If you thought Excel’s AI was already busy enough suggesting charts and answering questions in a side panel, Microsoft just pushed the assistant inside the grid itself. This month, the company began rolling out a new =COPILOT() formula that lets you type a natural-language prompt directly into a cell, point the formula at a range, and have an LLM return results that spill into the sheet — classifying feedback, summarizing text, generating descriptions, building mini-tables and more.

What it is (and how you actually use it)

Think of =COPILOT() like any other Excel function, except its “work” happens by calling a large language model. The basic syntax Microsoft shows is:

=COPILOT(prompt_part1, [context1], [prompt_part2], [context2], ...)

So to classify a column of customer comments, you might write:

=COPILOT("Classify this feedback", D4:D18)

or to make product blurbs from specs:

=COPILOT("Create a description for this product based on its specs", B2:B8)

Because the function is integrated into Excel’s calculation engine, results update automatically when source cells change, and you can nest COPILOT inside other Excel machinery (IF, SWITCH, LAMBDA, WRAPROWS, etc.). That makes it feel — for better or worse — like just another formula.

The engine under the hood

Reporting indicates Microsoft is running the function against an OpenAI model (reported as gpt-4.1-mini), which is why the results are natural-language first. Microsoft itself frames the feature as the successor to the experimental LABS.GENERATIVEAI function that it trialed inside Excel Labs in 2023, but now it’s a built-in function available to Copilot customers in preview.

Limits, guardrails and privacy

Microsoft is explicit about two practical limits right now:

• The COPILOT function is model-grounded: it cannot fetch live web pages or reach into enterprise stores (SharePoint, Teams, etc.) — at least initially. If you want enterprise or web sources included, you must import the data into the workbook or wait for future updates that expand its knowledge sources.
• Usage is throttled to keep compute predictable: the function currently supports 100 calls every 10 minutes and up to 300 calls per hour (Microsoft explains workarounds such as passing arrays so a single call can process many rows). Also, Microsoft warns that the function is not meant for numerically heavy-math or high-stakes legal/regulatory decisions because LLMs can return incorrect responses.

On privacy, Microsoft says data is passed through COPILOT is not used to train its models — the inputs remain “confidential” and are used only to produce the requested output. That’s an important distinction for businesses that have been nervous about shipping customer or internal text into third-party models.

Who can use it today (and how you get it)

The feature is rolling out to Microsoft 365 Beta Channel users who have a Microsoft 365 Copilot license — Windows and Mac builds are supported now, with web availability promised soon through Microsoft’s Frontier/web preview channels. Distribution is staged, so even some Beta users report waiting for the feature to be enabled on their accounts.

How it stacks up to Google Sheets

This isn’t a unique idea: Google added a similar cell-filling AI capability to Sheets earlier in the summer that uses Gemini and Workspace Labs tooling to generate text into selected ranges (Google’s rollout supports up to 200 selected cells per generation, for example). The two approaches converge on the same user problem — automating repetitive text work inside a spreadsheet — but each company’s limits and licensing models differ (Microsoft ties this to the Copilot add-on; Google’s functionality has been part of Workspace Labs / Gemini access paths). If you live in spreadsheets, this now becomes a platform choice as much as an AI capability.

Practical uses (and pitfalls)

What you’ll see people trying right away:

• Bulk classify and tag survey responses or support tickets inside a table.
• Summarize long comments into one-line takeaways for reporting.
• Generate marketing copy variations from product specs directly into a campaign spreadsheet.
• Auto-format a column of messy text into consistent labels, lists or mini-tables that spill across rows and columns.

But caveats matter: the model can invent plausible-sounding but incorrect facts (hallucinations), it won’t know things outside the cells you give it, and frequent recalculation can create compute costs and weird nondeterminism in results. Microsoft suggests reviewing outputs, and treating COPILOT as an assistant rather than an infallible calculator.

Why this matters

Putting LLMs inside cell formulas changes the mental model for spreadsheet power users. Historically, automation in Excel meant VBA, Power Query, or complex formula workarounds; now non-developers can ask the sheet to “do the thinking” in plain English. That broadens who can automate workflows — but it also pushes organizations to rethink governance, auditing and accuracy checks in a place where small mistakes can cascade into big business decisions.

If you live in Excel — and a lot of knowledge work does — =COPILOT() is the sort of change that quietly makes the grid more like a writable AI canvas. It won’t replace skilled analysts, but it will let more people ask spreadsheets to do the tedious parts of the job in plain English. Try it carefully, double-check the outputs, and expect Microsoft to iterate quickly: this is very much still the preview phase.