a minute.

• Give it the reasons. Only budget owners know why a number moved. The AI can only guess.

• Check numbers and meaning. A summary can keep every number and still change what one means.

• Results vary by tool. ChatGPT, Claude, Gemini and Copilot word things differently, but these prompts work in all of them.

A chat assistant predicts text. It often gets arithmetic right, as it did in every run below, but it isn't a calculator, and it doesn't know your ledger, your drivers or your policies unless you paste them in.

A few terms used below. A variance is the difference between the actual number and the budget. A reconciliation checks that two records of the same money agree, such as your bank statement and your books. For the general prompt method, see the pillar guide AI prompts for work.

Each task has a prompt you can copy, the model's real answer (trimmed with "…") and what to check. The answers come from gpt-5-mini, a small OpenAI model. All companies and numbers are made up. We checked every figure in the outputs by hand.

1. Draft variance commentary

The usual prompt returns about 900 words of template:

Write variance commentary for our Q3 operating expenses.

I don’t have your Q3 numbers, so below are two things to choose from:
…
- Personnel: $[Actual] vs $[Budget], $[Variance] ([Variance %]) [favorable/unfavorable].
  Explanation: [e.g., headcount growth delayed by X roles not filled; severance and one-time hiring costs; bonus accruals above plan]. …

Sample reasons like "severance" are exactly what ends up pasted into a real pack by mistake. Give it the data and the drivers:

Task: Draft variance commentary on Q3 operating expenses for the leadership pack.

Data (USD):
"""
Line,Q3 budget,Q3 actual
Salaries,1240000,1228000
Contractors,90000,156000
Software,145000,212000
Travel,60000,31000
Marketing programs,180000,149000
Rent,75000,75000
"""

Drivers (from the budget owners):
- Contractors: two data engineers hired on contract during the hiring freeze. Expected to continue in Q4.
- Software: the annual data-platform renewal was paid in Q3 instead of Q4. Timing only.
- Travel: the sales offsite moved from September to October.
- Marketing programs: a September trade show was cancelled. The money will not be spent.

Audience: The leadership team. They read this in two minutes.

Format: One sentence on the total variance. Then, for each line where the variance is more than 10% or more than $10,000: Line | Variance ($ and %) | Driver | Timing or permanent. Then one sentence on what to expect in Q4.

Rules: Show how you calculated each variance (actual minus budget, and divided by budget). Use only the drivers above. If a driver is missing, write "Driver not stated". Don't add explanations of your own.

Total variance: Q3 actuals were $1,851,000 vs budget $1,790,000, a +$61,000 variance (+$61,000 = $1,851,000 - $1,790,000; +3.41% = $61,000 / $1,790,000).

Lines with material variances (>10% or >$10,000):
- Salaries | -$12,000 (-0.97%) | ($-12,000 = $1,228,000 - $1,240,000; -0.97% = -$12,000 / $1,240,000) | Driver not stated | Timing or permanent: Driver not stated
- Contractors | +$66,000 (+73.33%) | … | Two data engineers hired on contract during the hiring freeze. Expected to continue in Q4. | Permanent
- Software | +$67,000 (+46.21%) | … | The annual data-platform renewal was paid in Q3 instead of Q4. | Timing
- Travel | -$29,000 (-48.33%) | … | The sales offsite moved from September to October. | Timing
- Marketing programs | -$31,000 (-17.22%) | … | A September trade show was cancelled; the money will not be spent. | Permanent

Q4 expectation (one sentence): Expect continued contractor expense in Q4 (upward pressure), lower software spend in Q4 due to the timing of the renewal (offsetting Q3), …

What to check: Every figure is right. Still recalculate them in your spreadsheet, because the next run may not be. "Driver not stated" for salaries is the rule working: ask the owner. Now look at the labels. The owners said contractors are "expected to continue in Q4", and the AI turned that into "Permanent". They said nothing about travel being timing. Labels like these change how leaders read the numbers, so confirm them with the budget owners.

2. Get help with a spreadsheet formula

Task: Write Excel formulas for a monthly summary sheet.

