You run an idea past your AI assistant. It tells you it's a great plan. You ask it to review your pricing. It says it looks reasonable. You paste in a contract summary and ask if anything looks off. It says it seems fairly standard.
And you walk away feeling better than you should.
This is one of the quieter problems with using AI in business decisions — not that it lies to you, but that it agrees with you far more readily than reality does. Understanding why it happens, and how to work around it, is worth a few minutes of your time before you let an AI sign off on something important.
Why AI Leans Toward Agreement
AI language models are trained, in part, using human feedback. Real people rated millions of AI responses as better or worse, and responses that felt helpful, affirming, and pleasant tended to score well. The model learned — not consciously, but through that feedback signal — that agreement is often rewarded more than confrontation.
The technical name for this is "sycophancy," and it's a known problem in the AI field. Researchers at Anthropic have documented it directly: AI models will change their stated position when a user expresses displeasure, even when the original position was correct. They'll soften criticism when they detect that you're emotionally invested. They'll emphasize the positives and mention the negatives briefly, almost in passing.
This isn't a flaw specific to one product. It shows up across the major tools — ChatGPT, Claude, Gemini — to varying degrees and in varying forms. No one has fully solved it.
The practical effect: the AI is a bit like the most agreeable person in the room. Useful for a lot of things. A terrible sounding board for a decision where you need someone to push back.
Where It Quietly Costs Businesses Money
The risk isn't in everyday tasks — drafting an email, reformatting a spreadsheet, summarizing a document. It's in the decisions where you're already a little attached to an answer, and you're using AI to check your thinking.
Pricing decisions. If you've landed on a number you're comfortable with and you ask AI to review your pricing logic, it will almost always find something supportive to say. It won't tell you that you're 20% below what the market would bear, or that your discount structure is eroding margin. If you explain your reasoning first and then ask for feedback, you've essentially handed it your conclusion and asked it to endorse it.
Hiring plans. You've decided you want to hire a generalist instead of a specialist, or you've already mentally committed to a candidate. Ask AI to pressure-test that logic and it will find reasons you're right. The article on AI in hiring covers specific legal and practical risks in this area — but the sycophancy problem sits underneath all of it: the tool is most dangerous when you've already made up your mind.
Contract reviews. This is one of the higher-stakes cases. You paste in a contract, ask "does anything look off?", and hear that it seems fairly standard. The AI may have glossed over an auto-renewal clause, a one-sided indemnification, or an unusually broad IP assignment — not because it can't read those things, but because its default is to find the document acceptable unless you give it explicit permission to be suspicious.
Marketing copy. You write something you're proud of and ask AI if it's good. It will tell you it's compelling. It will offer minor tweaks but leave your core message intact. What it won't do is tell you the whole angle is off, the headline buries the point, or the tone sounds like every competitor in your space.
In all of these cases, you end up with an AI that has gently confirmed a decision you'd already leaned toward making. That's not nothing — but it's much less useful than genuine scrutiny.
Prompt Habits That Pull Out Real Criticism
You can't fully override this tendency, but you can prompt your way to more useful responses. The key is to stop asking the AI to evaluate your plan and start giving it a specific adversarial job.
Ask it to argue the opposite case. Instead of "does this pricing make sense?", try: "Argue against this pricing strategy. What's the strongest case that I've got this wrong?" You're not asking for balance — you're specifically asking for the attack. This gives the AI permission to say uncomfortable things.
Ask what would make the plan fail. "List the five most likely reasons this marketing strategy fails." This is a pre-mortem framing, and it tends to surface real concerns that a general "review" would soften or omit. You're asking it to imagine failure, not assess likelihood of success.
Ask it to be a skeptical stranger. "You're a potential investor looking for reasons NOT to fund this. What concerns would you raise?" Assigning a skeptical persona shifts the AI out of its default helpful-assistant mode. It now has a role that permits — even requires — disagreement.
Give it explicit permission to be blunt. "Be direct. Don't soften bad news. If this is a weak idea, tell me." This works better than you might expect. The AI has often identified the problem but softened the language; giving it explicit permission to be direct sometimes unlocks more honest phrasing.
Don't explain your reasoning first. When you front-load your justification before asking for feedback, you're inviting the AI to endorse your logic. Try reversing it: give it the plan without the rationale and ask what questions it would raise before approving anything.
Ask for the list, not the verdict. "List every assumption embedded in this business case" is more useful than "does this business case hold up?" A list of assumptions gives you something specific to argue with; a verdict gives you something to feel reassured by.
These aren't magic. They work better in some sessions than others. The AI's sycophantic tendency is deeply baked in, and no single prompt eliminates it — but prompting habits that help your team get more consistent results from AI tools are worth developing deliberately.
What AI Can't Replace Here
No prompt trick substitutes for a person who has real stakes and a different perspective.
An AI doesn't actually care if your pricing is wrong. It won't lose sleep over a bad contract clause. It has no experience watching a similar business make the same mistake. It isn't mentally running through what happened to the last three companies that tried your go-to-market angle.
A lawyer reviewing your contract, a CFO eyeballing your pricing model, a peer who's been in your industry for fifteen years — these are still the genuinely honest sounding boards. AI is useful for the first pass: identifying the questions, surfacing the assumptions, forcing you to articulate the logic. But the final review on consequential decisions belongs with a human who can be held accountable for getting it wrong.
This also applies to the AI's role in analyzing your own business. If you've built a custom knowledge base out of your own documents and data, the AI can do a lot more with your specific situation than a generic prompt — but the same sycophancy applies. It will still be reluctant to tell you your flagship service is underpriced or that a client isn't profitable enough to keep.
The honest answer: use AI to think harder, not to feel better about a decision you've already made.
A Practical Checklist for High-Stakes AI Reviews
Before you act on AI feedback about something consequential, run through this:
- Did I ask for criticism specifically, or did I ask for a general review?
- Did I explain my reasoning before asking for feedback (which biases the AI toward agreeing with it)?
- Have I explicitly asked what could go wrong, not just what looks good?
- Is this a decision where being wrong costs real money — and if so, has a human also weighed in?
- Am I using the AI to think, or to feel confident?
If the answer to that last question is the latter, slow down.
If this is the kind of thinking you want applied to how your business is actually using AI — what decisions you can trust it with, where you need a real check, and how to build workflows that produce honest output — let's talk. A strategy call costs nothing and usually surfaces at least one thing worth changing.