Copilot for Data Analysis: Choose a Surface and Check the Work

Use Copilot for data analysis in the Microsoft app that holds the relevant data, such as Excel for workbook analysis or Power BI for report questions. Define the metric and date range before asking, then check the underlying calculation before using the conclusion.

fenbs can retain the analysis question, agreed data definitions, and outstanding review tasks between sessions. It does not run the analysis or connect to your Microsoft data. This guide covers choosing a work surface and reviewing the result within our AI at work library.

How do you start an analysis you can check?

1. State one decision the analysis should inform. For example: Which service category needs a closer review because its completed jobs fell compared with the previous quarter?
2. Identify the authorized workbook or report, its reporting period, and when its data was last updated.
3. Define the fields and measures that matter, including exclusions, missing values, and whether you are counting rows or unique jobs.
4. Ask for one calculation with its assumptions and supporting output. Reproduce a small part independently before expanding the analysis.

This order makes a confident answer easier to evaluate. A neat chart cannot tell you whether canceled jobs should count as completed work. Decide what the measure means first, then ask the assistant to calculate it.

Should you work in Excel or Power BI?

Choose by the data and the review you need. Excel is a practical starting point when your work is already in a structured workbook. Microsoft’s Excel analysis guide↗ (opens in a new tab) describes asking natural-language questions about tables or table-like ranges, reviewing responses, and inspecting generated analysis code.

Power BI is a different starting point: your organization may already maintain a report and a semantic model, which is the defined structure and calculations behind the report. Microsoft’s Copilot for Power BI overview↗ (opens in a new tab) describes report questions and summaries, as well as separate experiences for report creators. Check its requirements for your organization’s setup.

Microsoft also documents analyzing Power BI reports from Excel↗ (opens in a new tab) by adding a report to a Copilot request. If that option is available, keep the report identity and reporting context visible. Asking from Excel does not make two differently defined measures equivalent.

Do not treat the Copilot name as a guarantee that every app sees the same files or supports the same operations. Start from the controls and sources actually available to your work account. Our Copilot versus ChatGPT guide addresses the wider product choice; this workflow stays with Microsoft data surfaces.

What definitions belong in your analysis request?

Write the business meaning beside the column name. A field called Revenue might mean invoiced value, collected payments, or something else. A field called Date might mean order creation, completion, or payment. Ask the data owner when the distinction is unclear.

• Question: What are you comparing, and what decision might follow?
• Scope: Which source, reporting period, categories, and records belong in the calculation?
• Definitions: What do the metric, denominator, units, and dates mean?
• Treatment: How should duplicates, cancellations, refunds, and missing values be handled?
• Output: Which totals, supporting rows, and explanation would let someone check the answer?

Use Copilot instructions examples for reusable request patterns, but keep the actual definitions beside this analysis. A generic instruction to be accurate cannot resolve an undefined measure.

How do you verify the numbers and the explanation?

Start with record counts and totals. Confirm that the analysis used the intended source and period, then reproduce a small category or date range with an ordinary formula or known report measure. Check whether filters, duplicate records, or missing dates explain any difference.

Ask to see how a percentage was calculated. A change from 20 jobs to 30 jobs is a different claim from a change in that category’s share of all jobs. Check denominators and units before accepting either statement. Keep displayed rounding separate from the underlying calculation.

Inspect charts as well as numbers. Read the axes, labels, units, and any excluded categories. A chart can emphasize a small difference or hide an important missing group without containing a calculation error.

Finally, separate observation from cause. A lower completion count does not establish weaker demand, poor staff performance, or a need to change the budget. Those are possible explanations requiring more evidence. Record what the data supports and what a person still needs to investigate.

What should you retain after the result is reviewed?

Save enough information to repeat the review: the source link or identifier, reporting period, refresh date, agreed definitions, calculation method, checked findings, and unresolved questions. Keep sensitive row-level data in its approved system. A project note usually needs the meaning of the result rather than a copy of the entire dataset.

If a manager chooses to investigate one service category, record that choice separately from the calculation. The analysis established a count; the person chose an action. fenbs Decisions and rules can preserve that distinction alongside the review tasks, so another assistant does not turn an exploratory pattern into settled policy.

How to do this with fenbs

1. Create a free account, then add the project the analysis supports.
2. Manually save the question, source reference, reporting period, and agreed definitions as project context. Include only information appropriate for the kanban board’s members.
3. Add tasks for checks that remain, such as reconciling totals or confirming the treatment of refunds. Write what was checked and what evidence is still missing.
4. Record the actions a person approved in Decisions and rules. Connect a supported assistant using its integration guide if you want it to read the selected record.
5. Save Where we left off with the next check and any question waiting on a person. Keep the underlying workbook or report as the analysis source.

fenbs is not the right fit for spreadsheet calculation, statistical modeling, or live dashboard analysis. Use it for the work record when your question, definitions, and decisions need to survive a change of session or assistant.

Where should you go for a different analysis workflow?

For model-specific workflows outside Microsoft, see ChatGPT for data analysis, Claude for data analysis, or Gemini for data analysis.

The AI data analysis review guide covers the broader selection and checking process. For interviews, feedback, and other text evidence, use the separate guide to AI qualitative data analysis.

Within Microsoft, keep the same disciplined stopping point: a result is ready to inform a decision when you can explain its source, its definition, its calculation, and its remaining uncertainty. If one is missing, the next action is a check rather than a recommendation.