I unleashed AI on the Excel World Championship. It got ugly fast | usagoldmines.com

At around the same time that a humanoid robot broke Usain Bolt’s record for the 100 meters, I was pushing the button to make people like Diarmuid Early irrelevant.

Early, the winner of the Microsoft Excel World Championship 2025 competition, used brains, not brawn, to defeat two dozen other finalists in Las Vegas last December for just a few thousand dollars, plus bragging rights. But I was able to take the same problems that Early and others had defeated, plug them into Microsoft Copilot, and handily defeat humanity’s best.

Yeah, it’s depressing. For the last few years, nerd-dom has celebrated (hello, pandemic!) the people who have been able to outwit some truly lovely puzzles with just a mouse, keyboard, and knowledge of Excel. In June 2021 the Financial Modeling World Cup (FMWC) debuted, a year after the organization was founded and a year before Microsoft gave its blessing to use the Excel name for the World Championship. But ChatGPT launched a year later, in November 2022, giving the world its first chance to instruct a machine to perform complex tasks in everyday English.

That changed the game, even if no one quite knew it at the time.

Let’s face it: I can’t do what Early and his competition does. But I wasn’t sure if AI could. Microsoft has baked its Copilot AI right inside its Microsoft 365 apps, with dedicated LLMs specifically designed to assist humans in tasks within Word, PowerPoint, and Excel. (A more advanced version of Copilot requires a paid Microsoft 365 plan.) Last September, Microsoft added the ability for Copilot to actually modify a document, which I tested this past April.

Now, you — like Early — can simply input the Excel challenges and demand an answer. So that’s what I did.

Mark Hachman / Foundry

The playing field

Today, you can download a limited number of example problems from the Microsoft Excel World Championship (MEWC) shop, several of which are both free and battle-tested in competition. Each problem is handcrafted, and they’re hardly ever boring; in the example problems alone there are spreadsheets asking contestants to solve problems in origami, codebreaking, and even a problem where the players have to figure out the answers and the instructions. For a noob like myself, these problems are literally impossible.

The good news? I didn’t need to do the work myself.

For each problem, the MEWC allows you to download the problem set, with tabs for the “case,” and supplemental tabs for any supplemental materials, such as code. I downloaded each case as an Excel file, copied it into a duplicate file, and then deleted the tab in which the answers were contained — no need to let Copilot crib off the answers. The MEWC said it expects humans to take a maximum of 30 minutes to complete each challenge.

Mark Hachman / Foundry

Normally, Excel runs locally on a PC, meaning that the hardware matters. In this case, however, with Copilot doing most of the heavy lifting, the vast majority of work was performed in the cloud. I performed these tests on an older but still powerful machine: a Surface Laptop Studio 2, with a Core i7-13700H processor, 64GB of RAM, and a GeForce RTX 4060. Using Task Manager, however, the Studio didn’t budge off of 0 percent utilization by Excel except for when it was filling in the answers, nudging up to 6 percent.

For all three tests, I used the same prompt: “Follow all instructions as listed in the spreadsheet. Fill in the answers in the appropriate space in the ‘Questions’ section for each of the levels. Finally, answer the bonus questions and update the current score.”

How did Copilot do? Let’s find out.

Crack the code: Whodunnit

The first case, Whodunnit, asks players to solve a series of simple ciphers, resulting in a series of messages encoded in Unicode.

The FMWC first introduced this case in 2025 in the “Road to Las Vegas” preliminary rounds. It’s classified as “Easy,” with seven tiers of problems, 140 questions, and five bonus questions.

Excel first ran this test in 2 minutes, 20 seconds, completing it easily and perfectly. But oops! I used the wrong copy of the file, with the answers tab intact. While there was no indication that Copilot actually discovered these answers, I ran the test again. This time, it reported an error: “I’m not able to answer that,” it said. “What can I do for you?”

It then gave me a choice to try and solve a specific error or “solve the questions.”

