Microsoft is retiring the preview COPILOT() worksheet function in Excel on September 14, 2026, ending an experiment that put generative AI directly inside the spreadsheet calculation grid before the feature reached a full public launch.1
The decision points to a more constrained vision for AI in Excel: Copilot as an assistant that helps users analyze, summarize and transform data through a side pane or contextual controls, rather than as a callable function that behaves like SUM(), XLOOKUP() or IF(). TechRadar, citing a Microsoft 365 Message Center notice, reported that Microsoft is directing users to the Copilot side pane and the floating Dynamic Action Button as alternatives.1
For power users, the withdrawal matters less because a preview feature is disappearing than because it exposes a core tension in AI-enabled productivity software. Spreadsheets are built around repeatable calculations, visible dependencies and audit trails. Generative AI systems, by contrast, can produce probabilistic outputs, may change behavior as models are updated and often require judgment about whether an answer is acceptable.
Putting that kind of system inside a worksheet formula raises harder questions than placing it in an assistant panel.
Why the grid is different
Excel formulas carry a strong user expectation: the same inputs should reliably produce the same outputs, or at least outputs governed by clearly understood rules. A formula embedded in a cell becomes part of the workbook’s logic. It may feed charts, financial models, reports, forecasts or downstream calculations.
That makes an in-cell AI function very different from a chat prompt. If a Copilot side-pane response is imperfect, the user can treat it as a suggestion. If a COPILOT() formula sits inside a live model, its output may be copied, referenced, recalculated and shared as if it were deterministic spreadsheet logic.
TechRadar reported that the preview function allowed users to describe tasks in natural language and receive results inside cells, including as part of a formula.1 That is powerful for classification, summarization and content generation. But it also blurs the boundary between calculation and interpretation.
Reproducibility becomes a product problem
The first challenge is reproducibility. Spreadsheet users often expect that reopening a workbook next month will preserve its logic and allow another analyst to trace how results were produced. A generative AI formula complicates that expectation.
If the model behind a function changes, should old COPILOT() results remain frozen, or should they update? If they update, a workbook’s conclusions could shift without any traditional formula change. If they stay frozen, users may not understand why one AI-generated output reflects an older model while another reflects a newer one.
That problem is especially acute in shared workbooks. A sales operations analyst, finance manager and compliance reviewer may open the same file at different times. If AI outputs depend on model version, prompt interpretation, connected data or external knowledge, the workbook can become harder to verify than a conventional Excel model.
Auditability is not optional for power users
Excel is often used as an informal database, a reporting engine and a financial modeling environment. In those contexts, auditability is not a nice-to-have feature. Users need to know what changed, why it changed and whether the result can be defended.
A natural-language prompt embedded in a cell may not provide enough information. A formula such as =COPILOT(A2:A100, "summarize customer objections") might produce a plausible summary, but the reviewer still needs to know which rows mattered, what criteria were applied and whether any relevant detail was omitted.
Microsoft reportedly warned that the function should not be used for numerical calculations and advised against using it for certain legal and compliance-related content.1 Those cautions underline the mismatch: Excel’s grid encourages users to operationalize formulas, while generative AI output may require review before it becomes part of a business record.
Recalculation creates another layer of risk
Excel’s recalculation model is one of its strengths. Change an input and dependent formulas update. That is exactly what makes an AI formula risky.
For conventional formulas, recalculation is usually predictable. For AI formulas, recalculation may mean a fresh model call, a slightly different phrasing, a new categorization or a revised summary. Even small changes can matter if the output is used to group customers, flag risks or generate text for external communication.
The product design question is difficult: should an AI cell behave like a volatile function, a cached value, a refreshable query or a generated artifact requiring approval? Each answer creates trade-offs. Volatile behavior may surprise users. Caching may make workbooks stale. Manual refresh may undermine the convenience of a formula.
The side pane is a safer home for ambiguity
Microsoft’s apparent pivot toward the Copilot side pane and contextual buttons suggests a clearer division of labor. The spreadsheet grid remains the place for structured data, formulas and inspectable outputs. Copilot becomes a guided assistant that can propose transformations, explain patterns or generate drafts without necessarily entering the workbook’s calculation chain.
That approach fits the way many power users already work. They may want AI to suggest a formula, summarize a table, draft a Python or Office Script workflow, create a pivot table or explain anomalies. But they may not want a language model sitting silently in the middle of a dependent calculation path.
The distinction is subtle but important. An assistant can be wrong in a way that invites review. A formula can be wrong in a way that propagates.
A narrowing, not a retreat from AI in Excel
The retirement of COPILOT() does not mean Microsoft is backing away from AI in Excel. TechRadar noted that Microsoft is continuing to develop AI capabilities for Excel, including features aimed at finance workers and workflows such as modeling, forecasting and reporting.1
The more likely interpretation is that Microsoft is narrowing where generative AI belongs. AI can still help users understand data, automate repetitive preparation, generate explanations and accelerate analysis. But embedding open-ended model calls directly in worksheet cells may have created too many unresolved questions for a product whose value rests on trust, consistency and traceability.
For spreadsheet power users, the message is practical: expect Copilot to become more present around the workbook, but perhaps less like a native formula language feature. The future of AI in Excel may be less about typing =COPILOT() into a cell and more about asking Copilot to help build, explain and refine the workbook while leaving the grid’s core logic more auditable.
Comments