Excel's =COPILOT() function, the formula that let users pull generative AI text straight into a spreadsheet cell, is being discontinued. According to The Register AI, Microsoft is retiring the feature less than two years after unveiling it as one of the flagship additions to Microsoft 365 Copilot.
For a company that has spent years pushing Copilot into every corner of Office, walking back a native AI formula is a notable reversal. COPILOT() was supposed to be the moment generative AI stopped living in a chat sidebar and became a first-class citizen of the grid itself — a function you could drag across a range, reference from other cells, and chain together the same way you'd use SUM() or VLOOKUP().
That ambition, it turns out, ran straight into how spreadsheets are actually built to work.
What the function was supposed to solve
COPILOT() let users write plain-language instructions inside a cell and have Excel return AI-generated content wherever the formula sat, pulling in other cells as context. Typical use cases included:
- Summarizing a block of text or a column of customer feedback into a short paragraph
- Classifying rows by sentiment, category, or priority
- Extracting names, dates, or other entities from unstructured text
- Drafting short product descriptions or email snippets from structured data
The function required a Microsoft 365 Copilot license, ran only on Excel for Windows and the web, and for most of its life stayed inside limited preview channels rather than shipping broadly to every Microsoft 365 subscriber. That preview status is itself telling: the feature apparently never graduated to a fully supported, general-availability capability in the two years since it was introduced — which, in our estimation, likely made it an easier line item to cut than a mainstream feature with an established user base.
Why a formula-based AI call was always an awkward fit
Spreadsheets run on a simple contract: a formula in a cell should return the same result every time it recalculates, given the same inputs. VLOOKUP() and SUM() are deterministic. A live call to a language model is not — ask COPILOT() to summarize the same paragraph twice and there's no guarantee the wording matches, which becomes a real problem the moment a workbook needs to be auditable, reproducible, or shared with a colleague whose formulas should match yours.
There's also the cost structure. Any full recalculation of a workbook that leaned on COPILOT() across many rows meant firing off a fresh batch of model calls, with the latency and token spend that implies. In a sheet with thousands of rows, that's not a minor inconvenience — it's a file that takes noticeably longer to open and burns Copilot capacity on every keystroke that triggers a recalc.
What it means for anyone building on Office AI
The practical lesson isn't don't put AI in Excel — it's that live, cell-level AI calls fight the tool's own architecture. A few patterns hold up better:
- Treat AI output as a one-time enrichment step — via Power Automate, an Office Script, or a direct API call — that writes static values into cells, rather than a live formula that recalculates on every change
- Reserve Python in Excel or an external script for transformations that need to be repeatable and auditable, and keep generative text to genuinely one-off drafting
- Snapshot or cache AI-generated content before sharing a workbook, so recipients see exactly the numbers and text you saw
It's also a reminder that Microsoft will retire Copilot features that don't gain traction, even ones announced with fanfare. Anyone building against Microsoft 365 Copilot's expanding surface area should treat individual features as subject to change, not as stable long-term dependencies.
AiiN's takeaway
COPILOT() is a useful case study in what happens when generative AI gets bolted onto a tool built around deterministic recalculation: the idea of AI as a formula was elegant on a slide and messy in a real workbook. For teams building their own AI-in-spreadsheet tooling, the takeaway is to keep nondeterministic model calls out of the recalculation graph entirely and treat them as a data pipeline step instead. Excel's version of that experiment, at least, is now headed for the bin.