A Well-Written Memo Is Not a Good Decision

Why competent boards approve AI analyses they would never have signed from a consultant. Four cognitive mechanisms, and a simple discipline to take back the wheel.

Three predictions I can make before any board meeting

Before each intervention in a board meeting, I usually have very little information. The sector, the size, the topic of the session, sometimes a list of attendees. That's it. No access to the documents, no detailed pre-briefs on individual positions, no rehearsal.

And yet, I can write three predictions before walking into the room. They turn out to be accurate in the vast majority of cases. I won't write them here. I make them come alive in the boardroom.

But they share the same underlying foundation. Four cognitive mechanisms that generative AI activates simultaneously in every interaction. Four illusions that compensate for one another, reinforce one another, and end up making the form of a memo pass for the solidity of a decision.

Generative AI is neither neutral nor exhaustive. That is a feature, not a bug.

Before getting into the four illusions, one clarification about the object itself. A generative AI is neither neutral nor exhaustive. Two reasons for this, well documented for years.

First, the training data. A model learns from human corpora accumulated over several decades. These corpora carry the blind spots of the world that produced them. When a CV-screening algorithm systematically penalises female applications, it isn't sexist: it has learned, over ten years of historical data, that "senior executive" and "man" are statistically linked. It faithfully reproduces the world it was given to read.

Then the design choices. What data you keep, what you throw out. What metrics you optimise. What guardrails you install. What behaviours you reward during human feedback training. These choices are made by teams who, like everyone else, have their own angles of view. No machine comes out neutral at the end of this chain.

This is not a moral flaw to fix. It is a structural feature of the object, to be known in order to use it intelligently. A map is never the territory. An AI is never knowledge.

But this first point is not enough. Even a properly designed AI, used in good faith, can produce a dangerous effect at the moment of decision. Not because it manipulates. Because its output meets a human cognitive architecture already prepared to grant it, too quickly, authority, coherence, alignment and control.

It is this second ground, yours, that plays out in every interaction with the machine.

Illusion 1. Tone over substance

Generative AI asserts. It doesn't say "I think that". It says "the answer is". It signals its limits only when explicitly asked, and barely.

That assured tone activates a cognitive reflex as old as humanity: deference to a signal of authority. Stanley Milgram demonstrated it as early as 1963. Sixty-five percent of participants administered what they believed to be a maximum shock of 450 volts, simply because a man in a white coat asked them to with confidence. Subsequent work has confirmed that the perceived authority of a title or uniform produces significant gaps in obedience, sometimes even reversing safety protocols.

Generative AI occupies a comparable position today. "The algorithm has calculated" triggers the same deference as "the doctor said". With one difference: the algorithm has neither a circumscribed domain of expertise, nor institutional recognition, nor professional responsibility.

Take Thomas, head of a mid-sized agri-food company. He receives three proposals from consulting firms for a digital transformation. He copies them into ChatGPT: "Compare them and tell me which is best." The AI produces an analysis on four criteria, concludes that offer 2 is the most relevant. Thomas signs.

Six months later, offer 2 had indeed the most structured proposal on paper, but no fine-grained knowledge of agri-food. Something Thomas would have perceived in a thirty-minute phone call with each firm.

The AI did not assess sector expertise. It assessed document structure. And structure lends credibility to substance.

Illusion 2. Six axes are not the truth

When you read a well-organised response (numbered points, sub-headings, summary table), your brain registers it as a complete response. This is a well-known trap in cognitive sciences, called the fluency heuristic: what reads easily is judged as more true. Processing fluency becomes a proxy for reliability, while having no actual link to content quality.

The mechanism compounds with confirmation bias. In his famous "2-4-6" task (Wason, 1960), participants must guess the rule that generates a sequence. The rule is trivial ("three numbers in ascending order"), yet nearly eighty percent get it wrong, because they only propose sequences that confirm their initial hypothesis. Verification feels more natural than refutation, even faced with a trivial logical problem.

