A repeated order number can be a valid line
You combine two exports and see the same order number several times. It looks like a quick cleanup, until you remember that an order can contain several items. Comparing the order number alone could discard a valid line. Today’s Reel introduces that distinction; here we work through the review.
Our nine rows were created for this case and contain no customer data. Each has a source row ID, order, line, item, quantity and amount in euros. The first output we want is a list of issues with reasons. A shorter table is not yet the goal.
Choose the key before counting matches
For this case, the combination of order and line number identifies an order line. P-2103 has lines 1 and 2, both valid. Another dataset might also need a store, year or revision. That rule comes from how your records are created; AI should not guess it.
Keep the source file and work on a copy. Add review columns alongside the received values. For these identifiers, we have agreed to ignore outer spaces and letter case: “ P-2101 ” is compared as P-2101, and p-2102 as P-2102. Do not apply that rule to case-sensitive codes or remove leading zeros without checking what they mean.
Nine rows, four different situations
R01 and R03 share order P-2101, line 1, item CAR-01, two units and €48. After the agreed normalisation, the displayed data repeats. Flag both rows to establish whether the same import was included twice. A match alone does not authorise deletion.
R02 and R05 share P-2102 and line 1, but their amounts are €18 and €20. This is a conflict to check against the source record. Choosing the larger or later amount, or averaging them, would change the data without explaining what happened.
R04 and R06 belong to P-2103, with lines 1 and 2 and amounts of €27 and €24: their complete keys differ. R07 shows €18 but has no order number, so it remains incomplete. R08, P-2104 line 1, shows €9. R09, P-2105 line 1, has zero units and €0. Zero is a recorded value, not a gap to fill.
Give AI a review task with a traceable result
Explain the columns, the chosen key, the permitted formatting changes and what must remain untouched. Ask for a review showing source row IDs, matching fields, differences and the next check. Identifying R02 and R05 lets you return to the source; “duplicates cleaned” does not tell you what changed.
RavenPrompts’ existing prompt for detecting duplicate Excel records can help frame this task. Start with a description of the structure or an authorised sample without personal information. Before sharing a work file, check your organisation’s and tool’s rules. Review any suggested formula or script and try it on the copy before applying it more widely.
Choose an operation that preserves what you need
Microsoft distinguishes highlighting, filtering and removing duplicates. Conditional formatting lets you inspect matches without deleting them. Remove Duplicates removes records using the selected columns and keeps the first occurrence. The first row is not necessarily the correct business record. Microsoft recommends keeping a copy before removal.
COUNTIFS can count rows meeting criteria across several columns. Formula names and separators must match your Excel installation. Check the covered range as well: a calculation limited to the sample must not be described as covering the entire worksheet.
Microsoft — Find and remove duplicates
Microsoft — Filter for unique values or remove duplicate values
Reconcile the review before changing the source
This sample has four flagged rows in two pairs: one pair repeats data and the other has conflicting amounts. One row is incomplete and four have neither issue. That accounts for all nine rows. Four flagged rows do not mean four deletions: both occurrences in each pair are included.
Check that P-2103 retains both lines, R07 remains pending and R09 keeps its zeros. On a copy, change a key and add a row at the end, then check whether the result and covered range respond as intended. Record each decision, its reason and a reference to the original entry.
Recent Excel developments still need your judgement
The official Excel Blog described new Copilot data-preparation assistance on 12 October 2026 and updated the post on 16 October. That is recent context, not an announcement made today. Availability and limits depend on the account and rollout. You can practise this review with tools you already have access to.
Microsoft’s Copilot documentation warns that results can be inaccurate and need review. AI can help describe a rule or explain an issue; checking it still means returning to your records. When you close the file, you should be able to see what matched, what remains unknown and why each row was changed.
Excel Blog — Beyond Analysis: Copilot Gets to Work in Excel (12 October 2026; updated16 October)
Microsoft — Frequently asked questions about Copilot in Excel