Setup:
- Sheet "Data" has transactions in rows 2 to 5000. Column A: Date. Column B: Department. Column C: Account. Column D: Amount.
- Sheet "Summary": cell B1 holds the first day of a month, as a real date (for example 1-Sep-2026). Departments are listed in A3:A12. Budgets for that month are in column C (C3:C12).

I need:
1. In B3 (to copy down to B12): the total Amount for the department in A3, for the month in B1.
2. In D3: the variance in dollars (actual minus budget).
3. In E3: the variance as a percent of budget, showing blank if the budget is 0 or empty.

Format: Each formula on its own line, then a one-line explanation of each part. Then two ways I can test the formula myself.

Rules: Excel 365. No macros. No helper columns.

=SUMIFS(Data!$D$2:$D$5000, Data!$B$2:$B$5000, $A3, Data!$A$2:$A$5000, ">=" & $B$1, Data!$A$2:$A$5000, "<" & EDATE($B$1,1))

Explanation (parts): SUMIFS sums Amounts in Data!$D$2:$D$5000 where Department in Data!$B$2:$B$5000 equals the department in $A3, the Date in Data!$A$2:$A$5000 is >= the first day in $B$1, and the Date is < the first day of the next month (EDATE($B$1,1)) — this limits the sum to that calendar month.

= B3 - C3
…
=IF($C3=0, "", (B3 - C3) / C3)
…
Two ways to test the formulas yourself:
1) Add a test transaction row on the Data sheet with Department equal to A3, Date set to a day inside the month in Summary!B1, and Amount = e.g. 100. Confirm that B3 increases by 100 …
2) Add another transaction with the same Department but a Date outside the month … and confirm it does NOT affect B3; also try setting C3 to 0 to confirm E3 becomes blank.

What to check: These formulas are correct. SUMIFS adds up the rows that meet all the conditions, and "on or after the 1st, before the 1st of next month" catches every date in the month. An empty budget cell counts as 0, so E3 shows blank as asked. Format E3 as a percent. Run both tests it suggests. Note that you only described the layout. You never pasted real data, and formula help never needs it.

3. Summarize a management report

Task: Summarize the monthly management report below for the CEO.

Format: 5 bullets, most important first, each under 25 words. Then one line that starts with "Watch:" for the biggest risk the report mentions.

Rules: Copy every number exactly as it appears in the report. Don't calculate new numbers. Don't add opinions or advice. If something in the report is unclear, add a final line that starts with "Unclear:".

Report:
"""
September management report (unaudited)

Revenue was $4.82M, up 3% on August and 2% below budget ($4.92M). Subscription revenue grew to $4.31M; services revenue fell to $0.51M because two implementation projects slipped into October.

Gross margin was 71.4%, down from 72.9% in August, mainly because of higher cloud hosting costs after the September traffic spike.

Operating expenses were $3.65M, 4% under budget, mostly from lower travel and a cancelled trade show.

Cash at month end was $18.2M. Net burn was $0.41M, compared with $0.63M in August.

Days sales outstanding rose to 52 days from 44. Three enterprise customers paid late; one of them, with $380K overdue, has asked for a payment plan.

Headcount ended at 212, against a plan of 220. Six open roles are in engineering.
"""

- Revenue $4.82M, up 3% on August and 2% below budget ($4.92M); subscription $4.31M, services $0.51M slipped to October.
- Cash at month end was $18.2M; Net burn was $0.41M, compared with $0.63M in August.
- Gross margin 71.4%, down from 72.9% in August, driven by higher cloud hosting after a September traffic spike.
- Days sales outstanding rose to 52 days from 44; three enterprise customers paid late, one with $380K overdue requesting a payment plan.
- Headcount ended at 212 versus plan of 220; six open roles are in engineering.

Watch: $380K overdue enterprise customer requesting a payment plan.

What to check: Every number was copied correctly, and one bullet is still wrong. "Services $0.51M slipped to October" says the $0.51M moved. The report says services revenue was $0.51M because two projects slipped. Compressing text can change its meaning, so check what each bullet says, not just its numbers. The summary also dropped "unaudited". Keep labels like that when you forward numbers. For more on summaries that keep the facts, see Chain of Density summaries.

