Excels Risky AI: Why It Fails You

Key insights

• =COPILOT() — A new experimental worksheet function that runs a generative AI model inside a cell to create text, classify data, summarize ranges, or build tables.
  It accepts natural-language prompts and optional cell ranges as context and returns an in-grid result rather than a chat response.
• Model & Behavior — Reports say the function uses a variant of OpenAI’s gpt-4.1-mini and can produce non-deterministic outputs that may change on recalculation.
  Microsoft warns it can give incorrect answers and says users should not rely on it for precise numerical or compliance-critical work.
• Scope & Data Limits — The function uses only the prompt text and any referenced ranges; it does not automatically pull data from other workbooks or external enterprise sources.
  This design limits what it can access but does not guarantee accuracy or reproducibility.
• Risks & Best Uses — Use it for drafting, summarizing, quick classification, and exploratory text tasks where exact reproducibility is not required.
  Avoid it for finance, legal, regulatory, or other high-stakes calculations because of accuracy limits.
• Controls & Practical Settings — Excel exposes Copilot settings where users can disable formula suggestions and adjust behavior; some features may require AutoSave or cloud-connected accounts to run.
  Turning off in-grid suggestions or using the Copilot pane can reduce unwanted AI prompts during routine work.
• Product Roadmap — Microsoft is testing more agent-like Copilot actions inside Excel but appears to be shifting toward side-pane experiences; reports indicate the experimental =COPILOT() function may be retired and users steered to the Copilot pane in the future.
  Expect ongoing changes as Microsoft balances convenience with safety and enterprise needs.

Overview of the video

In a recent YouTube video, Wyn Hopkins [MVP] examines Microsoft’s new in-grid AI feature for Excel and explains why it has sparked concern. He demonstrates the experimental =COPILOT() worksheet function and highlights how it can generate text, classify data, summarize ranges, and create tables directly inside cells. Moreover, he stresses that Microsoft warns the feature may return incorrect answers and should not be used for high-stakes numerical, legal, or compliance work. Consequently, the video frames the function as powerful but potentially risky for traditional spreadsheet uses.

Hopkins also walks viewers through the settings and practical toggles, showing how to turn off persistent suggestions and find Copilot controls within Excel. He emphasizes simple, hands-on steps and points out an AutoSave-related requirement that many users may overlook. As a result, the tutorial feels aimed at both power users and administrators who must balance productivity gains with governance needs. Overall, the video blends demo, caution, and practical advice.

How the =COPILOT() function works

First, Hopkins clarifies that =COPILOT() is not a traditional deterministic formula but an AI-driven worksheet function that accepts a prompt and optional cell ranges as context. He explains that the function relies on a generative model, reported to be gpt-4.1-mini, and returns text-based outputs into cells rather than numeric calculations. Furthermore, Microsoft states the function confines itself to the provided prompt and referenced ranges rather than searching the entire workbook or external enterprise data. Therefore, users should treat the outputs as exploratory rather than authoritative.

Hopkins shows how the function accepts natural language prompts and can work with a companion feature called Formula by Example, which helps construct formulas from demonstrations. He demonstrates examples where Copilot summarizes feedback rows or classifies categories, which are tasks suited to text outputs. At the same time, he notes recalculations can yield different answers, underscoring the non-deterministic nature of the feature. For this reason, Hopkins urges caution when using the function for repeatable reporting.

Benefits and practical use cases

Despite the risks, Hopkins acknowledges clear benefits: speed and convenience for text-oriented tasks such as summarizing customer feedback or creating draft content directly in cells. He points out that natural language prompts lower the barrier for users who otherwise would write complex formulas, and that embedding AI in the grid keeps workflows familiar. Consequently, analysts can prototype faster, automate routine tagging, and generate draft tables without leaving Excel. These capabilities can improve productivity, especially for exploratory work and drafting.

Furthermore, Hopkins highlights that the feature fits a specific niche—tasks where exact reproducibility and numeric precision are not essential. In such scenarios, the tool can reduce repetitive manual work and help users iterate rapidly. However, he also reminds viewers that using it for financial models or compliance reports would be inappropriate given the stated accuracy limits. Thus, the value lies in accelerating creative and text-heavy chores rather than replacing verified calculations.

Risks, limitations, and tradeoffs

Hopkins delves into the central tension: Excel users expect precision, while the AI returns probabilistic outputs. He warns that Microsoft itself advises against relying on the function for critical calculations, which creates a conflict between convenience and trust. Moreover, the feature’s tendency to vary results on recalculation undermines reproducibility, posing a challenge for auditing and regulatory scenarios. Therefore, organizations must weigh faster workflows against potential errors and governance gaps.

He also raises privacy and deployment concerns, noting that AutoSave and cloud-based processing can affect where data is sent and stored. Consequently, administrators need to check tenant-level settings, evaluate data residency rules, and decide whether to enable or restrict the function. In addition, usage limits—such as reported caps of about 100 functions per 10 minutes—mean heavy automation could hit throughput constraints. Overall, the tradeoffs involve balancing productivity gains with control, security, and auditability.

Practical guidance and final takeaways

Finally, Hopkins offers actionable tips: locate the Copilot settings to turn off formula suggestions, test Formula by Example in safe datasets, and avoid using the function for financial or compliance outputs. He suggests enabling the feature for exploratory teams while applying stricter controls for production work and recommends documenting any AI-generated results before relying on them. Furthermore, he notes Microsoft may retire the experimental =COPILOT() worksheet function in favor of the Copilot side pane, which could change how organizations adopt these capabilities.

In conclusion, Wyn Hopkins’ video provides a practical tour and a clear warning: the in-grid AI can boost productivity in the right scenarios, but it also introduces non-determinism and governance challenges. Consequently, users and IT leaders should test it carefully, set clear policies, and treat outputs as drafts unless validated. Ultimately, the video encourages measured adoption rather than blind trust, offering a helpful roadmap for teams that want to experiment safely.

Keywords

Excel AI feature risks,Excel AI hallucinations,Excel Copilot mistakes,Excel AI data privacy concerns,Excel AI errors and bugs,Excel hallucination examples,Excel Copilot reliability issues,Dangerous Excel AI features