AI at Work
Copilot in Excel: Review the Plan and Validate the Workbook After AI Edits
Distinguish current chat, plan and edit modes, then inspect scope, formulas, source inputs, independent controls and recalculation using an actual local synthetic workbook.
Before letting Copilot edit an Excel workbook, review its proposed ranges, calculations, and exclusions. Afterward, inspect the actual cells and formulas, compare the results with an independent control, and test a representative input change. A convincing explanation in the chat pane does not establish that the workbook is correct or that unrelated inputs survived.
This LandOffer.ai guide combines current Microsoft documentation with an original synthetic inventory exercise. We created, recalculated, rendered, exported, and reimported a local workbook using a separate spreadsheet tool. Copilot did not edit it, and we did not test a Microsoft account or native Excel session. The completed examples demonstrate an acceptance method, not Copilot's measured reliability.

Confirm the mode and environment before submitting the task
Microsoft's getting-started guide distinguishes three modes. Chat keeps the response in the conversation; plan produces an approach to review before editing; edit can change the workbook directly. The guide currently says the default is edit mode. Read the mode selector in your own session before sending a request that you intend only as a discussion.
A plan is useful because it exposes assumptions before they become formulas or deleted rows. It is not a spreadsheet permission system. After editing, compare the input cells with the original values even when the plan says to preserve them. Nor does reviewing a plan approve a separate use of company data, a new external source, or access your organization has not authorized.
As checked on October 25, 2026, Microsoft lists edit, chat, and plan on Windows, Mac, and the web. Its mobile table differs by device and includes edit functionality still rolling out. Check the documented platform table and the controls actually available to you rather than assuming a desktop workflow exists on your phone.
The Copilot in Excel FAQ lists several eligible subscription routes and says availability can depend on license and tenant configuration. Verify your account's entitlement and organizational settings with the appropriate administrator. A visible Copilot button does not prove access to every feature described online. The same FAQ says editing requires Automatic calculation. Check that condition before diagnosing a calculation problem as a bad prompt.
Define the workbook result before asking for a plan
Our fictional operations analyst, Dara, wants closing stock and a suggested reorder quantity for four synthetic supply items. This is a count exercise, not a purchasing authorization or inventory valuation. All quantities use the same unit within each row. There are no prices, pack-size conversions, reserved stock, or supplier lead times.
The input columns are SKU, Opening, Received, Issued, and Target. The intended closing quantity is opening plus receipts minus issues. Reorder is the positive shortfall between target and closing quantity, with zero when closing stock already meets the target. A target is a supplied teaching assumption, not an optimized stocking policy.
Before editing, Dara identifies Inventory!A7:E10 as original inputs, F7:G10 as the permitted formula area, and F4:G4 as the two summary cells. Headers occupy row six. There are exactly four records, A01 through A04. No sorting, row deletion, external imports, or new source values are needed.
Microsoft's prompting tips recommend naming relevant columns or ranges. Explain what the referenced output measures: ask for closing stock in units, with its calculation rule. Ask the assistant to surface missing inputs rather than silently substituting zero.
Completed asset: a bounded edit request and reviewed plan
The following request and plan are original, fictional teaching assets. They were not submitted to Copilot. They specify the local workbook exercise precisely enough for a human to inspect the intended result.
Request: Propose a plan first. On Inventory, preserve A7:E10 and its four SKU records. Add Closing formulas in F7:F10 and Reorder formulas in G7:G10. Closing is Opening plus Received minus Issued. Reorder is the positive difference between Target and Closing, otherwise zero. Add totals in F4:G4 that include all four rows. Keep a missing numeric input visible in affected results and summaries. Do not sort, delete records, fetch data, add prices, or place orders. Show the ranges and assumptions before any editing.
Dara's completed plan review records the following decisions:
| Proposed action | Reviewed scope | Acceptance evidence required |
|---|---|---|
| Read four stock records | A7:E10 only; existing SKU order retained | Same four identifiers and input values afterward |
| Calculate closing stock | F7:F10; one formula per row | Individual quantities 24, 6, 5, and 23 |
| Calculate target shortfall | G7:G10; no negative reorder quantity | Individual quantities 0, 4, 3, and 2 |
| Summarize results | F4:G4; all four rows, no subtotal counted twice | Closing 58; reorder 9 |
| Handle absent inputs | Missing differs from numeric zero | Blank receipt keeps affected result and totals unavailable |
| Preserve original material | No source changes, extra records, or external data | Input comparison and saved formula inspection |
If a proposed plan instead says “clean the source data” or “remove unused rows,” Dara should ask which cells and why. A broad cleanup step could change the record population or turn a valid zero into missing data, invalidating the agreed calculation. Narrow that step or exclude it before moving into an editing mode.
Decide how to return to the original workbook before applying changes. Microsoft's tips describe AutoSave choices and undo or previous-version recovery. Use the recovery route available for your file and storage arrangement, and keep an identifiable original copy where appropriate. Do not assume a chat transcript contains the original workbook values.
