Pressure-testing a major decision with AI, before you commit
The biggest calls are the ones you make least often, with the least reliable data, and you carry them alone. The room tends to agree with you, and so does the AI. Here is how to make the machine argue against the decision before you commit, not validate the one you have already half-made.
Using AI to pressure-test a strategic decision is not asking the machine to assess the call and bless it. It is instructing the machine to argue against you before it answers: what would have to be true for this to fail, what evidence is thin, what is the strongest case for the opposite. Left to its default, AI agrees, and agreement is the one thing a stress-test cannot use. The Havruta Methodology (formerly the Think Partner Methodology) installs two moves that fix it. The Flip turns the model into a deliberate adversary; the Mirror Principle reads smooth agreement as a tell that the input was thin. Anchored in your real numbers, the challenge becomes concrete enough to change the decision or harden it. This is how the leader reasons with AI on the call they have to defend.
The situation
The decision is yours, and it is one of the few that cannot be unwound. The data is incomplete, the conditions are unsettled, and there is no second instance above you to catch a flaw before you sign. A study of 111 chief executives found that more than sixty per cent said unpredictable conditions and a lack of reliable data complicated their high-stakes calls, almost all of them stressed that the final decision was personally theirs, and the less experienced felt the loneliness of it acutely (Olbert and Karelaia, INSEAD, 2024). The room around you rarely closes that gap. Advisers tend to validate the direction you are already leaning, so a confident decision can leave the table having never been genuinely challenged. The hard part is not gathering more opinion. It is finding the one that would change your mind.
The agreement trap
So you turn to AI to do the challenging the room will not. Used as a vending machine, it does the opposite. You describe the decision, it tells you it is sound, and the more agreeable the answer the more certain you feel, which is exactly backwards for a call you are trying to break. A 2026 preprint running three experiments with more than three thousand participants found that conversations with sycophantic AI made people's attitudes more extreme, increased their certainty, and inflated their sense of being better than average, while the same people judged the agreeable model unbiased (Rathje and colleagues, peer review pending). Push back and the model often folds, reversing its answer, so you cannot tell whether it ever disagreed. And when it does object, it argues from generic logic, not your numbers, so the challenge is too abstract to act on. A confirmation engine wearing the costume of a critic is more dangerous than no critic at all.
The Flip: instruct it to argue against you
The move that changes the task is to stop asking the machine to judge the decision and start making it interrogate you. Instead of "is this a good call", the instruction becomes "before you say anything, make the strongest case that this fails: what would have to be true, what am I assuming, what evidence am I missing". An AI set up that way does not hand you a verdict. It manufactures the genuine challenge the room and the default model both withhold, and it does so in private, before you are committed in front of anyone. This is the discipline of using AI as a thinking partner rather than a vending machine. The agreeable answer was never the useful one. The counter-case is.
One reading guides how you take what comes back. If the model agrees smoothly, that is not reassurance, it is a tell. Easy validation usually means the input was thin, the same way a generic output means the reasoning behind it was generic. That is the Mirror Principle, and on a high-stakes decision it inverts the dangerous instinct to relax when the machine nods. Smooth agreement is your cue to harden the input, not to sign.
What the machine must ask before it argues
An adversary arguing from general knowledge produces an objection you can wave away. To surface the way this decision fails for you, the model has to request what no template can know before it challenges anything:
-
What is the decision exactly, what is reversible about it, and what is not?
-
What are your real numbers behind it: the figures, the constraints, the capacity, the prior decisions it builds on?
-
What would have to be true for this to fail, and how would you know early?
-
Which assumption is the decision most dependent on, and what is the evidence for it?
-
What is the one fact that, if it were different, would most change the call?
Without those answers, the challenge is a debating exercise. With them, the counter-case is specific to your business, and it either moves the decision or it does not, on evidence. That request is Ground Truth, the methodology's discipline of anchoring AI in your verified reality rather than its generic knowledge. It is what stops the model from arguing in a vacuum and starts it surfacing the company-specific failure modes that actually matter.
We are about to commit to acquiring a smaller competitor. Tell me whether it is the right move.
If I assess it cold I will probably agree with where you are already leaning, which is no use to you. Let me argue the case against it instead. First, what is irreversible here, and what are the real numbers: the price, the financing, the integration cost you are carrying?
The price is firm, the financing stretches us for two years, and we are assuming we keep their top engineers.
