ChatGPT for finance: examples, prompts and controls
ChatGPT can help a finance team comment on a budget, draft document requests and analyse files with calculation tools. How reliable it is depends on the data you supply, the checks you run and the context you give it: fluent prose is no proof that a figure is right.

English version published on 2 October 2026. Sources rechecked on the same date. The corrected teaching kit was added to the French guide on 30 September 2026.
How do you comment on a budget-to-actual variance?
This fictitious example works in thousands of euros. Expenses are shown as positive amounts; the balance is a simplified difference, not accounting EBITDA.
| Indicator | Budget | Actual | Actual − budget variance |
|---|---|---|---|
| Revenue | 200 | 180 | −20 (−10%) |
| Expenses | 150 | 155 | +5 (+3.33%) |
| Simplified balance: revenue − expenses | 50 | 25 | −25 (−50%) |
Prompt to adapt:
Using this fictitious table in k€, write a commentary of no more than 120 words. Use the variances already calculated. Separate facts, hypotheses and questions to ask. Do not invent any cause or missing data. Finish with two questions: one for the sales lead, one for the expenses owner.
Expected answer: it should note that lower revenue and higher expenses both reduce the balance. A lost customer, a pay rise or seasonality must stay hypotheses until documented. A percentage against a zero budget must be flagged as not calculable.
The data model and upstream controls are explained in the automated reporting guide.
How do you prepare a more precise document request?
Give the expected document type, the period and the exact reason for rejection. Strip out unnecessary identifiers before you test. The Armanino case published by Anthropic describes a tool for drafting explanations when an audit document is wrong.
Write a courteous request of fewer than 100 words. Document received: June trial balance. Document expected: July trial balance after closing. Explain the difference and specify which document to send. Add no amount, deadline or commitment that has not been provided.
Check the recipient, document and period before you send anything. This draft must not trigger a follow-up automatically. The case details and limitations separate the tool actually deployed from the benefits the publisher estimates.
How do you prepare board questions?
Work from an approved version of the reporting, once access has been authorised. Define the decision you are preparing: a hire, a funding round or a spending trade-off.
Using only the indicators and commentary supplied, suggest five questions a board member could ask. For each one, cite the row or passage that prompts it, identify what information is missing and distinguish an observed risk from a hypothesis. Recommend no decision without the necessary data.
Set the questions against the cash-flow forecast. Positive profitability does not rule out a short-term funding need.
Can ChatGPT calculate and analyse an Excel file?
Yes: OpenAI’s data-analysis documentation describes file analysis and code execution. Distinguish generated prose from a calculation actually run by a tool. Ask for the code or formulas, the columns used and the exclusions.
Check totals, signs, currencies, empty rows and duplicates separately. Rerun the calculation in Excel, a controlled script or your BI tool. The presence of code validates neither the assumptions nor the quality of the file. Never delegate signing off the accounts, approving a payment or taking a tax decision to a model’s answer.
Which data and plan should you use?
Start with fictitious or public data. For internal data, the company must authorise the tool, its access and the purpose of the processing. Removing a name is not always enough to anonymise a file: amounts, dates and job titles can identify someone.
OpenAI documents no training by default on Business, Enterprise and API data. That does not mean no retention, nor permission to upload every document. Check permissions, retention periods, connectors and contractual terms too. Whether the DPO or security owner needs to sign off depends on your company’s context.
Record the authorised uses, excluded data, the reviewer and the error procedure in your policy. Use our LLM selection framework to compare environments on your own documents.
Try fictitious reporting with its answer key
This teaching kit contains no client data. It describes one month for a fictitious company, in thousands of euros, with positive expenses. The revenue-minus-expenses balance is not accounting EBITDA.
Download the complete kit (ZIP) or open the CSV file.
- Open the data in Excel or your spreadsheet: semicolon delimiter, UTF-8 encoding. Check period, entity, unit and unique identifiers.
- Aggregate revenue and expenses, then calculate variances. Compare with the supplied answer key.
- Use the prompt below on this fictitious file. Reject any cause absent from the data.
- Then test the file containing a duplicate: the control must identify the repeated row before validating the total.
| Indicator | Budget | Actual | Variance |
|---|---|---|---|
| Revenue | 200 | 180 | −20 (−10%) |
| Expenses | 150 | 155 | +5 (+3.33%) |
| Simplified balance | 50 | 25 | −25 (−50%) |
Analyse this fictitious reporting in k€. Check duplicate identifiers, periods, units and empty fields before any calculation. If an anomaly exists, identify it and suspend validation of totals. Otherwise, calculate revenue, expenses and the simplified balance, then their budget variances. Show the operations or code used. Write commentary of no more than 120 words, separating facts, hypotheses and questions. Invent no cause. A percentage with a zero budget is not calculable.
Reference commentary: revenue is €20k below budget and expenses exceed budget by €5k. The simplified balance therefore falls by €25k. The data does not allow attribution to price, volume or hiring: these causes must be checked with the owners.
Passing this exercise does not validate deployment on your accounts. The pilot protocol describes controls to organise with your team.
Frequently asked questions
Can ChatGPT be used for a cash-flow forecast?
It can help structure the assumptions and prepare the calculations using tools. Due dates, expected receipts and scenarios must be documented, then checked in the reference financial model.
Does a paid plan guarantee confidentiality?
Price alone proves nothing. Check the plan’s contract, settings, retention, access and integrations, and follow company policy.
Do these prompts describe Iter Advisors assignments?
No. The exercises use fictitious data. Public cases are credited to their publishers and kept separate from documented Iter engagements.
Scope your project with a CFO
Share your needs, tools and difficulties. We can clarify the scope, deliverables and controls for support.
Discuss my projectSources & references
- Data analysis with ChatGPT — OpenAI Help Center, source rechecked on 2 October 2026.
- Business data protection — OpenAI, source rechecked on 2 October 2026.
- Armanino: drafting document requests — Anthropic / Claude, source rechecked on 2 October 2026.