Point Copilot at a clean table, ask for the patterns that matter, then verify every total and cause yourself. That loop turns a spreadsheet into an answer you can defend.
What you'll learn
Prepare and name an Excel table Copilot can analyze reliably
Write insight prompts that carry goal, context, source, and expectations
Match the evidence you request to the type of claim being made
Verify a Copilot insight with your own arithmetic
Separate an observation from an unproven explanation
You've exported three months of sales and the status meeting starts in an hour. There isn't time for a full analysis project. You need to see what changed, which region carried the quarter, and which number is likely to draw a question. Copilot in Excel is useful for that first pass, provided the question is specific and you check the answer.
Clean, named tables come first
Copilot works with structured worksheet data, so start by making the table uneventful. Give each column one field, each row one record, and the table one descriptive header row. Dates must be dates and revenue must be numeric. Keep title rows, notes, and subtotals outside the records. Decide what a blank means as well. "Not recorded," "not applicable," and zero describe different situations, and Copilot can't choose among them for you.
Convert the range with the table command in your version of Excel, then use a specific name such as SalesData. That name fixes the source and the column set in every prompt. It also gives you a straightforward place to return when an answer looks wrong. A named table doesn't guarantee accuracy. It does remove one common ambiguity and makes the result easier to audit.
Before you ask a question, scan for blank headers, numbers stored as text, inconsistent dates, stray subtotal rows, and accidental duplicates. Five minutes here can spare you from reviewing a polished answer built on a damaged table.
Name the table, then name it in the prompt
A selected cell is not a reliable way to tell Copilot what to analyze. Convert your range to a table, give it a name, and repeat that name in every prompt. When an answer looks wrong, the first thing to check is whether Copilot used the source you meant.
A useful insight prompt
An insight is a meaningful observation about your data: a trend, a comparison, a concentration, or an unusual value. "Revenue increased in May" is an observation. "Revenue increased because the sales team ran a promotion" is an explanation, and a table of dates, regions, and revenue cannot prove it. Keep that line sharp, because Copilot will offer explanations that sound reasonable and aren't grounded in anything on the sheet.
Useful insight prompts include a goal, context, source, and expectations. The Anatomy of a Great Prompt explains that structure from the beginning. In Excel, tell Copilot what you need, why you need it, which named table to use, and what the answer must show or avoid. You might require the rows behind each pattern and add "do not infer causes."
Work in two passes. Start with a broad scan for the patterns worth your attention. Then choose one observation and request the exact rows and arithmetic behind it. The first pass finds a candidate. The second gives you something you can defend in the meeting. Packing both jobs into one oversized prompt often produces an answer that's difficult to inspect.
Scan a table for its three biggest patterns· excel
Bad example
Analyze this and give me the main insights.
Good example
Analyze only the SalesData table for a monthly sales review. Identify the three most important patterns in Revenue, grouped by Region and OrderDate. For each pattern, list the supporting dates, values, and columns, and separate what the data shows from possible explanations. Do not change the worksheet.
Why this works: Three ranked patterns, each with the rows and values behind it, and any "why" kept separate as a hypothesis still awaiting proof.
Evidence has to match the question
Different claims need different proof, and naming the evidence you expect is how you catch a weak answer before it reaches a slide. Ask for the evidence that fits the shape of the claim:
If the claim is about…
Ask Copilot to show…
A trend over time
Every period in order, a total for each, and the arithmetic of each change
A group comparison
Count, total, and average for every group over the same date range
An unusual value
The value, its source row, and the comparison group used to judge it
A concentration
Every contributing row and how the largest compares with the group total
A possible cause
A column or source that directly records that cause, otherwise label it unverified
Notice how much these differ. A trend needs ordered periods and per-period totals. Two isolated numbers are only a comparison. A group comparison needs counts as well as totals, because a big total made of many small orders means something different from one made of a single huge deal. Once the broad scan hands you a candidate, narrow in and make Copilot defend it row by row.
Put one claim on trial· excel
Bad example
Check whether West is doing best.
Good example
Recheck the claim that West has the highest total Revenue, using only SalesData. List every West row with OrderDate, Product, and Revenue. Add those Revenue values and show the arithmetic, compare the largest row with the West total, and state whether one record dominates the result. Do not infer a cause absent from the table.
Why this works: A West total you can add up yourself, plus a flag if a single large transaction is carrying the region, distinguishing a real trend from a one-off.
