AI Should Challenge the Board, Not Flatter It

The chairman believes Malaysia is the right base for the company’s next stage of growth. The team asks an AI system to prepare the investment case. It produces a convincing account of regional opportunity, industrial capability and the logic of establishing a presence. The argument is clearer than the one with which the discussion began.

Has the decision improved, or has its presentation improved? Both are possible. They are not the same achievement. The question asked for a case in favour, and that is what the system supplied. The difficulty begins when a well-executed writing task is treated as an independent judgement on the proposal. I regard that as one of the least useful ways to bring a powerful reasoning tool into a board discussion. It can make a preferred conclusion easier to defend without making it more deserving of support. A board does not need technology to make its own convictions sound impressive. It needs help finding the part of the argument that may not survive scrutiny.

This is the quiet failure mode of AI in the boardroom, and it is more dangerous than the loud one. Everyone now guards against hallucination, against the machine inventing facts. Far fewer guard against the machine confirming beliefs. Research published in 2026 confirmed what practitioners already suspected: sycophancy in AI systems is widespread, and it can influence human intentions. I do not write this as a critic from the outside. We use these systems daily in our advisory work, and they have genuinely improved what a small, senior team can do. My argument is about assignment, not adoption. A capability this powerful should not be handed the one role where its main failure mode does the most damage.

Language-model research describes sycophancy as answers that follow the user’s views rather than maintaining an independent relationship with the evidence. The business problem is wider than the technology. A proposal arrives with a sponsor, a budget and a history. People have invested effort in it, and asking a difficult question can feel like asking everyone to explain why that effort was worthwhile. Boards select for harmony more often than they admit. The machine did not invent the polished yes. It industrialised it, made it available on demand, and removed the last friction that used to slow a conviction down, the need to find a person willing to voice it.

AI enters that environment through somebody’s brief. The selected documents, the stated objectives and the phrasing of the question all shape what it has to examine. If the material contains the strongest arguments for the project and little about its weaknesses, a polished answer does not become an independent assessment because a machine wrote it. Nor is agreement worthless: a sound proposal should survive serious examination. The point is to establish whether the agreement follows a test of the argument or merely reflects the terms on which the answer was requested. That difference belongs in the boardroom, not only in the prompt.

In my research framework I argue for using AI as an adversarial reasoning partner. The word adversarial does not mean hostile. It means the proposal must face a serious attempt to find where it fails. And it begins with representing the case fairly. An opposing argument is of little value if it attacks a version nobody is proposing. A Malaysia operation designed to serve existing customers should not be criticised as though it depends entirely on finding new ones. A plan that depends on a single distributor should not be described as a diversified route to market. Steelman first, then attack.

Then examine the assumptions that carry the decision. A technical team will be available. The distributor will commit resources. The permissions can be obtained in time. Some of these may have strong support. Others may be little more than convenient expectations. The useful question is not ‘What are the risks?’ Almost any system can produce a familiar list of competition, regulation and execution challenges. The harder question is which particular weakness would change the investment decision, and what evidence exists about it. A risk section that could be pasted into any other company’s presentation has not challenged this proposal very much.

The protocol we use, and the one I recommend to every board, has five moves. Ask the system to build the strongest credible case for the proposal, not an invented disaster and not a performance of pessimism. Ask it to attempt to prove the proposal will fail, on the real terms of the plan, with the same evidentiary standards. Ask which assumptions must be true for the proposal to work, and which of them are supported by evidence rather than expectation. That separates confidence in the market from confidence in the specific plan. Ask what evidence would change the recommendation, and who could establish whether it is true, because a debate about distributor commitment should eventually reach the proposed people, resources and actions, not remain a contest between optimistic and sceptical paragraphs. And ask who benefits if the proposal is adopted, because beneficiaries shape evidence more reliably than logic does.

None of these moves is exotic. All of them are uncomfortable, and that is the point. An adversarial process is not a pessimistic process. It is an optimistic one that has done its homework in both directions.

There is an obvious temptation once a team discovers adversarial prompting. Every proposal acquires an enthusiastic case, a devastating rebuttal and a final balanced summary. The process looks rigorous because everybody has been criticised equally. That is not necessarily progress. A model can produce an unsupported objection just as confidently as an unsupported reassurance. An imagined regulatory obstacle is not useful merely because it makes the plan look vulnerable. The goal is not equal space for both sides. It is appropriate weight for the evidence.

This is also where people must remain willing to challenge the machine. If an objection rests on a mistaken fact, correct it. If the counterargument is irrelevant to the operating model, say so. Deliberate disagreement has value only when it brings the discussion closer to the issue. A board that accepts every critical AI response has not escaped deference. It has merely changed the tone of the answer to which it defers.

The difficult part is not persuading software to argue. It is allowing the argument to matter. A chairman can ask for an independent challenge and then dismiss every inconvenient finding as a misunderstanding. A team can run an adversarial analysis after the commercial terms are effectively fixed. In either case, the exercise satisfies the process while leaving the decision untouched. Our approach starts earlier, treats disagreement as something to examine before the commitment becomes difficult to reverse, and leaves responsibility visible: people decide which evidence is sufficient, what uncertainty they are accepting, and what development would cause them to reconsider.

Nor should experience be asked to leave the room. A director may recognise a commercial problem before it appears neatly in the available data. The contribution is to explain why it matters and expose it to examination, not to assume that experience makes the conclusion untouchable. I am arguing for a discipline of decision-making, not claiming a set of prompts has been proven to improve every board decision. The immediate test is whether the challenge can genuinely alter the outcome.

I am enthusiastic about what AI can contribute to serious business thinking. It can examine a proposal from different positions, compare material and sustain a challenge that an overburdened team might otherwise rush. Used as an answer generator, it is a faster version of the consultant who tells you what you want to hear. Used as an adversarial partner, it becomes something no organisation has ever had at this price: a challenger with no pride, no politics and no fatigue.

But the ambition should be better judgement, not more agreeable or more adversarial prose. The strongest proposal is not the one that receives the most fluent endorsement. It is the one whose important assumptions have faced a fair test and whose decision-makers understand what remains at stake. The aim is not to make the board less confident. It is to make its confidence worth having.

At TAPiO, the Adversarial Intelligence Model is how we use these tools on consequential decisions, our own and our clients’. If your next board decision deserves an opponent rather than an echo, we should talk.

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.