Excel turns AI into a spreadsheet formula with =COPILOT

For the past two years, using AI in Excel mostly meant awkward sidebar chatting: copy a block of data, ask the assistant to make sense of it, paste the text back, and pray nothing broke when source numbers changed.

Microsoft is now pulling that intelligence straight into the grid. The Excel team introduced the native =COPILOT function for Windows and Mac, turning large language models into standard spreadsheet formulas.

Formulas that read and write plain English

The syntax behaves just like any traditional formula: =COPILOT(prompt, [context]). You hand it an instruction in plain language and point it to a cell or a range.

If you have a column of messy coffee maker reviews in cells D4 through D18, typing =COPILOT("Classify this feedback", D4:D18) drops sentiment and category tags straight into your sheet. It can also generate product descriptions based on spec sheets, summarize lengthy footnotes into reporting snippets, or brainstorm test tables that spill across multiple rows and columns.

Because it lives inside Excel's calculation engine, it updates dynamically. When underlying cell values change, Copilot recalculates without requiring users to rerun Python scripts, trigger macros, or refresh add-ins.

Users can also nest it inside functions like IF, SWITCH, WRAPROWS, or LAMBDA, piping calculated outputs straight into the AI prompt.

Rate limits and reality checks

According to The Verge, the function is powered by OpenAI's gpt-4.1-mini and acts as the official successor to Microsoft's experimental LABS.GENERATIVEAI function from 2023. The Decoder notes that Microsoft plans to bring larger next-gen models to the tool down the road.

For now, availability is tightly gated. It is restricted to Microsoft 365 Copilot license holders on the Beta Channel, running Windows version 2509 or Mac version 16.101 and above. A web rollout via the Frontier program is slated for later.

There are also hard usage ceilings: users can run 100 calls every 10 minutes, capped at 300 per hour. Microsoft recommends passing entire ranges rather than writing individual formulas cell by cell, since one range call only burns a single request from the quota.

On the limitation front, the function cannot access live web data or internal corporate files yet. All external context must be manually imported into the workbook first. Results are returned strictly as plain text, and Microsoft explicitly warns against using the function for precise mathematical calculations or high-stakes compliance workflows where hallucinations could cause real harm.

Microsoft confirmed that data processed by the formula remains private and will not be used to train AI models.

Why it matters

Moving an LLM from a sidebar chatbot into an actual formula engine feels like the right architectural shift for spreadsheet productivity.

Previously, running batch semantic tasks like customer ticket tagging or copy variations required API scripts or brittle third-party plugins. Now, a marketing manager or ops lead can perform lightweight natural language processing within familiar spreadsheet syntax. Given the current 100-call limit per ten minutes, it won't handle enterprise-scale datasets overnight, but it is immediately useful for ad-hoc qualitative analysis.

Competitively, Microsoft had to answer Google Sheets, which rolled out similar in-cell AI capabilities back in June. By baking Copilot directly into formula dependencies, Microsoft maintains its grip on power users who live inside complex Office workbooks.

Yet, making cells probabilistic introduces fundamental tension. Spreadsheets earned trust across finance and operations because formulas are deterministic—same inputs always yield identical outputs. When a formula's output can drift or hallucinate between recalculations, auditing models becomes significantly harder. We think teams will welcome the speed for rough text processing, but keeping unreliable generative text from corrupting mission-critical models will require serious user discipline.