Microsoft Copilot just failed a major exam for spreadsheet AI. In a hands-on evaluation that pitted it against three other AI tools, Copilot—the assistant built directly into Excel—came in dead last, stumbling through tasks that professionals consider table stakes for financial modeling and data analysis. The test, conducted by YouTube creator Kenji Explains and reported by Geeky Gadgets, shows that being baked into Microsoft 365 isn’t enough to make an AI useful for the kind of complex workbook tasks that actually fill an accountant’s day.
The loser in the test isn’t the AI itself—it’s the user who trusted it with a critical spreadsheet.
What Actually Happened
The test put four AI assistants through five real-world Excel scenarios: extracting a balance sheet from a 92-page PDF, comparing two similar workbooks for differences, building a dynamic business model with live scenarios, detecting errors in a multi-tab financial model, and shaping a messy dataset into a clean pivot table analysis. Each task demanded more than a formula suggestion; they required the AI to understand context, produce accurate formatting, and deliver outputs that a human could review and use immediately.
Tracelight—a specialized AI built for finance and spreadsheet work—won the challenge, delivering accurate, well-structured results across all five tasks. ChatGPT showed speed and often solid reasoning, but required manual cleanup. Anthropic’s Claude was methodical and clear, but slow and less adaptable. Copilot, meanwhile, couldn’t complete four of the five tasks without heavy user intervention. When it did produce something usable, like a scenario model, the formatting was so messy as to be almost worthless in a client meeting.
Here’s how the tools stacked up in a nutshell:
| Scenario | Tracelight | ChatGPT | Claude | Microsoft Copilot |
|---|---|---|---|---|
| Extract balance sheet from PDF | Clean, accurate ratios and layout | Fast but minor errors | Accurate but slow; formatting issues | Failed to produce formulas; needed manual rebuild |
| Compare Excel files | Side‑by‑side comparison with clear summaries | Useful insights but no side‑by‑side view | Somewhat helpful, no dedicated comparison | Incomplete, unclear results |
| Build scenario model | Accurate, customizable, fully functional | Fast but incomplete; required manual finishing | Polished but limited customization | Functional but poorly formatted and unclear |
| Error detection in financial model | Detailed error report, user‑guided fixes | Fast and accurate but auto‑edited model without consent | Found errors but interface cluttered | Couldn’t complete the task |
| Data manipulation (unpivot, pivot) | Accurate transformation, clean pivot and slicers | Issues with formatting and completeness | Similar issues, needed extra work | Errors; couldn’t finish the analysis |
The takeaway isn’t just that one tool beat the others. It’s that for advanced spreadsheet work, the most familiar and most integrated tool—the one Microsoft is selling as a reason to upgrade to Microsoft 365 Copilot subscriptions—is the least reliable option in the field.
What It Means for You
If you’re a home user who occasionally asks Copilot to sum a column or suggest a VLOOKUP, this test probably doesn’t break your workflow. Copilot’s basic features still work for simple tasks. But if you rely on Excel for anything that touches financial decisions, operational reports, or client-facing models, the risks suddenly look concrete.
For power users, analysts, and finance teams, the lesson is clear: Copilot cannot yet be trusted to handle complex, multi-step spreadsheet jobs without constant supervision. An AI that silently alters formulas during error detection—as ChatGPT did—or one that simply fails to produce a usable comparison output—as Copilot did—creates more work than it saves. In the best case, you waste time fixing sloppy output. In the worst, you base a business decision on a faulty analysis.
The test also exposes a hidden tax. When a tool produces output that looks right but contains small, hard-to-spot errors, the human reviewer has to double-check everything. That review effort can erase any time savings the AI should have provided. For professionals, the ideal AI assistant is one that consistently produces “review-ready” work. Tracelight came closest. ChatGPT and Claude offered speed or elegance, but still demanded vigilance. Copilot demanded a rewrite.
How We Got Here
Microsoft launched Copilot for Microsoft 365 in 2023 with a promise to bring generative AI into the flow of work. In Word, PowerPoint, and Outlook, the assistant has shown real value for summarization, drafting, and quick edits. But Excel has always been a different beast. Spreadsheet logic, model architecture, and data integrity are unforgiving. A single wrong formula can cascade into millions of dollars in misallocated budgets or mispriced contracts.
For the first year, Copilot in Excel was limited to basic operations: summarizing data, asking natural-language questions, and proposing simple formulas. It was never designed for advanced financial modeling. Microsoft gradually added capabilities—Python integration, for instance—but the core AI still seems to struggle when asked to think like an analyst rather than a bot.
Meanwhile, specialized tools like Tracelight have emerged, purpose-built for the workflows that make Excel the backbone of global finance. They don’t try to be general assistants; they train on the specific patterns of financial statements, workbook comparisons, and data transformation. This focused approach is paying off.
ChatGPT and Claude, while generalist, benefit from massive training data and rapid iteration. Their models can reason about spreadsheet logic because they’ve swallowed the internet’s worth of Excel tutorials, forums, and documentation. But they lack the deep integration and dedicated spreadsheet features that a tool like Tracelight built from the ground up.
Copilot sits in an awkward middle. It has deep integration into Excel’s object model—it can navigate workbooks, insert sheets, and run formulas—but its underlying language model appears less capable of the kind of multi-step reasoning that a specialized tool or a larger general model can muster. The result: it knows Excel’s plumbing but can’t solve the hard problems.
What to Do Now
If advanced Excel is a daily part of your job, you need to pick an AI companion based on your specific workflow, not on which logo is already built into the ribbon. Here’s a practical guide from the test data:
- For strict financial modeling and auditing (balance sheets, scenario analysis, workbook comparison): Give Tracelight a hard look. Its outputs are the most complete and trustworthy of the batch, and it saved testers significant cleanup time. It isn’t free, but for a full-time analyst, the subscription cost can be a rounding error compared to the time saved.
- For rapid idea generation and formula help where you’re willing to verify everything: ChatGPT is a strong choice. It’s fast, conversational, and often gets the job done if you’re an experienced Excel user who can spot mistakes. Pay close attention to any changes it suggests to your file—never let it auto-edit a financial model.
- For presentations and explanatory models where clarity matters as much as accuracy: Claude produces beautifully formatted outputs and explains its reasoning well. It works best when you need to communicate results to a non-technical audience, but it may slow you down if you need to iterate quickly.
- For basic Office assistance: Copilot in Excel can still help with simple tasks like summarizing a table, creating basic charts, or generating boilerplate formulas. But for anything that goes beyond entry-level use, hold off until Microsoft proves it can handle the heavy lifting.
In general, adopt a “trust but verify” policy. Treat any AI-generated spreadsheet as a first draft, not a finished product. Audit formulas, spot-check numbers, and always test scenario logic before presenting to a client or boss. This practice applies even to Tracelight, the top performer. No AI is infallible.
Outlook
Microsoft isn’t standing still. Insiders tell us that future updates to Copilot will incorporate more specialized models and deeper reasoning capabilities. The company also recently invested in spreadsheet-specific AI research and is expected to bring OpenAI’s latest models to Office apps. By late 2026, Copilot’s Excel skills could look very different.
For now, though, this test serves as a reality check. The spreadsheet on your screen deserves an assistant that respects its complexity. Copilot, the built-in helper, isn’t that assistant yet. Choose your AI the way you choose a formula: based on what actually works, not what’s easiest to click.