ChatGPT for finance: examples, prompts and controls

This fictitious example uses thousands of euros. Expenses are positive amounts; the balance is a simplified difference, not accounting EBITDA.

Prompt to adapt:

Using this fictitious table in k€, write commentary of no more than 120 words. Use the calculated variances. Separate facts, hypotheses and questions to ask. Do not invent any cause or missing data. End with two questions for the sales lead and the expenses owner.

Expected answer: note that lower revenue and higher expenses reduce the balance. A lost customer, a salary increase or seasonality must remain hypotheses until documented. A percentage with a zero budget must be marked as not calculable.

The data model and upstream controls are explained in the automated reporting guide.

Provide the expected document type, period and exact reason for rejection. Remove unnecessary identifiers before testing. The Armanino case published by Anthropic describes a tool for drafting explanations when an audit document is incorrect.

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 the document to send. Add no amount, deadline or commitment that has not been provided.

Check the recipient, document and period before sending. This draft must not automatically trigger a follow-up. The case details and limitations distinguish the deployed tool from the benefits estimated by the publisher.

Work from an approved reporting version after access has been authorised. Define the decision to prepare: recruitment, funding or expenditure trade-offs.

Using only the supplied indicators and commentary, suggest five questions a board member could ask. For each, cite the row or passage that motivates it, identify missing information and distinguish an observed risk from a hypothesis. Recommend no decision without the necessary data.

Compare the questions with the cash-flow forecast. Positive profitability does not rule out a short-term financing need.

Yes: OpenAI’s data-analysis documentation describes file analysis and code use. Distinguish generated prose from a calculation executed by a tool. Request the code or formulas, columns used and exclusions.

Check totals, signs, currencies, empty rows and duplicates separately. Rerun the calculation in Excel, a controlled script or BI. The presence of code validates neither assumptions nor file quality. Do not delegate account sign-off, payment approval or a tax decision to a model’s answer.

Start with fictitious or public data. For internal data, the company must authorise the tool, its access and the processing purpose. Removing a name is not always enough to anonymise a file: amounts, dates and job roles may identify someone.

OpenAI documents no training by default on Business, Enterprise and API data. This does not mean no retention or permission to upload every document. Also check permissions, retention periods, connectors and contractual terms. DPO or security-owner approval depends on the company’s context.

Record authorised uses, excluded data, the reviewer and error procedures in your policy. Use our LLM selection framework to compare environments on your own documents.