Inspect formulas and inputs before trusting the totals
In our actual local exercise, the four input rows and final outputs were:
| SKU | Opening | Received | Issued | Target | Closing | Reorder |
|---|---|---|---|---|---|---|
| A01 | 30 | 12 | 18 | 20 | 24 | 0 |
| A02 | 15 | 0 | 9 | 10 | 6 | 4 |
| A03 | 0 | 8 | 3 | 8 | 5 | 3 |
| A04 | 40 | 5 | 22 | 25 | 23 | 2 |
For a complete numeric row, the calculation in F7 is B7+C7-D7; G7 takes the greater of zero and E7-F7. The saved workbook also guards required numeric inputs, so an absent receipt is not treated as a genuine zero. Its closing formula is =IF(COUNT(B7:D7)=3,B7+C7-D7,"Missing input"). The reorder formula checks that closing and target are numeric before calculating the shortfall.
Inspect the first, middle, and final formulas. A correct first row does not establish that the formula was filled to the bottom or that its references advanced correctly. Inspect F4 and G4 for the last included row, not just their displayed totals. Formatting a result as a whole number does not establish that the underlying source was numeric.
Compare original inputs separately. In this exercise, we checked every value in A7:E10 against the original fixture, including the zero receipt for A02 and zero opening quantity for A03. The local check also confirmed exactly four distinct SKU values in A7:A10. These small-fixture checks do not establish a general duplicate-detection system for a larger workbook.
An error scan is useful, but a plausible wrong number may contain no Excel error. Our final local scan found no formula-error matches. That finding was combined with expected values and input changes, rather than treated as proof by itself.
Completed asset: the independent result audit
The control calculation was established from the synthetic inputs independently of the worksheet's summary formulas. Opening stock totals 85, receipts total 25, and issues total 52. Therefore closing stock should total 58. Separately, the four target shortfalls are 0, 4, 3, and 2, which sum to 9. Comparing F4 to another cell that simply links to F4 would not provide that independence.
We deliberately inserted an omitted-last-row error into the local summary during validation: SUM(F7:F9) returned 35. It looked like a valid number and raised no ordinary formula error, but disagreed with the control of 58. We then restored the correct summary. This was an intentionally seeded test defect, not an error observed from Copilot.
| Check actually performed locally | Observed result | Interpretation |
|---|---|---|
| All four closing results against fixed controls | 24, 6, 5, 23 | Matches the synthetic input arithmetic |
| All four reorder results against fixed controls | 0, 4, 3, 2 | Matches the stipulated shortfall rule |
| Summary against independent control | 58 closing and 9 reorder | Includes the expected complete population |
| Seeded summary omits A04 | 35 instead of 58 | Independent control detects the range error |
| Change A02 opening from 15 to 17 | Closing total 60; reorder total 7 | Dependent outputs recalculate in the local tool |
| Replace A02 receipt zero with blank | Affected closing and both summaries show Missing input | Missing information is not accepted as zero |
| Change final-row target from 25 to 26 | Reorder total 10 | Final-row calculation affects the summary |
| Restore inputs, export, and reimport XLSX | Original inputs, formulas, and totals match | Saved structure retained in this local round trip |
The validation script recorded twenty-two assertions, including restoration checks. Their number is an audit detail, not a quality score. This exercise covers these inputs, formulas, and local engine behavior. It does not certify native Excel recalculation, Copilot performance, enterprise permissions, procurement decisions, or arbitrary future rows.

Test a changed input and preserve a truthful result record
A workbook can show the expected answer because someone typed the answer into a cell. Change a relevant input in a disposable copy and watch its dependents. Choose a change whose expected effect you can calculate beforehand. Then restore it, recalculate, and confirm the original outputs return before sharing the result.
Our A02 opening change exercised both closing and reorder behavior. The blank-receipt check exercised a different question: whether a missing prerequisite could disappear into a convincing summary. Because we restored every input, the delivered synthetic workbook returns to closing 58 and reorder 9. The intentionally wrong summary was also removed.
For a real workbook, extend this check to the features you actually use. A new record may require range or table expansion; an edited date may change a reporting period; a lookup may depend on unique identifiers. Do not claim those behaviors were tested merely because a four-row example worked. Inspect affected charts and PivotTables if they are part of the requested result.
Save a short result record that names the changed ranges, input comparison, formula checks, independent control, perturbation result, and remaining limitations. A suitable note for this exercise is: “Local synthetic workbook validated against closing 58 and reorder 9. Missing-input and two input-change cases checked; original values restored. No Copilot session or native Excel test performed.” A real Copilot-assisted edit should identify the actual workbook version and tests performed in that environment.
Stop when the workbook exceeds the agreed scope
If editing touches original records, creates an unexplained sheet, or changes a rule you did not request, pause and compare with the plan. Preserve enough evidence to understand the difference before trying another broad repair prompt. Recover the affected version when necessary, then ask for a narrower correction and repeat the relevant checks.
Microsoft explicitly recommends verifying generated material before relying on it. Treat the assistant's completion message as a pointer to inspect its work, not the acceptance record. Before sharing, state which workbook version was checked and which dependencies or issues remain unverified.
Before your next AI-assisted workbook change, write down its allowed ranges and one independent expected result. Use the allowed ranges to check scope and the expected result to investigate a discrepancy. For further workplace reading, visit LandOffer.ai.
Sources and Further Reading
- Microsoft Support: Get Started with Copilot in Excel — modes and platform availability.
- Microsoft Support: Frequently Asked Questions about Copilot in Excel — eligibility, calculation settings, and verification guidance.
- Microsoft Support: Copilot in Excel Tips — range-specific requests and recovery considerations.