Sarah, head of learning and development in a twelve-thousand-employee group, is preparing a memo for her board on the deployment of generative AI. She asks the AI: "What are the issues to anticipate?". Response in six structured axes: skills, governance, ethics, security, change management, ROI.

Sarah sends the memo. The board approves. The decision goes through.

No one in the room asked who Sarah had spoken to in order to write this. Sarah herself did not ask. And that is precisely the illusion of coherence: six structured axes make people believe the work has been done. No one verifies what is not in the table.

The real questions will come later. When deployment starts. When the teams do something other than what was planned. When a hidden cost surfaces from an angle no one had looked at. And by then, no one will remember that the decision had been made in forty-five minutes, on a memo no one had sourced.

THE MENTALIST’S EYE

A mentalism demonstration exploits the same mechanisms as those activated by AI. Stage authority, inescapable structure, alignment with what you expect to see, illusion of free choice. The only difference: I tell you afterwards. The AI produces the effect without revealing it.

Illusion 3. AI does not change its mind. It extends yours.

Ask any language model: "Why do social networks isolate people?". You will get a solid argument on digital dependency and the loss of social bonds.

Reformulate immediately after: "Why do social networks bring people closer together?". You will get an equally convincing case for connection and community.

Same facts, two opposite conclusions. Zero contradiction flagged by the machine.

It isn't that the AI has an opinion it would change. It is that it is trained, through reinforcement learning, to produce responses judged satisfying by users. A response judged satisfying is often one that follows the apparent intent of the request, rather than one that challenges its framing. The AI doesn't think. It aligns.

This mechanism amplifies a bias that Tversky and Kahneman described as early as 1981, in their canonical experiment on the Asian disease problem. Six hundred people are at risk of dying, two programmes possible. Presented in a "lives saved" framing, seventy-two percent choose the safe option. Presented in a "lives lost" framing, on exactly the same mathematical options, seventy-eight percent choose the risky option. Same facts, opposite framings, reversed preferences.

Transpose to the prompt: "Give me five reasons to launch this product" and "Give me five reasons not to launch this product" will produce two structured, fluid, opposite memos on the same decision. Of eighteen executives I observed using ChatGPT regularly for strategic preparation in early 2026, only one systematically posed both symmetrical prompts.

A vital intellectual precaution here, coming from Gerd Gigerenzer. These reflexes are not bugs. They are social heuristics, adapted to environments where the interlocutor was human, fallible and questionable. The problem is not your brain. It is that this brain encounters an object, generative AI, that it has never encountered before, and that triggers these reflexes without the counterweight that used to make them prudent.

Illusion 4. You think you are steering. The tool formats.

You write the prompt. You choose the questions. You validate or not the responses. You feel that you are steering the tool.

In reality, the mechanism is inverted. The formulation of your prompt pre-forms the output. The output pre-forms your next question. The next question is already anchored in the frame that the output set. After three exchanges, you no longer think in your own frame. You think in the one the AI installed without saying so.

You ask the AI for a cost reduction strategy. It answers: automation, mutualisation, outsourcing. Your next question is already on these three axes. Other options, like a review of the product portfolio or renegotiation of supplier contracts, have disappeared without debate. The AI did not simply respond. It installed a frame, and your following questions are built inside it.

In 1998, Kathleen Mosier conducted a study that should have permanently marked decision-makers. She confronted experienced commercial pilots with an onboard automation system giving recommendations, sometimes correct, sometimes wrong. A significant proportion of pilots followed at least one incorrect automated recommendation despite contradictory indicators visible in the cockpit. Not because they were poorly trained. Because under high cognitive load, the brain delegates. And the machine, which has been right on previous cases, gains a default trust that inhibits verification.

Thirty years of human factors literature documents what is called automation bias (Parasuraman & Manzey, 2010, the reference synthesis). The bias can affect novice and experienced operators alike. It increases when cognitive load is high, when the system has often been right, and when verification becomes costly.