Treat Copilot's answer as a claim
Review every response as a claim. Read the observation and the source it names, find the contributing rows, and recompute the measure yourself in a scratch cell, on paper, or by hand. Compare that result with Copilot's number. If they differ, ask Copilot to revise the insight using only the rows you've listed, then run the check again.
Asking Copilot to check its own arithmetic isn't independent verification. It may repeat the same mistake with the same confidence. The wording tells you nothing about whether the total is right. A West total of 12,610 that you've added from the source rows does.
Watch for two particular problems. Copilot may supply a reason the table can't support. In that case, require a column that records the reason or keep the observation without the explanation. It may also build a result on one dominant row. Report that concentration as a limitation. These are small checks, and they keep you from carrying a number into the meeting that you can't defend.
A cause is not free
When Copilot explains why a number moved, check whether a column records that reason. "Discounts drove the June increase" needs discount-versus-revenue evidence. A DiscountPct column sitting nearby isn't enough on its own. If the evidence isn't there, keep the observation and label the explanation an unverified hypothesis.
Turn a vague answer into an evidence table· excel
Bad example
Put your answer in a useful table.
Good example
Make your previous answer specific. Return a table with the columns Observation, Evidence, Values, Calculation, and Limitation, using exact OrderDate, Region, Product, and Revenue values from SalesData. Do not infer any cause that is not in the worksheet.
Why this works: A structured answer where every claim is tied to real rows and arithmetic, and every gap is named as a limitation, ready to paste into a review.
A worked example, verified
Say Copilot scans a ten-row sample and reports three patterns: West is the largest region, Analytics is the largest product, and one West Analytics record looks unusually large. Each is a hypothesis until you check it.
Group the West rows and add them: 2160 + 1750 + 7200 + 1500 = 12,610, across four records, for an average of 3,152.50. Do the same for East and North and you get 8,520 and 4,910, so West is the largest region. But look closer at those four rows: the single 7,200 Analytics transaction is larger than the other three combined, and more than half the West total. West's lead is real, and it is also fragile, resting on one deal.
Now the time view. April totals 9,610, May 10,970, and June 5,460, so revenue rises 1,360 from April to May and then falls 5,510 in June. That's a genuine month-to-month movement, but across ten rows and three months, it is far too small a sample to call a trend. The verified conclusion states the numbers, names the concentration, and admits the limitation. That's what "checked" looks like, and it took about a minute.
Try it yourself
Audit an insight in ten minutes
Work through a claim that is deliberately a little wrong. You'll see exactly where the arithmetic stops supporting the answer.
01
Open a small sales table of your own, or the SalesData sample, and confirm it is a named Excel table with clean headers.
02
Ask Copilot for the single region with the highest total Revenue, and to list every supporting row with its Revenue value.
Hint: Name the table in the prompt so scope is never in doubt.
03
In an empty cell, add those Revenue values yourself and compare your total with Copilot's.
Hint: A quick =SUM over the listed rows is enough.
04
Check whether one row is more than half the total. If it is, write one sentence naming it as a concentration limitation.
05
If your total and Copilot's disagree, ask it to correct the insight using only the rows you listed.
A verified regional claim with your own arithmetic beside it, plus a feel for how one dominant transaction can masquerade as a trend.
Key takeaways
A clean, named table is the real starting point: Copilot analyzes what you point it at.
Good insight prompts carry a goal, context, a named source, and clear expectations.
Work in two passes: scan broadly, then make one claim defend itself row by row.
Recompute the number yourself. Asking Copilot to check its own math is not verification.
Keep observations and explanations apart, and report a dominant row as a limitation.
Check your understanding
1. Copilot reports, "Revenue increased in June because discounts attracted more customers." Your table has only Month, Region, Product, Revenue, and DiscountPct. What should you do?
2. Which statement is an observation your sales table can support?
3. A strong Copilot insight prompt should identify which four things?
4. Copilot lists the West rows and reports a West total of 12,610. What's the strongest way to verify it?
5. One 7,200 transaction accounts for more than half of West's 12,610 total. How should you report the West result?
Frequently asked questions
Terms used in this lesson
insight
A meaningful observation about data, such as a trend, comparison, concentration, or unusual value, as opposed to an explanation of why it happened.
Excel table
A structured range with a header row, data rows, and a name you can reference in prompts and formulas.
outlier
A value that differs substantially from a stated comparison group. It may be a legitimate large transaction rather than an error.
concentration
When a single record or small group accounts for a large share of a total, making a group result depend heavily on one row.
