Use Copilot in Excel to Find Outliers Before You Write the Summary

You have a spreadsheet full of sales rows and a short update due this afternoon. The first question is not “What story can I tell?” It is “Which rows deserve a second look?”

This guide uses Microsoft Copilot in Excel to produce a small outlier review from fictional data. It is a documentation-based guide, not a recorded personal test. Microsoft says Copilot in Excel can analyze formatted tables, surface trends and outliers, and return summaries, charts, or PivotTables. Its direct-answer workflow can also show the analysis code behind a response. The result still needs a human check.

What you will create

A review table with the unusual rows, the reason each row was flagged, and a check against the original values.

You need Excel for Microsoft 365 with Copilot available on your account, a workbook containing a formatted table, and data you are allowed to analyze. Do not use confidential customer data for a practice run.

Fictional input

Create a sheet named April sales and paste this table into Excel:

Order ID	Region	Product	Units	Revenue	Refunds
A-101	North	Desk lamp	18	900	0
A-102	North	Monitor stand	15	750	1
A-103	South	Desk lamp	16	800	0
A-104	South	Monitor stand	14	700	0
A-105	East	Desk lamp	17	850	0
A-106	East	Monitor stand	13	650	0
A-107	West	Desk lamp	2	100	0
A-108	West	Monitor stand	12	600	0
A-109	Central	Desk lamp	19	950	0
A-110	Central	Monitor stand	14	700	0

The low Units value for A-107 is deliberate. It is a practice case, not evidence that Copilot found it.

Steps

Select any cell inside the pasted range.
Format the range as a table and keep the headers enabled. Give it a clear name such as AprilSales.
Select the Copilot button in the Excel ribbon.
Ask:

Review the table AprilSales for unusual values or rows. Return a review table with Order ID, the field that looks unusual, the source value, the comparison or reason for the flag, and a question for a human reviewer. Do not change the workbook. Do not call a row fraudulent or incorrect. Separate observations from guesses.

If Copilot offers a table or visualization, insert it only on a new sheet named Copilot review. Microsoft notes that some returned tables and visualizations are static, so do not treat the inserted result as a live calculation.
Ask:

For every flagged row, show the original Order ID, Region, Product, Units, Revenue, and Refunds values from AprilSales. Explain the comparison used. If the table does not contain enough information to explain a flag, write "needs human context".

Check each flagged row in AprilSales. A flag is usable only when the Order ID, field value, and comparison match the original table.
Add a Human check column. Use confirmed from source, needs context, or not supported by source.

Acceptance check

The workflow passes when every flagged Order ID exists in the source table, every copied value matches the source row, the review distinguishes an unusual value from a confirmed business problem, the analysis has not changed the source data, and a reviewer can explain each flag without reopening the chat.

For this fictional input, A-107 is a sensible row to inspect because its units are much lower than the other rows. That is an editorial expectation from the sample, not a Copilot result.

If something goes wrong

Copilot cannot see the data. Confirm that your cursor is inside a formatted table and that the headers are clear. A loose block of cells may not meet the documented input requirements.

The answer is vague. Name the table and columns. Ask for Order IDs and source values, not just “interesting insights.”

A chart cannot be refreshed. Treat it as a static artifact. Recalculate or rebuild it from the source table before using it in a recurring report.

Copilot calls a row wrong. Use “flagged for review” instead. An outlier needs context. It is not automatically an error, fraud case, or business decision.

Limits and privacy

Use fictional or approved data while learning. Review any generated formula, code, table, or chart against the source. Do not paste secrets or personal data into an unapproved workbook. Availability and exact controls depend on your Microsoft 365 setup.

For a source-linked review workflow, see Use ChatGPT to Turn a Spreadsheet into a Data Dictionary. For a simple review contract before automation, see Before You Give AI a Work Queue, Define What Good Looks Like.

Want one practical workflow each week? Join The Weekly Verdict.

Sources

Evidence label: Documentation-based guide with an illustrative example. No personal test, reproduced demo, measured result, or client result is claimed.