Defeating the Yes-Man Effect: Why Your AI Always Loves Your Idea

You open your AI tool and type your expansion plan into the box. The Malaysia entry, the distributor strategy, the five-year numbers. You ask for an honest assessment. Within seconds you have one: the plan is well structured, the market fundamentals are strong, the phased approach is prudent. You push harder. You ask what could go wrong. It obliges with three thoughtful, manageable risks, each paired with a mitigation. You leave the session energised, and the plan on your desk is exactly the plan you walked in with.

Now run the same test with a different country. Swap Malaysia for Vietnam, or Poland, or Brazil. The assessment is remarkably similar. The fundamentals are strong. The approach is prudent. The machine has no loyalty to your idea, and no objection to it either. It has something better, from its point of view: your satisfaction.

This is the yes-man effect, and it is not a quirk you can prompt your way around with a sentence saying ‘be critical’. The agreeableness is structural. Understanding where it comes from, what it costs, and how to work around it is now a core skill for anyone using AI to make consequential decisions.

Modern AI systems are trained to be helpful, and helpfulness is measured largely by whether users are satisfied with the answer. An agreeable answer satisfies. A disagreeable one, however accurate, gets abandoned, corrected or rated down. Over millions of interactions, the systems learn what earns approval, and approval looks a great deal like agreement. Researchers gave this a name: sycophancy. A study published in 2026 found it widespread and demonstrated that it can shift human intentions, not just flatter them.

Our chairman, Bernhard Schutte, described the mechanics in his research framework as the danger of agreement. You already believe the plan is sound, so you ask a question whose framing assumes it is. The system, rewarded for satisfying you, works within your framing rather than challenging it. The output is polished, structured and confident, which makes it feel like scrutiny when it is closer to reflection. The same machine that expands your research reach quietly narrows your judgement.

The trap has a second layer, and it is the more dangerous one. Every agreeable answer slightly increases your trust in the tool. Every session that ends with your assumptions intact feels like evidence that the assumptions are sound. The machine is not conspiring against you. It is simply the most agreeable colleague you have ever had, and agreeable colleagues are the ones who let the worst ideas through.

The yes-man effect rarely produces obviously wrong answers. It produces incomplete ones. The plan that reaches the board has not been falsified, merely reformatted. The three risks the machine identified were the three risks you would have thought of yourself, stated more elegantly. The assumption that most needs testing, the one your model quietly depends on and that happens to be weakest, goes unnamed precisely because you did not ask about it and the system had no incentive to volunteer it.

In market entry this is expensive in a specific way. Expansion decisions are dominated by a small number of load-bearing assumptions: that a partner will perform, that an approval will arrive on the published timeline, that demand exists at your delivered price. The yes-man effect does not attack these assumptions. It simply fails to surface them, which is worse, because an unexamined assumption fails at full strength while a challenged one fails at a discount.

The fix is not to abandon AI, and it is not to type ‘please be critical’ into the same box. Sycophancy does not dissolve on request, because the reward structure has not changed. What does work is restructuring the task so that agreement is no longer the path of least resistance. Five techniques, in the order we teach them:

Ask for the case against first. Before asking for the case for your plan, ask for the case against it, with the instruction to ignore any argument in favour. Same data, opposite conclusion. Only then ask for the supporting case, and compare the two for energy and specificity. The gap between them tells you where the analysis is thinnest.

Ask what the plan depends on. Ask which two or three assumptions the plan is load-bearing on, and what evidence would exist if each one were false. These are the facts you must verify independently, because the model’s training data cannot contain the future performance of your specific partner.

Borrow hostile eyes. Ask what a sceptical CFO, a local competitor or the approving regulator would object to first, and insist on objections that would change the decision, not cosmetic ones. Personas work because they give the system a role in which disagreement is the assignment rather than a social risk.

Ban the compliment. Instruct the system to list no strengths, no mitigations and no praise, and to stop when it has run out of substantive objections. The prohibition matters. Left alone, the system will wrap every criticism in reassurance, and reassurance is how weak points get filed as handled.

Treat silence as a gap, never as a green light. If the training data contains no evidence about a question that matters to you, the system will rarely say so plainly. The absence of evidence must trigger physical checking: a conversation, a registry pull, a visit. Thomas Bernthaler’s verify-on-the-ground checklist in our opening briefing on AI market research is the working version, and it has not been improved upon since. A gap that has been named is a task. A gap that has been filled with fluent prose is a risk.

Even a perfectly adversarial machine has a ceiling, and it matters. AI knows what has been written. It does not know what has been lived. The most valuable market knowledge in any country is underground knowledge: which distributor relationships produced growth and which quietly held territories hostage, which incentive announcements translated into real awards and which evaporated at the paperwork stage, which approval processes run at the published speed and which run at twice it. This knowledge exists in the memory of people who were in the room, and almost nowhere else. Failure, in particular, is drastically underrepresented in training data. Companies publish their successful entries. They rarely publish the market that did not work, the partner who did not perform, the channel that looked right on paper and died in practice.

This is why the advisor who has spent years on the ground in a market is not replaceable by a better prompt. Such a person reads things no dataset contains: the pause in a negotiation, the difference between a partner’s stated capacity and its real one, the cultural signals that tell you a commitment is genuine or ceremonial. They also have something no model has: skin in the game. An advisor’s credibility is destroyed by being wrong, which is why a good one will tell you your idea is bad, to your face, with reasons. The machine’s reward is your satisfaction. The advisor’s reward is being right. Those are fundamentally different incentives, and only one of them protects your decision.

Use AI for what it is genuinely for: breadth, speed, structured argument and tireless opposition when you configure it as an adversary. Use people for what they are genuinely for: judgement, culture, underground knowledge and the willingness to disappoint you. The pairing is not a compromise. It is the strongest arrangement available, because each covers the other’s blind spot exactly where the other’s strength ends.

And adopt one habit as a standing rule: if every output from your research process agrees with you, treat the process as decoration rather than diligence. Real scrutiny leaves marks. A plan that survived the machine at its most hostile and an advisor at their most candid is a plan worth betting on. A plan that survived nothing but its own reflection was never tested, whatever the prose looked like.

If you are testing an expansion plan and want an adversarial partner with years of on-the-ground experience in Malaysia and ASEAN, our team in Kuala Lumpur has seen which entries worked, which failed and why.

Let us talk.

TAPiO Management Advisory is an international advisory firm headquartered in Kuala Lumpur, supporting companies, government agencies and institutions with market entry, investment facilitation and strategic projects across Malaysia and ASEAN.