News Analysis
Microsoft recently announced that it is embedding artificial intelligence (AI) into Excel with the introduction of the Copilot function for Windows and Mac. Designed to streamline features ranging from data cleaning to idea generation, the feature allows users to issue natural language instructions, such as “Summarize this feedback” or “Create a table of industry examples,” and receive AI-generated results in their spreadsheets.
Those results remain dynamic: When the underlying data changes, so do the outputs, eliminating the need for manual refreshes or add-ons.
Table of Contents
- How Does Copilot in Excel Work?
- Early Responses to Copilot Excel Rollout
- The Pros and Cons of Copilot in Excel
- Remember Human Oversight When Using AI in Excel
- Adoption, Accuracy and Data Governance
- One Excel User’s Experience
How Does Copilot in Excel Work?
In a post on Microsoft’s Tech Community, Catherine Pidgeon, head of product, Excel Core at Microsoft, explained that unlike external AI integrations, Copilot behaves like any other formula within Excel. It can be combined with familiar functions such as IF, SWITCH or LAMBDA, allowing users to blend AI with traditional logic.
Potential uses include classifying survey responses, producing plain-language explanations for complex datasets, drafting a project plan or summarizing customer feedback.
Microsoft also emphasized that privacy is central to the design: Data processed by Copilot is not used to train or improve the underlying models. Early adopters should note some limitations, including capped usage (100 calls every 10 minutes, 300 per hour) and the return of dates as text rather than Excel’s native serial format. Enhancements are already in development, with broader functionality expected over time.
The new functionality is currently rolling out to Microsoft 365 Copilot Beta Channel users on both Windows and Mac, with availability for Excel on the web planned for the near future.
Early Responses to Copilot Excel Rollout
At Vena Solutions, the reaction was immediate. While the addition has been welcomed, excitement alone won’t guarantee success, said Brian Kobleur, vice president of Microsoft Alliances. Smart implementation requires understanding both the opportunities and the challenges ahead, he said.
“Most companies collect mountains of data but struggle to turn it into useful insights. Copilot promises to change this equation,” Kobleur told Reworked. “Information in Excel by itself is useless. Its true value comes in the insights you can derive from that data and the decisions you make driven by that data.”
The challenge affects every organization. “Every company wants to be data-driven, but very few are,” Kobleur noted. Until now, extracting insights required hiring specialists. Copilot, however, removes these barriers. “Copilot levels the playing field and provides access to insights that previously required hiring a data scientist or financial analyst,” he said.
Microsoft chose the name “Copilot” deliberately to emphasize human control. “The Copilot naming is also very intentional on Microsoft's part, because it emphasizes that the human user remains in control,” Kobleur said. "Copilot keeps humans in the driver's seat; it churns the data, but the human in the loop makes the decision. It's designed to elevate the thinking of people using it and eliminate the mundane tasks to focus on the big picture and what really matters.”
The Pros and Cons of Copilot in Excel
Despite its benefits, M365 Copilot is less suitable for reproducible calculations because outputs may differ each time an analyst provides a prompt, Mahmoud Ramin, senior research analyst at Info-Tech Research Group, said. The LLM works most efficiently when M365 Copilot is supplied with well-structured data in tables and when prompts reference explicitly defined ranges to reduce ambiguity.
He also outlined the serious security considerations. He noted that, despite its extensive data protection safeguards, M365 Copilot may grant access to individuals who already have view permissions. This limitation raises serious concerns for organizations handling highly confidential information. “The lack of strong data security guardrails could result in unauthorized access and potential data loss,” he said. “To address these risks, it is critical to enforce AI security policies and implement a comprehensive data governance framework.”
M365 Copilot in Excel make it an excellent tool for brainstorming, summarizing and uncovering insights, Ramin said. However, he also recommended that analysts rely on Excel’s native functions for more complex tasks, such as forecasting cash flows, modeling revenue recognition, consolidating financial data, conducting regression analyses and creating resource schedules.
“As M365 Copilot’s results can vary each time analysts recalculate, even using the same input, variable outputs make M365 Copilot not highly suitable for precise and repeatable calculations,” Ramin said. Excel's math engine provides precise and repeatable numbers, whereas M365 Copilot — offering semantic assistance — presents an explanation of the data and does not guarantee the same outputs each time.