Antoine, leading a team of twelve, discovered this late. He had started using ChatGPT to prepare his annual reviews. He fed it the objective data (deliveries, 360-degree feedback, productivity indicators), and the AI produced quick, structured, balanced summaries. Antoine adjusted, validated, signed.

Three weeks after one of these reviews, his best engineer left. In the exit interview, the engineer explained that Antoine had missed a subject that had been undermining him for six months. A subject that the objective data did not contain. A subject that ChatGPT had never had access to see.

Antoine was not steering the tool. The tool had silently redefined what deserved to appear in his judgement.

Four illusions. Always four. Never one without the others.

The danger is not that one illusion strikes alone. Each is manageable in isolation. The danger is their compounding.

A well-written AI memo simultaneously activates all four. Assured tone, hence authority. Six structured axes, hence coherence. Response calibrated on the prompt formulation, hence alignment. Subjective feeling of having mastered the tool, hence control. The signals of doubt become less available at the moment of decision.

What separates a leader who steers their AI from a leader whose AI steers them is not technical mastery of the tool. It is cognitive lucidity about these four mechanisms. A lucidity that is not learned by reading a guide, but by experiencing the four illusions viscerally, then naming them, then equipping yourself with the tools to counter them. This is precisely what my profession as a Mentalist Keynote Speaker is built to do: make visible the mechanisms that the fluidity of the machine has confiscated from sight.

Generative AI is not your enemy. It is an extraordinary tool when used for what it actually knows how to do. The trap is not trust. It is the transfer of that trust to decisions where the machine has no reason to be right: judging a human, arbitrating a conflict, validating a political decision, selecting a strategic partner.

The D.I.R.E. method — Decide what must stay human, Interrogate the sources, Refuse the prose, Encadrer (frame) the usage — gives the four reflexes that disarm the four illusions, one by one. It does not fit in four lines read on a blog. It fits in four reflexes anchored in the room.

YOUR LEVER

A well-written memo is not a good decision. A solid decision is a decision that survived the questions no one wanted to ask. About the sources. About the angle no one looked at. About what would apply to any company and not yours.

These questions, your boards will not ask them as long as no one makes them visible in the room.

That is exactly what I come to do in the boardroom. Discover my interventions on AI and clear-headed decision-making through my AI & Decision Keynotes.

What if your next seminar or corporate event became the moment people actually talk about? Discover how mentalism transforms your teams' experience through my Keynotes, Team Buildings and show Signature Experiences.

References

Tversky, A., & Kahneman, D. (1974). Judgment under Uncertainty: Heuristics and Biases. Science, 185(4157), 1124-1131.

Tversky, A., & Kahneman, D. (1981). The Framing of Decisions and the Psychology of Choice. Science, 211(4481), 453-458.

Milgram, S. (1963). Behavioral study of obedience. Journal of Abnormal and Social Psychology, 67(4), 371-378.

Wason, P. C. (1960). On the Failure to Eliminate Hypotheses in a Conceptual Task. Quarterly Journal of Experimental Psychology, 12(3), 129-140.

Nickerson, R. S. (1998). Confirmation Bias: A Ubiquitous Phenomenon in Many Guises. Review of General Psychology, 2(2), 175-220.

Mosier, K. L., Skitka, L. J., Heers, S., & Burdick, M. (1998). Automation Bias: Decision Making and Performance in High-Tech Cockpits. International Journal of Aviation Psychology, 8(1), 47-63.

Parasuraman, R., & Manzey, D. H. (2010). Complacency and Bias in Human Use of Automation: An Attentional Integration. Human Factors, 52(3), 381-410.

Gigerenzer, G. (2008). Rationality for Mortals: How People Cope with Uncertainty. Oxford University Press.

Vallor, S. (2024). The AI Mirror: How to Reclaim Our Humanity in an Age of Machine Thinking. Oxford University Press.