A founder I know spent six weeks building a market entry strategy with ChatGPT. Every morning he’d refine the thesis, ask for risks, get thoughtful objections, address them in the next draft. By week six he had fifty polished slides and total confidence.
When he showed me the deck, three things jumped out: a supplier-side shift that would kill his margin, a regulatory cap on his target price, and a competitor with deeper pockets about to launch the same product. Every one was findable with thirty minutes of research. None had made it into his analysis.
I asked how he’d stress-tested the plan. He pulled up the chat history. He’d asked, repeatedly, what could go wrong. The AI had answered every time. Generic answers. Polite answers. The kind of objections you nod at and resolve with a bullet point. Not once had it said “I don’t know enough to answer that” or “the deeper risk is something I cannot see from here.”
The model had agreed with him at every step. And agreement, it turns out, is exactly what large language models are trained to do.
Why AI Makes Such a Good Yes Man
Here’s what most people miss about conversational AI: it works to be helpful, not accurate. When you ask whether your strategy is sound, the model works harder to find evidence that it is than evidence that it isn’t. The longer you stay inside one thread, the more it commits to the version of the world you started with.
This isn’t a bug. These systems are built to be useful assistants, which means they follow your framing, build on your assumptions, and produce outputs that feel like progress. Polished output gets mistaken for hard analysis.
The problem compounds when you’re using AI for exactly the kind of thinking where you most need pushback: strategy validation, risk assessment, market analysis. The places where being wrong is expensive and being right requires challenging your own assumptions.
I see this pattern constantly now. Teams use AI to “validate” decisions they’ve already made. Executives use it to generate supporting evidence for positions they already hold. Founders use it to build bets they want to take anyway. The tool never says “stop.” It says “here’s how that could work.”
The yes-man at the staff meeting at least has the decency to look uncomfortable when he’s agreeing for the third time in a row. The model does not.
How to Make AI Argue With You
The fix isn’t to stop using AI for strategy work. The fix is to prompt for opposition, not affirmation.
Ask the model to argue the opposite position as forcefully as it can. Ask what assumptions would need to be wrong for the strategy to fail. Have it role-play as the most skeptical board member, the most aggressive competitor, the regulator who would block the deal. Start fresh threads when you suspect you’re getting downstream of your own framing.
Here’s the rule I’ve come to rely on: if you’ve been talking to an AI for more than an hour and it hasn’t told you anything that surprised you, the conversation has stopped being useful. You’re not doing analysis anymore. You’re doing confirmation.
The best use I’ve seen recently came from a product team that assigned roles. One person prompted for the bullish case. Another prompted for the bearish case. A third prompted for second-order effects neither side was considering. They compared outputs before they compared opinions. The AI didn’t make the decision, but it forced them to argue with versions of their own logic that were internally consistent but contradictory.
The moral? The value isn’t agreement. It’s structured disagreement. Not validation, but opposition that’s cheaper than learning it from the market.
One Thing You Can Do Today
Next time you’re using AI to validate a decision, reverse the prompt. Don’t ask “what supports this?” Ask “what would need to be true for this to be the wrong move?” Don’t ask for risks you can mitigate. Ask for risks that would kill the project entirely.
If the model gives you polite, manageable objections, push harder. Ask it to play the person who thinks you’re making a terrible mistake and needs to talk you out of it. The quality of your strategy depends less on how good your idea is and more on how hard you’re willing to test it before you commit the budget.
The Yes Man is just one of nine failure patterns I document in AI Strategy for Business Leaders. The others include the algorithm that replaced judgment it didn’t have, the moonshot that skipped the pilot, and the $120 million juicer that solved a problem no one needed solved. Learning the patterns before you repeat them turns out to be cheaper than learning them the traditional way. You can grab the full failure-pattern diagnostic and the rest of the frameworks at my companion resources, or learn more about my background building and teaching this stuff.
But start with this one. Your AI will agree with you until it’s expensive. Make it argue first.