4. Draft a reconciliation checklist

Task: Write a month-end bank reconciliation checklist for our team.

Context:
- Two people: an accountant prepares, the controller reviews.
- We use QuickBooks Online. We have 3 bank accounts and 1 company credit card.
- The books should be closed by business day 5.
- Our rule: any reconciling item older than 60 days goes to the controller with a note.

Format: A checklist grouped by business day (Day 1 to Day 5). Each item: what to do | who | what "done" looks like.

Rules: This is our internal process. Don't cite accounting standards, laws or audit rules. Don't add tools we don't use.

Day 1 — Prepare bank and credit card reconciliations
- Pull bank and credit card statements (all 3 bank accounts + 1 company credit card) | Accountant | Statements downloaded and saved to the month folder; statement ending dates noted in the reconciliation worksheet.
…
Day 2 — Complete reconciliations and resolve routine reconciling items
…
- Investigate reconciling items ≤ 60 days: contact vendors/branches/employees for missing info, locate receipts, ask for copies of cleared checks if needed | Accountant | …
…
Day 3 — Escalate old items, prepare reconciliations package for review
- Flag reconciling items older than 60 days and prepare a summary note for the controller (amount, age, explanation, actions taken, recommended resolution) | Accountant | 60+ day items compiled in a single document and emailed/placed in the review folder with clear notes.
…
Day 4 — Controller review and feedback
- Controller reviews each account reconciliation, supporting documents, Outstanding Items list, and 60+ day notes | Controller | …

What to check: The structure is useful, and your 60-day rule is in the right place. But the full answer has 23 items, more than a two-person team will follow, so cut it to what you actually do. "Branches" came from nowhere. The AI drafts the list. Your controller decides what the process must include.

5. Find mismatches between two lists

Task: Compare the bank statement lines with our ledger entries for the same account and month. Find what doesn't match.

Bank statement:
"""
Date,Description,Amount
Sep 2,Payment from Acme Corp,12500.00
Sep 5,ADP payroll,-48210.55
Sep 8,AWS,-3125.40
Sep 12,Payment from Lumen LLC,1520.00
Sep 15,Office lease,-9800.00
Sep 20,Payment from Harbor Inc,7340.00
Sep 25,Zendesk,-1180.00
Sep 30,Bank service fee,-35.00
"""

Ledger:
"""
Date,Description,Amount
Sep 2,Acme Corp invoice 1041,12500.00
Sep 5,Payroll Sep 1-15,-48210.55
Sep 8,AWS August,-3125.40
Sep 12,Lumen LLC invoice 1038,1250.00
Sep 15,Rent September,-9800.00
Sep 20,Harbor Inc invoice 1044,7340.00
Sep 25,Zendesk annual,-1180.00
Sep 25,Zendesk annual,-1180.00
Sep 29,Check 2207 Office supplies,-412.80
"""

Format: Three tables: (1) Matched, (2) In the bank but not in the ledger, or amounts differ, (3) In the ledger but not in the bank. For each unmatched line, add one short note on what a person should check.

Rules: Match on amount and date (within 3 days). Don't decide which side is right. Show the amounts exactly.

…
In the bank but not in the ledger, or amounts differ
Date | Bank description | Bank amount | Note
Sep 12 | Payment from Lumen LLC | 1520.00 | Amount differs from ledger (ledger shows 1250.00). Check payment remittance, invoice number, or possible short/overpayment.
Sep 30 | Bank service fee | -35.00 | No matching ledger entry. Check whether bank fees are recorded in the ledger or need to be entered.

In the ledger but not in the bank
Date | Ledger description | Ledger amount | Note
Sep 25 | Zendesk annual (duplicate) | -1180.00 | Duplicate ledger line with no second matching bank item. Check for duplicate recording or a split payment not on bank statement.
Sep 29 | Check 2207 Office supplies | -412.80 | No corresponding bank debit. Check whether the check cleared in a different month, was voided, or is missing from the bank statement.