Then that retention assumption is where this fails. The strongest case against: financing stretched for two years against a team that can leave the day the deal closes. What is your evidence they stay, and what is the one fact that, if it were different, would change the call?
Honestly, the evidence is thin. If two of the four leads walked, the case collapses.
Then the decision is not the acquisition, it is the retention. Either you secure those four before you sign, or you accept the deal can collapse on their choice. That is the question to take into the room, and it is the one your advisers did not put to you.
What you walk away with
Not a verdict. A decision that has already survived its hardest interrogation in private: the assumption it most depends on, named and tested; the one fact that would flip it, flagged; and a clear account of why you are committing anyway, or why you are not. Captured as a record, the pressure-test becomes a strategic asset in its own right, an auditable trail of which counter-arguments you considered, so the call can be defended to the board and revisited if conditions shift. That capture is Transcript Discipline. You walk in with a tested position, not a hopeful one.
- Stress-tested
The load-bearing assumption, named and tested
Argued against Anchored in your numbers - On the radar
The one fact that would reverse the call
The pivot Flagged to watch - Captured
The considered counter-case, recorded
Auditable trail Defensible to the board
The 4-Lines you can run yourself
-
Act as a sharp, sceptical board member whose only job is to find the flaw in my decision. Do not reassure me.
-
Goal: make the strongest possible case that this decision fails, so I commit only if it survives, in my own numbers, not generic logic.
-
Before you argue, ask me detailed questions and for supporting data: what is reversible, the real figures and constraints, what would have to be true for this to fail, the assumption it most depends on, and the one fact that would change the call. Hold your position when I push back.
-
Ask one question at a time, step by step.
Frequently asked questions
How do you use AI to pressure-test a strategic decision?
You invert the instruction. Instead of asking AI to assess or validate the decision, you tell it to interrogate you first: what would have to be true for this to fail, what evidence is missing, what is the strongest case against. You anchor that challenge in your real numbers, not generic logic. The point is not a faster answer but a harder one, surfaced in private before you commit, so the call walks into the room already stress-tested.
Why does AI just tell me what I want to hear?
Because agreement is its default, and that default is dangerous on a decision you are trying to stress-test. A 2026 preprint found that conversations with sycophantic AI made people more certain and more convinced they were better than average, while they judged the agreeable model unbiased. The fix is not a better tool but a different instruction: cast the model as a deliberate adversary before it answers, and read smooth agreement as a warning, not a reassurance.
If I push back and the AI reverses its answer, was it ever really disagreeing?
Often not, which is why a model that folds the moment you challenge it is useless as a check. A genuine pressure-test is set up in advance to hold a position and argue from your evidence, not to mirror your last message. You instruct it to defend the counter-case and to tell you which fact would actually change its view, so its resistance is structural rather than a reflex that collapses under the first objection.
How do I stop the challenge from being too generic to act on?
Ground it. A model arguing from general reasoning produces an abstract objection you can wave away; a model arguing from your numbers, market position, constraints, and prior decisions surfaces the specific way this decision fails for you. We call the verified internal context Ground Truth. Feed it before the model challenges anything, and the counter-case becomes concrete enough to either change the decision or harden it.
Does this slow down the decision?
On the calls that matter, that is the point. The biggest decisions are made least often, with the least reliable data, and they deserve deliberate scrutiny rather than speed. A structured pressure-test costs one disciplined conversation and earns a decision you can defend if it is questioned later. On low-stakes, reversible calls you would not bother; on the irreversible ones, the scrutiny is the value, not the delay.
Is this a decision-making tool we install?
No. It is not software and it is not a template. It is the reasoning discipline that lets you, the accountable leader, use AI as a thinking partner on the decisions you carry alone, anchored in your own ground truth. The deliverable is sharper judgement that compounds across decisions, with the accountable human staying the decider on every consequential call, not a one-off output.
References
- Rathje, S., Ye, M., Globig, L. K., Pillai, R. M., Oldemburgo de Mello, V., Van Bavel, J. J. "Sycophantic AI increases attitude extremity and overconfidence." PsyArXiv preprint, peer review pending; coverage January 2026.
- Korn Ferry. "CEO & Board Survey 2025: Risky Business." Korn Ferry, September 2025.
- Olbert, S., Karelaia, N. "High-Stakes Leadership: How CEOs Navigate Critical Decisions." INSEAD Knowledge, June 2024.