Open the official Claude for Microsoft 365 listing and select Get it now. Check that the publisher is Claude by Anthropic for Office before installing. The listing describes an add-in that can read and change your document and send data over the internet.

Open Excel, activate the add-in, and sign in with Claude. Follow Anthropic's installation steps: on Windows, check Home → Add-ins; on Mac, check Tools → Add-ins.

Open the practice workbook and save a separate working copy. Keep the checked reference closed while you try the exercise. Open only the files you need and let each task finish before editing cells yourself.

The practice workbook starts with three sheets:

Start by stating what each row means and what the report should count. The practice data uses one row per order, US dollars, and US month/day/year for dates written with slashes.

Paste this into the Claude sidebar:

Read the Read Me, Raw Data, and Old Report sheets. Make no changes yet. We need a September 2026 sales report. Count only Paid orders dated September 1 through September 30. Net sales means Sales USD minus Refund USD. Count each order once. Check for duplicate order IDs, inconsistent labels, dates stored as text, amounts stored as text, missing fields, and formulas that break these rules. Cite the sheet and cells for each issue. Explain how Old Report calculates its total. List the fixes you propose. Flag unclear values for my review. Wait for my approval before editing. Use only this workbook; do not browse, connect other apps, or send anything.

Read the response beside the source cells. Follow a cited cell reference and check that it supports the claim.

A useful diagnosis should identify both repeated order IDs, the text-formatted amounts, inconsistent labels, and the missing region. It should also question a formula that simply adds the raw sales column.

The meaning of the data comes from you. A column called “sales” does not tell Claude whether it includes tax, canceled orders, or refunds. State those rules before it starts.

Review the proposed fixes, then send:

Create a Clean Data sheet and a Change Log. Keep Raw Data and Old Report unchanged. Follow the Read Me rules. Convert dates into real Excel dates and amounts into numbers. Trim spaces and standardize the listed region and status names. A missing region becomes Unassigned; keep that order and its amount. Keep one row per Order ID only when all normalized source fields agree. Record the source rows you exclude. If two rows with the same ID disagree, stop and ask me. Do not guess missing amounts. Put the cleaned rows in an Excel table named Sales. Retain the original six columns and add a Net sales USD formula column and a Raw source rows column. Keep canceled and out-of-period orders in Clean Data so the report can filter them. Record each correction and its source cells in Change Log.

In the supplied workbook, 22 raw rows should become 20 distinct orders. Two extra rows repeat orders already present. The order with no region must remain, labeled Unassigned.

Removing a repeated export row differs from removing a real order. A shared customer name, date, or amount is not enough to decide that two orders are duplicates.

Pro member resource: Use the separate answer key to check the cleaning decisions, including the exact repeated rows and changed labels. Keep it out of Claude's first attempt so you can judge the result fairly.

The old report returns a number without showing a formula error. That makes the mistake easy to miss: it adds raw amounts without applying the reporting rules.

Ask for a new report:

Create a Report sheet using formulas that reference the Sales table. Leave Old Report unchanged. Use a visible reporting-month cell set to September 2026. Include only Paid orders within that month. Show paid-order count, sales before refunds, refunds, net sales, and net sales per paid order. Explain the formula for each metric. Add a regional summary that includes Unassigned. Its total must match the headline net sales figure. Use formulas for calculated results, not pasted totals. Handle a month with no paid orders without dividing by zero. List any remaining data concerns.

Inspect the formulas before asking for charts. The net sales per paid order calculation divides total net sales by the number of eligible paid orders.

Once the figures pass your checks, request the chart:

Add one column chart showing net sales by region. Link it to the visible regional summary so it changes when those formulas recalculate. Label the chart with the reporting month and USD. Start the value axis at zero. Keep the source table visible and the chart clear of other cells. Format dollar amounts consistently, use clear headings, and freeze the data headers. Keep the layout simple enough to read without zooming in.

Excel tables can extend calculated columns as rows are added. That makes them useful for a report you intend to reuse, but you still need to test the formulas and chart.

For the supplied fictional workbook, these are the correct September results:

We calculated these figures independently from the sample inputs. They are an answer key, not a claim about what Claude produced in a live test.

Spot-check two orders. S1006 has $1,200 in sales and a $120 refund, giving $1,080 net sales. S1011 has $650 in sales and a $50 refund, giving $600. Check that each appears once in the report.

The $300 Unassigned order must remain in the total. Two canceled September orders and two orders outside September must stay out of the report.

If something differs, ask:

Trace the difference between this report and the expected result. Show the source rows, filters, and formulas involved. Explain the cause before changing anything. Do not replace a formula with the expected number just to make the check pass.

A matching total is one check. Also inspect the rows and formulas. Two separate mistakes can cancel each other out.

Save another copy of the reviewed workbook. Inside its Sales table, add one test order:

Confirm the new row is part of the table and its net-sales formula fills in. The report should now show 17 paid orders, $15,480 in net sales, and $4,400 for East. The East bar should update too. Net sales per paid order should display $910.59.

If those values do not change, inspect the table boundary, formula coverage, chart source, and Excel's calculation setting. In Manual calculation mode, results can remain stale until recalculation.

Remove only the test row and confirm the report returns to its original figures. Keep a record of both checks.

Then ask Claude to add a short monthly-use note:

Add instructions explaining where new exports belong, which cleaning rules need review, how to change the reporting month, and how to run the checks again. Keep the sample source data intact. Do not create a schedule or change other files.

Next month's export still needs review. New region names, duplicate IDs, and missing values can break assumptions that worked for this sample. A reusable report does not clean a fresh export by itself.