"Microsoft retired Excel's COPILOT() function — and that's a sign AI in spreadsheets is maturing"

A year after it began letting testers drop an AI prompt straight into a spreadsheet cell, Microsoft is quietly retiring the COPILOT() function in Excel. The function — which let users type something like =COPILOT("summarize this paragraph") and have Copilot fill the cell with a generated result — will stop working after September 14, 2026, before it ever reached general availability. On its face, that reads like a retreat. Read a little closer, and it looks more like a product team finally admitting where AI actually belongs in a spreadsheet.

It is worth remembering how natural the idea seemed at the time. Google had already shipped a Gemini-powered AI() function in Sheets, and the formula-as-AI-entry-point looked like a clean way to bolt generative intelligence onto the tool people already understood best: the formula bar. If you can write =VLOOKUP(...) and =SUMIFS(...), why not =COPILOT(...)? The appeal was obvious, which is exactly why Microsoft's decision to walk away from it deserves more attention than the "feature killed" headline suggests.

Microsoft's stated reason is straightforward: the function was redundant. Copilot in Excel already lives in two richer places — the floating Dynamic Action Button and the dedicated Copilot side pane — and both do everything the function did, with more room to breathe. Nothing the function could do is actually disappearing; it is just being rerouted to surfaces that are better suited to the job. That framing is not spin. It points at a real insight about how AI fits into productivity software.

The list of limitations Microsoft attached to the function is the giveaway. The company said the COPILOT() function was not recommended for numerical calculations — an almost comical admission inside a spreadsheet application — nor for lookups against data inside your own workbook, nor for anything touching legal or compliance content. It also needed an internet connection and a premium subscription. Strip those caveats away and what remains is a function that was really only good at generating, summarizing, classifying, and pulling text from the web. That is not a spreadsheet calculation; it is a text box wearing a formula costume.

Here is the first insight worth extracting: the formula cell is the wrong shape for AI. A spreadsheet function is, at heart, a pure, deterministic, recomputable thing. Change an input and the result updates; copy the formula down a column and it scales a thousand times. Generative AI is the opposite — probabilistic, context-hungry, and not guaranteed to return the same answer twice. Squeezing it into a cell forced a model built for open-ended reasoning into a container designed for reproducibility. The mismatch was structural, not cosmetic, and it is why the function spent a year in testing without ever feeling essential.

A second, related point is that a single cell simply does not hold enough context for Copilot to be genuinely useful. The most valuable things Copilot does in Excel — understanding a whole table, catching the column you meant rather than the one you named, explaining the steps behind a transformation, iterating when the first answer is wrong — all require seeing the workbook, not a single cell reference. The side pane and the dynamic action button offer that context; a formula cell cannot. Moving Copilot out of the formula bar is, in effect, giving it the headroom to be the tool it was always trying to be.

There is also a portfolio-management lesson hiding in the timing. The function had been slated for general availability in January 2027, which means Microsoft chose to kill a planned public feature rather than ship it and support it. That is a disciplined move — and a rare one. "Put Copilot everywhere" is an easy strategy to announce and a hard one to maintain, because every entry point multiplies the testing, support, and confusion surface. Retiring the function before launch signals a maturing strategy: fewer, richer Copilot surfaces rather than a thin layer of AI sprinkled across every menu.

The competitive context makes the decision more interesting, not less. Google is still running its own experiment with the AI() function in Sheets through its Workspace Experiments program, and the two companies are now essentially testing opposite hypotheses in public. Microsoft is betting that AI belongs in a conversational side panel; Google is betting there is still room for a cell-level function. Neither is obviously wrong — Sheets has a more collaborative, less formula-centric heritage than Excel, so the same function may land differently there — but the divergence is a useful natural experiment to watch.

For everyday users, the practical upshot is reassuringly dull: nothing is being taken away that they will miss. Anyone who leaned on Copilot's text abilities in Excel can reach them through the side pane, usually with a better result, since that surface can see the whole sheet. The only real casualty is a bit of muscle memory for the handful of testers who had gotten comfortable typing prompts into cells.

The broader takeaway is that the spreadsheet industry just ran a clean little experiment in AI interface design, and one of the two giants concluded the formula bar was the wrong place for a language model. AI in productivity tools is still young enough that getting the entry point right matters more than shipping every possible one. Microsoft's willingness to retire a feature before it launched is less a retreat than a sign that the team now understands what Copilot is for — and a conversation, it turns out, is a better interface for a language model than a function call.

Read more: the original Microsoft announcement introducing the COPILOT() function, Google's workspace updates on its rival AI() function in Sheets, and the Neowin report on the retirement.