Copilot in Excel: what it does well and where it struggles

Copilot in Excel now means the chat pane, not a formula. Microsoft retired the in-cell COPILOT() function on September 14, 2026, about a year after it launched, and folded its use cases back into the side pane. Microsoft's own documentation had already said not to rely on the function for numerical calculations, lookups inside your workbook, or anything with legal or compliance weight, which was most of what an SMB actually needs from a spreadsheet.

What remains is the Copilot pane: a chat-and-edit assistant that can draft or fix formulas, summarise a tab, explain why a number changed, and, on paid tiers, run Python directly against your data for analysis a formula cannot easily do.

It is genuinely fast at the drafting and explaining work that eats an analyst's morning. Give it a plain-English request and it writes the formula, and it can walk back through an existing one and explain what each part does, which is useful for anyone who inherited a workbook they did not build.

Two features make it safer to hand to a team than a bare chat window. Workbook Rules let you set naming conventions and formula standards once, and Copilot applies them on every request in that file afterwards. Plan with Copilot shows which ranges and formulas it intends to touch before it touches them, and every edit lands in the Show Changes pane attributed to Copilot, so a reviewer can see exactly what an AI wrote versus what a person wrote.

The core problem is that it can write a formula confidently and still get it wrong, with no error flag to tell you. Reviewers testing the retired COPILOT() function found it would return a fluent, plausible, factually wrong answer with nothing to distinguish it from a correct one. That risk does not disappear when the same model moves from a cell formula into the chat pane.

It also misses the kind of silent logic error a careful analyst catches on sight: a SUMIF that quietly excludes rows because the criteria text does not match exactly, or a lookup that returns the wrong row on a table with merged cells, blank rows, or inconsistent headers. A workbook that looks fine to a human but has stray spaces or mixed data types in a key column will trip it up more often than it will admit to.

Yes, for the drafting and explaining work, and only with a human checking the output before it leaves the building. Treat it the way you would a fast junior analyst on their first month: useful for a first pass, not yet trusted to sign off a number that goes in front of a client or a board. That means using it to speed up the model, not to skip the review step that catches the model's mistakes.

The practical sequence that keeps this safe is short: draft with Copilot, read what it changed in the Show Changes pane, spot-check the logic on anything with a lookup or a conditional sum, and only then share the file. Skip the middle two steps and you are shipping an AI's confident guess under your own name.