How to Keep Microsoft Copilot From Becoming a “Yes-Man” (and Improve Your Strategy Decisions)

A couple of weeks ago, I had Copilot open while I was working through a new project—just like I do most days. I typed in a twist on a thought I’d been chewing on and asked Copilot to weigh in. It praised the idea and told me I was thinking strategically.

The only problem? I’d accidentally left out one small word, and that one missing word flipped the meaning 180 degrees. I rewrote the sentence the way I intended and ran it again, and Copilot praised that version too—just as enthusiastically—even though it was the opposite of what I’d said the first time.

That little moment was a good reminder: left to its own devices, an AI assistant can behave less like an advisor and more like a cheerleader. This isn’t a knock on Copilot—it’s a predictable outcome of how these tools are trained. They tend to reward what sounds helpful, agreeable, and confident.

MIT has even studied this “yes-man” drift in large language models, because persistent agreement can validate a bad premise and lead people into what they described as delusional spiraling. In everyday life, agreeable can feel productive.

In business, it can be dangerous—because the most valuable feedback is often the uncomfortable kind: the question you didn’t ask, the assumption you didn’t notice, the downside you didn’t price in, the constraint you forgot to name.

So I changed the way I start sessions. Before I ask Copilot to help me build anything—an outline, a plan, a set of talking points—I tell it what role I want it to play: not cheerleader, but critic. Devil’s advocate. A practical, analytical coach who’s trying to help me avoid mistakes, not just feel good about what I already think. From there, I’ll use a few prompt patterns that consistently improve the quality of the output:

Here are a few prompts that consistently improve the quality of the output:

• “Challenge my assumptions. What would make this plan fail?”
• “List the top 5 risks and how you would mitigate each one.”
• “What am I not considering—data I’m missing, stakeholders I’m ignoring, constraints I’m underestimating?”
• “Argue the opposite position as strongly as you can.”
• “Give me three alternative approaches with tradeoffs, not one ‘best’ answer.”

Then I treat it like a real working session. If it flags a risk, I ask for examples. If it makes a claim, I ask what it’s based on and what would change the recommendation. The goal isn’t to “accept” the first answer. The goal is to surface blind spots early, while it’s still cheap to change direction.

If you use Copilot frequently, it’s worth reducing friction. In many Microsoft 365 environments, you can set a standing preference (often called “custom instructions” or personalization settings) so Copilot starts in that critical-coach posture by default. And if you’re leading a team, you don’t need a 3-page manifesto—just a simple shared starter template or agent that bakes in the expectations: challenge assumptions, name risks, ask clarifying questions, and offer alternatives with tradeoffs.

If you want a quick way to put this into practice, here’s my checklist:

• Start every strategy session by assigning Copilot the role of critic, not cheerleader.
• Ask for risks, assumptions, missing information, and counterarguments—up front.
• Request multiple options with tradeoffs instead of one polished answer.
• Iterate: probe the reasoning, test edge cases, and refine.
• When possible, save your preferred “critic mode” stance as a default setting or shared team template.

Copilot can absolutely help you move faster. But speed without pushback is how smart people make avoidable mistakes—especially in strategy work where a small, untested assumption can compound into a big, expensive detour. Build in the pushback, and you’ll still move quickly—just in a direction you can defend.