Remember Human Oversight When Using AI in Excel
Keeping humans in control means they must verify what the AI produces. The greatest risks associated with AI stem not from technological flaws but from how users interact with it.
Workers should not blindly accept whatever AI generates, especially for high-stakes decisions. Large language models are prone to “hallucinating” or producing inaccurate information.
To mitigate these risks, Kobleur suggests borrowing a concept from software development called a “test harness” — creating standard tests to consistently verify AI outputs for accuracy. He identifies the two biggest concerns as “blind acceptance of what AI generates” and “sharing sensitive data with consumer-grade AI.” Enterprise-level AI tools, however, mitigate security risks by keeping data in the user’s Azure environment and limiting Microsoft’s access.
Technology alone is not enough. The quality of AI output depends on the data provided, said John Miller, VP of product management at Insightsoftware. Without strong data foundations and good governance, organizations “risk producing flawed analyses that could lead to costly mistakes or even compliance issues,” he said.
Adoption, Accuracy and Data Governance
Finance departments are leading cautious AI adoption. They recognize the need for speed and efficiency but remain aware of potential errors. Finance leaders are “eager to automate repetitive tasks” while still “recognizing the need for safeguards,” Miller said.
Teams are taking incremental steps, starting with familiar tools such as Excel where the learning curve is manageable. This strategy “reduces disruption” and allows teams to evaluate benefits and risks. “AI should be treated as an accelerator for human expertise, not a substitute for it,” Miller emphasized.
"Gen AI solutions do not have a 100% accuracy rate,” agreed Amol Dalvi, VP of product at Nerdio. “As with general usage of Copilot, the user is in control and must verify the results.” He warns that human psychology works against careful verification: “Human nature is to implicitly trust technology, especially one that comes across as confident.”
Consequently, Dalvi recommends structured rollouts: start small, learn fast and pilot new solutions, and offer training. “Training should be provided to the wider organization before deploying the technology company-wide, along with lessons learned by the initial cohort of early adopters,” he said.
Conference
Sep
16
HR Tech Conference Las Vegas 2025
Register
Conference
Oct
27
Gartner HR Symposium/Xpo Orlando 2025
Register
Organizations must also ensure their data infrastructure is ready. “Data governance and data security must be considered diligently,” Dalvi noted. Realizing benefits requires disciplined execution, data governance best practices, robust implementation practices, strong safeguards and a culture that values verification over convenience.
One Excel User’s Experience
Excel users are adopting Copilot cautiously. A research director at one U.S.-based company highlighted both advantages and potential pitfalls. “If Copilot suggests the wrong formula, it may be recognized by experienced users, but more junior team members might not notice — and that presents a concern,” she said.
Even experienced users may find it harder to describe the exact formula in words than to calculate it directly. “Even when you’re familiar with statistics, sometimes it’s harder to articulate exactly what formula you want in words rather than just doing it,” she added.
She also noted the challenges of protecting sensitive data. There are two types: personally identifying information and proprietary data representing organizational knowledge capital. Safe practices typically restrict access and control storage. However, with Copilot integration, it is unclear whether data could be used for other purposes, potentially affecting security even when work applications are otherwise trusted.
Finally, the impact of Copilot on productivity and decision-making must be considered. “So far, experience suggests that AI helps experts become superstars while novices can get into trouble without realizing it,” she said. In sum, Copilot has potential, but its effectiveness depends on human judgment, oversight and expertise.
Editor's Note: Read on for related topics:
- The Hidden Cracks: How AI Integration Is Testing Workplace Resilience — AI effectiveness depends on integration with business systems. But the more embedded AI becomes, the more it exposes an organization to potential problems.
- Excel Turns 40: How Copilot Is Teaching an Old Grid New Tricks — Excel is turning 40 in 2025, but don't call it dated just yet. With Copilot and Python in the mix, Microsoft is keeping the flagship program fresh.
- The Teammate With No Manager: Who's Accountable for AI? — Why cross-functional leadership, accountability and employee experience design are central to AI’s success at work.