What to check: It found all four differences, including the swapped digits (1,520 against 1,250), and it didn't decide which side is right. That stays with a person. This works for a short list or a handful of leftover items. For a full month, use your accounting software's matching and the AI to explain what's left. Bank data is confidential, so use only an approved tool.

6. Explain the numbers to a colleague outside finance

Task: Write a short email that explains her Q2 budget to the head of customer support. She isn't from finance.

Facts:
- Q2 budget for customer support: $150,000. Q2 actual: $177,000. That is $27,000 (18%) over.
- The reason: the annual Zendesk renewal ($42,000) was planned for July (Q3), but the vendor moved the renewal date, so it was paid in June (Q2).
- Without the Zendesk payment, Q2 spending was $135,000, which is $15,000 under budget.
- Because Zendesk is already paid, her Q3 report will show about $42,000 less spending than budget. That is not extra money to spend.
- Her full-year budget hasn't changed.

Format: Under 150 words. No finance jargon. If you must use a term like "timing difference", explain it in plain words.

Rules: Use only these facts and numbers. Don't add advice about what she should do.

Subject: Q2 customer support spend — explanation

Hi [Name],

Quick note on Q2 numbers: budget was $150,000 and actual spend was $177,000 — $27,000 (18%) over budget. The reason is the annual Zendesk renewal ($42,000) was planned for July (Q3) but the vendor moved the renewal date, so it was paid in June (Q2).

If you exclude the Zendesk payment, Q2 spending was $135,000 — $15,000 under budget. Because Zendesk is already paid, Q3 will show about $42,000 less spending than budget; that lower Q3 total is due to timing, not extra funds. Our full-year budget remains unchanged.

Thanks,
[Her Name]

What to check: The numbers are right and the message is clear. Three small fixes: "Our full-year budget" should be "your", the sign-off should be your name, not "[Her Name]", and "due to timing" is still finance language. Say "because we paid it earlier than planned". Plain-language emails like this prevent a lot of worried meetings. They're one of the best uses of AI in finance.

7. Flag expense lines for review

Task: Review these expense lines and flag any that a person should look at before approval.

Flag a line if:
- it looks like a possible duplicate (same employee, vendor and amount within 7 days), or
- the amount is $500 or more and a round number (ends in 00), or
- the description is missing.

Expense lines:
"""
Row,Date,Employee,Vendor,Amount,Description
1,Sep 3,E104,Delta Air Lines,412.60,Flight to Chicago client visit
2,Sep 3,E104,Hilton Chicago,289.00,Hotel 1 night
3,Sep 4,E221,Uber,38.75,Taxi to airport
4,Sep 6,E104,Delta Air Lines,412.60,Flight to Chicago client visit
5,Sep 9,E317,Best Buy,1200.00,Monitors
6,Sep 10,E221,Panera,46.20,Team lunch
7,Sep 12,E405,Amazon,89.99,
8,Sep 15,E317,Staples,64.30,Printer paper
9,Sep 18,E512,Marriott,600.00,Hotel 2 nights
10,Sep 19,E221,Uber,38.75,Taxi to airport
11,Sep 22,E104,Hertz,233.18,Car rental
12,Sep 30,E512,Restaurant Nora,515.40,Client dinner
"""

Format: A table: Row | Rule | Why | What to ask the employee. Then a count of flagged rows.

Rules: Flag for review only. Don't say anything is wrong or fraudulent. Use only the rules above.

Row | Rule | Why | What to ask the employee
--- | --- | --- | ---
1 | Possible duplicate | Same employee (E104), vendor (Delta Air Lines) and amount (412.60) as Row 4 within 7 days (Sep 3 vs Sep 6). | …
4 | Possible duplicate | Same employee (E104), vendor (Delta Air Lines) and amount (412.60) as Row 1 within 7 days (Sep 6 vs Sep 3). | …
5 | Round high amount (≥ $500, ends in 00) | Amount is $1,200.00 — a large round number that should be verified. | …
7 | Missing description | Description field is empty for the Amazon charge of $89.99. | …
9 | Round high amount (≥ $500, ends in 00) | Amount is $600.00 — a large round number that should be verified. | …

Total flagged rows: 5

What to check: Correct, including what it left out: the two Uber rides (rows 3 and 10) are 15 days apart, so the duplicate rule doesn't apply. Written rules make the result easy to check with a spreadsheet filter. A flag is a question, not a finding, and the person approving decides. The lines use employee IDs instead of names, which is a good habit.

