Microsoft's Copilot Data Reveals What People Actually Use AI For

Here's the thing nobody expected: when Microsoft actually looked at how people use Copilot, the top query category wasn't code or productivity hacks. It was health advice.

Yeah, health advice. 🏥

The IBM Technology team broke down Microsoft's Copilot usage report on their Mixture of Experts podcast, and honestly, the patterns are wilder than you'd think. Turns out we're not using AI the way the marketing departments thought we would. We're using it like we used to use Google—except now the stakes are way higher and someone's definitely paying for it.

The Health Advice Thing Makes Sense (Sort Of)

Kush Varshney, IBM Fellow, wasn't surprised at all. "That's what I've been using it for, too," he said on the podcast. Which tracks—if you've got medical test results and you want a second opinion before your doctor's appointment, why wouldn't you paste them into an LLM?

But here's where it gets interesting: the usage patterns follow human rhythms in ways that feel almost eerie. People ask for philosophy advice at night. They flip from gaming queries on weekends to coding questions during the week. Valentine's Day queries spike before February 14th. The AI isn't just answering our questions—it's basically becoming a mirror of our entire lives.

"It's like synchronizing with our lives or we're synchronizing our lives with it," Varshney noted. Which, yeah, that's the question, isn't it?

Your Company Is Paying For Your Gaming Tips (Probably)

Lauren McHugh Olende, program director for AI open innovation at IBM, pointed out something that's going to become a real problem: if people are using Copilot for work during the week and gaming advice on weekends, and they're on an enterprise license... that means companies are footing the bill for all of it.

"Google is free and I have my own Google account," she explained. "But here it's definitely not free. It's not free for me to use and then it's not free to like actually produce these things in the back end."

The token costs add up. Fast. And unlike Google search, which costs fractions of a penny per query, these AI systems are burning actual money every time you ask them to explain your Elden Ring build or help you plan your D&D campaign.

Olende even joked about a post she saw suggesting people should negotiate their token budget when they get a job offer. But honestly? That might not be a joke for much longer. Either companies are going to get a lot more restrictive about AI usage, or we're headed toward a "bring your own AI plan" world.

This Isn't Google 2.0—It's Something Weirder

The natural comparison is to search engines. We've seen usage reports like this before, right? People searching for health stuff, relationship advice, hobby tips—this is just what humans do online.

But Olende pushed back on that: "With Copilot, I mean, there's definitely sensitivity that way. You're uploading things, you're generating things, you're much more precise. You're not just kind of putting topics in, but you're putting like full questions in."

You're not typing "best laptop under $800" into Copilot. You're uploading your actual work documents, your actual code, your actual medical test results. The data sensitivity is on a completely different level.

And yet—weirdly—there hasn't been a huge privacy backlash to Microsoft publishing this usage data. Kush Varshney thinks the norms might actually be shifting. "Kids these days are like more than happy to be like tracked," he said, pointing to location sharing with friends and parents. Maybe we're all just getting more comfortable with companies knowing everything about us?

His take: "I'm not there yet." Same, honestly.

The Ralph Wiggum Strategy: Let The AI Figure It Out

The second half of the podcast dove into something called the "Ralph Wiggum strategy" for coding agents, and this is where things get genuinely fascinating for anyone who's tried to use AI for actual work.

The traditional approach to AI coding assistants is to be very specific. Tell it exactly what you want, review every change, micromanage the process. It's exhausting and kind of defeats the point.

The Ralph Wiggum strategy? Give the AI simple instructions and just... let it loop until it figures it out. Don't babysit. Don't check in constantly. Just say "make this work" and walk away.

Olende pointed out the obvious problem: "You have to give it much better instructions. It's kind of just frontloading the work, right?" You're trading constant supervision for better initial prompts. And honestly? She's started to just do things herself because writing the perfect prompt takes longer than doing the actual task.

But Volkmar Uhlig, CTO and VP of data platforms at IBM, had a different angle. He compared managing AI agents to managing junior engineers—or kids. You have to be more patient, more explicit. "We throw something over the fence. We are not describing what we really want. There's a lot of mind reading going on," he said.

AI can't read your mind (yet). So it just shows you how bad we all are at giving clear instructions. Which means either we get better at communicating, or the AI gets better at mind-reading. One of those seems way more likely than the other.

What This Actually Means For Normal People

Here's what I keep coming back to: Microsoft has detailed usage data showing that people are using expensive enterprise AI licenses for personal health advice, gaming tips, and Valentine's Day card ideas. And... nobody seems to care?

That's the real story. Not that AI usage patterns mirror human psychology (of course they do). Not that we're bad at giving instructions (we've always been bad at that). But that we've collectively decided this is fine. The privacy concerns, the cost concerns, the work-versus-personal-use concerns—they're all out there, floating around, acknowledged by everyone on this podcast.

And yet adoption keeps climbing. The Super Bowl had AI ads. Enterprise contracts keep getting signed. We're all just... moving forward.

Maybe that's the actual pattern worth watching. Not what we're using AI for, but how quickly we're all getting comfortable with it knowing everything about us—as long as it gives us health advice at 2 AM when we're spiraling about that weird rash.

—Tyler Nakamura