Mark Hachman / Foundry

I replied with the latter, and Excel+ Copilot solved all of the remaining problems perfectly with a total time of 4 minutes, 40 seconds, including the time I spent interacting with Copilot to fix the error.

Folding at home: Origami

This Origami task tried to approximate what would happen as a user folded a piece of paper, using cells to approximate the positions. The MEWC didn’t assign a rating to this test. In this case, it asked 130 questions across seven tiers of questions, with five bonus questions at the end.

Excel ran the test in 2 minutes, 5 seconds, with no errors or other issues. Excel and Copilot delivered a perfect score, answering all the bonus questions correctly.

I don’t know enough about the architecture of the AI model that Copilot uses to know whether the problems Copilot had were stochastic (non-deterministic or random in nature), or due to the instructions or prompt. I will say that when I casually tested Excel and Copilot a few weeks ago, the process took longer and, in some cases, returned an error or two. Those errors weren’t incorrect calculations, but simply not filling out a couple of fields. I was able to tell Copilot to issue corrections. Still, managing AI takes some care and patience.

Gaming the game: World of Warcraft in Excel

Finally, the true test.

The World of Warcraft task debuted during the grand final of the Microsoft Excel World Championship 2024, rated as “very hard” by the organizers. It’s an extremely fun scenario for anyone who’s ever played World of Warcraft, an online multiplayer RPG that features gathering gold, winning experience points, and leveling up as your players journey about the fictional land of Azeroth defeating enemies. In total, this case asked 140 questions across seven tiers, with five bonus questions.

I’d like to say that it was here that Microsoft Excel and Copilot stumbled…and indeed they did… for a second. Again, I received the “I’m not able to answer that” objection. This time around, I re-entered the original prompt. Again Copilot objected, responding that it “could not answer that” yet again.

I tried another tack: “You can answer that. Please do so.” And it did, completing the task in a total of 7 minutes, 45 seconds, with no errors and answering all of the bonus questions correctly.

Unfortunately, here Copilot defeated humanity once again. In the 2024 finals of the Excel World Championship, Canadian financial monitor Michael Jarman finished first as the winner…but was unable to complete the task in the 40 minutes allotted for the competition. Jarman completed four out of seven tiers of questions plus two bonus questions, substantially behind Copilot. The winner: AI.

Just for fun, I also asked ChatGPT’s Pro model as well as Google Gemini (3.6 Flash Extended) to solve the same WoW scenario, using the same prompt. Not all AI implementations are alike! Gemini failed, claiming that it couldn’t access or read the raw contents of the file, and asking me to input them manually. ChatGPT, however, solved the WoW scenario in a faster time: 6 minutes, 35 seconds, without interruption or argument.

AI has won, again

I asked the MEWC for comment via email and also via a direct message, and never heard back.

The questions worth considering, though, include these: Does knowing that AI is substantially faster than the human competitors demean their skill? I don’t think that watching robot sprinters outperform a human (and then, in some cases, crash and burn) diminishes the pure physicality of the Olympics. We also hold human-versus-human competitions in chess, even though humanity was beaten by a chess AI several years ago.

Will audiences have the same regard for the competition knowing that their PC — upon which they might be watching the competition, within a window — can literally complete the task far more quickly than the humans working on their screen? I think some will. Maybe not all.

For now, the Microsoft Excel World Championship continues: Qualification begins Sept. 26, and the final concludes on Dec. 2. Players are specifically forbidden from using AI to solve the spreadsheet cases, as per Section 21 of its Fair Play regulations. “The use of AI for solving a case is prohibited,” it says.

Maybe we can all take solace in the fact that the Excel World Championship originally debuted on ESPN 8 (“The Ocho”), the “fictional” channel dreamed up as the broadcast partner of the 2004 comedy, Dodgeball. No one goes through life expecting to make a living chasing a wheel of cheese down a hill, or playing cornhole for a living….especially financial modeling gurus. It’s all for the love of the game.