AI use case · Finance

AI for data analysis

A practical data analysis workflow with a synthetic example, reusable prompt, expected output, review checklist and common failure to avoid.

By NMR Infotech · Updated 4 October 2026 · Editorial standards

Workflow overview

What can AI help with?

Teams need help structuring questions about tabular data.

Use a synthetic dataset; ask for an analysis plan; inspect formulas and missing values; verify results independently.

Tools to explore: ChatGPT, Microsoft Copilot. Choose an organization-approved account and verify the features available to it before using the workflow.

Example input

Start with a clearly bounded input

Illustrative example · not client work

Synthetic table: April actual 120, budget 100; May actual 90, budget 100. Values are in the same unspecified unit. No explanation for either variance has been supplied.

Identify the person responsible for the task, the source material the tool may use and the format needed by the next person in the process. Keep missing facts visible instead of asking the model to fill them from guesswork.

Reusable starting point

Try this prompt with permitted material

Replace the bracketed fields with approved context. Use the example to practise the method before considering a connection to a live business system.

Analyze the synthetic table using only the provided data. Calculate actual minus budget and percentage variance using budget as the denominator. Show your arithmetic, state units as unspecified if absent, and separate numerical results from possible explanations. Do not invent business causes. List checks for missing data and zero denominators.
Table: [insert synthetic rows]

Expected result

What should a useful output contain?

April

Absolute variance 20; percentage variance 20%. The cause is not supplied.

May

Absolute variance -10; percentage variance -10%. The cause is not supplied.

Data questions

Confirm units, definitions, period boundaries and whether the dataset is complete.

Before the output is used

What should the reviewer check?

Recalculate each result independently.
Check denominators, units, rounding and sign conventions.
Inspect missing values and unusual records before summarizing.
Keep explanations clearly separate from calculated results.

Learn from an error

A failure worth testing for

The narrative attributes April’s variance to a marketing campaign that is not in the data. Delete the invented cause and state what additional evidence would be needed.

Keep a record of the error and the correction. Re-test the revised instruction with a different input so a change that fixes one example does not hide another problem.

Evaluate the whole task

How can you assess whether this helps?

Use known-answer examples to measure calculation correctness and review effort. Do not use the output as a financial recommendation or assume that a plausible chart is accurate.

Agree the quality criteria before comparing task time. If the reviewer cannot verify the answer or the workflow repeatedly exceeds its source boundary, narrow the scope or return to the established process.

Common questions

Can we use real business spreadsheets?

Only when the organization has approved the data and tool environment. Synthetic data is sufficient for learning the analysis and verification method.

Is this a completed client project?

No. This is an illustrative learning workflow. It shows a way to frame and review a task, without claiming client deployment, measured savings or guaranteed results.

Related learning