Your First Finance AI Use Case: Start with a Report You Can Check

"Use AI in finance" is an ambition, not a task. A workable first attempt is much narrower: one output, built from approved inputs, with checks agreed before anyone writes a prompt. This article works through one such task from start to finish.

Hypothetical example. The team, figures and illustrative output below are invented. None of it was produced by an AI model or by FinanceWorkroom's tools.

1. Choose a task with a checkable result

Our example is a finance team that wants a first draft of the results section of its monthly management report, built from an approved, prepared dataset. The scope is one month, one entity, three metrics, a compact table and a short summary. A person reviews the draft before it goes anywhere.

In our editorial view, this is a reasonable first task because:

• The output is specific.
• The inputs exist and are approved for use.
• A named reviewer understands both the numbers and the intended message.
• Errors and omissions can be checked against a reference.
• The output stays a draft.

This is a worked example, not proof that management reporting is the right first use case for every team. The same criteria can be used to test any candidate task.

Be clear about who does what. Spreadsheet formulas or rules calculate repeatable totals and variances. AI can help organize, summarize and draft. People own definitions, business explanations, review and release. If a task only needs a known calculation or a fixed transformation, a formula or rules-based process may be all you need.

2. Complete a task card before writing the prompt

The card becomes the request: each field turns into part of the instruction. For example:

Using only the attached September file and the definitions below, draft a results table and a three-to-five-sentence summary for the CFO, in USD thousands. Do not state causes, forecasts or recommendations. If an input needed for a requested figure is missing, say so rather than estimating.

An instruction like "do not estimate" is a requirement you test. It does not guarantee how the model behaves.

3. Prepare inputs and a reference result

For a first learning exercise, synthetic data is enough. Put company data into an AI tool only under your company's approved handling.

State the period, currency, units, sign convention and metric definitions in the input itself, so none of them is left to interpretation. The account-to-metric mapping is supplied and approved as part of the input. It should not be inferred from account names. Define the calculation convention as well:

• Amount variance = actual minus budget.
• Percentage variance = (actual minus budget) / budget × 100, shown to one decimal place.
• If a budget is missing or zero, no percentage is presented as a valid result, and the limitation is flagged.

Then build a reference result independently, using formulas in a workbook you already trust:

For operating expenses, a positive variance is unfavorable.

You can also give the AI the variances already calculated by formula and ask it only to draft. That narrows what the test must cover. Either way, the reference is what you check against.

Include explanations from business owners only if they exist and are approved. Otherwise, list what's missing (here, any explanation of why revenue fell short) rather than asking the AI to fill it in. If your source figures disagree before you start, resolve that first; our guide to checking finance data when reports disagree covers how.

4. Test the normal case and one exception

Run the task, then check the output against tests agreed in advance. The missing-context test matters because generative AI can produce confidently stated but erroneous content, including logic that appears to justify an answer (NIST AI 600-1).

Run the exception test in a fresh conversation or another isolated test context that contains only the modified input and the instructions. The removed budget must not be available anywhere the AI can see it, including an earlier message, the normal-case reference table, a calculated variance column or another attached file. The reviewer keeps the complete reference result. A usable output keeps the operating expenses actual of 1,640, marks the budget as not supplied, and leaves both the amount variance and the percentage variance unavailable. Any reconstructed or estimated budget, or any valid-looking variance, counts as a failure, even if the rest of the report is right.

Illustrative summary that would pass the normal case (written by us, not generated). All amounts in USD thousands.

"September revenue was 4,820, which is 180 (3.6%) below budget. Gross profit was 2,170, which is 130 (5.7%) below budget. Operating expenses were 1,640, which is 40 (2.5%) above budget. The supplied data does not explain these movements; business owners' input is needed before causes are stated."

Three cautions apply:

• Spot checks, such as confirming the revenue variance of (180) and (3.6%), are only a start. Before circulation, the reviewer assesses every material figure and statement relevant to the intended use.
• Asking the same AI to check its own work does not replace the independent reference or human review.
• A passing test is evidence about that output, with that input, those instructions and that tool. It does not establish reliability for future files, model versions or runs. NIST notes that current testing approaches may not reflect real-world use, a gap it says prompt sensitivity can widen (NIST AI 600-1).

5. Continue, revise or stop

• Continue with the bounded draft task if it meets the acceptance checks and the review requirement.
• Revise the inputs or instructions and rerun all the checks if the task is useful but the output fell short.
• Stop or narrow the task if evidence, permissions, definitions or adequate review aren't available.

Passing review approves a draft. It does not authorize autonomous reporting, posting entries, sending emails or publishing.

To reuse the task for October, keep the task card, update the period and inputs, look for structural changes such as new accounts or renamed departments, rebuild the reference, and repeat the relevant tests. A saved conversation or prompt is not evidence that the next report is correct.

Try your first management report with sample data. Select "Create My Monthly Management Report." The guided flow gives you prompts and steps to use with an external AI tool, including a sample-data option and prompts for comparing selected figures with the source. Its sample data is separate from the figures in this article.