Add these lines to the end of a finance prompt, then run the checks below before you use the answer.

Rules:
- Use only the numbers and facts above. Don't estimate or fill in missing numbers.
- Copy numbers exactly as written, with their units and periods.
- Show how you calculated every new number (the formula and the inputs).
- If a reason or driver isn't given, write "not stated".
- Label anything you infer as "assumption".
- Don't give tax, investment or accounting advice.

My checks before I use the answer:
[ ] Totals tie to the source report or system
[ ] Every calculated number is recalculated in the spreadsheet
[ ] Signs (+/-), units ($K vs $M) and periods (month, quarter, YTD) are right
[ ] Labels such as "unaudited", "forecast" or "preliminary" are kept
[ ] Every reason matches what the budget owner said

• Follow your company's data rules. Many companies treat unpublished results, forecasts and bank data as confidential or restricted. Use only a tool your company has approved for that data, with your work account.

• Don't paste: bank or card numbers, payroll by person, customer payment details, unpublished results, deal or acquisition information, or tax documents, unless your approved tool and your company's rules allow it.

• Paste less. Totals, rounded figures and IDs instead of names are often enough. Formula help needs only the column layout.

• Check every number against the source before it goes into a report, an email or a pack. A person signs off, as with any other draft.

• Take tax, investment and accounting-rule questions to qualified people. AI answers on these sound authoritative and can be wrong or out of date. Your controller, auditor or advisor is the right source. Some uses of AI are regulated in some places. Ask your legal and compliance team what applies to you before you roll this out.

The guide to using ChatGPT safely at work covers what each vendor says it keeps. For why models produce confident wrong answers, see hallucination basics.

When the same prompts run every month, teams usually build a shared prompt library, then switch on the AI features in their spreadsheet and accounting tools. Some want more: software that prepares bank matches for review, drafts the monthly commentary from the ledger, or chases missing receipts. Software that takes steps in your systems is an AI agent. In finance it must never post entries, approve spend or move money without a person's approval. That approval step is called human-in-the-loop. Start with what is an AI agent? and the overview of AI agents for business leaders.

Can ChatGPT do financial analysis?

ChatGPT can help you analyze numbers you give it: comparing periods, listing the biggest changes, drafting commentary and writing formulas. It doesn't know why your numbers moved, and its math must be checked. Use the spreadsheet for calculations, and people for the explanation and the sign-off.

Is ChatGPT good at math?

It is often right on small calculations, and every figure in this guide's examples was correct. It is still a text predictor, not a calculator, and errors grow with the size of the data. Ask it to show its math, and recalculate anything that matters in a spreadsheet.

What are the best ChatGPT prompts for finance?

The best finance prompts paste the numbers and the reasons, name the reader, set the format, and add rules: use only these numbers, show your math, write "not stated" when a reason is missing. Good starting tasks are variance commentary, formulas, report summaries, checklists, list comparisons, plain-language explanations and expense flags.

Can I paste financial data into ChatGPT?

Only into a tool your company has approved for that kind of data, with your work account. Unpublished results, bank details and payroll are usually confidential. For many tasks you can use totals, rounded figures or just the column layout instead.

Can AI give tax or investment advice?

Don't rely on it for that. AI answers about tax, investments and accounting rules can be wrong or out of date and still sound certain. This guide doesn't cover those topics. Ask your controller, auditor or a qualified advisor.

Can AI write Excel formulas?

Yes, and it's one of the safest finance uses, because you only describe the layout and you can test the result. Say which tool and version you use, describe the columns, and ask for two ways to test the formula. Then test it on rows where you know the answer.

• Learn the method behind these prompts: AI prompts for work

• Check numbers and facts before you send: how to check AI answers

• Know what you can paste: is it safe to use ChatGPT at work?

• Write clearer instructions: clear asks, roles, constraints and format

• The other role playbooks: AI for HR, AI for marketing and AI for sales