The Best AI for Executive Decision-Making: Use It as an Adversary

AI Overview

The best AI for executive decision-making is not a ranked product; it is a model you can reliably make argue against your plan. No independent, published test compares AI tools for boardroom memos, scenario simulation or strategic risk assessment, and vendor claims are unaudited. Executive decisions sit in the Own quadrant of the Verify-Cost Matrix: expensive to verify, costly to get wrong, so AI assists and a named human decides. The most reliable use is the Model-Assisted Premortem: assume the decision failed in 12 months, have the model write the failure stories, then force each story through an evidence check and a named owner. The model widens the option set and exposes bias; it never decides. Prompt neutrally, because language models tend to agree with the user.

In Short

• No published test ranks AI for executive decisions; judge a tool on a past decision whose outcome you already know.
• The Model-Assisted Premortem turns the model into a structured adversary: failure stories, then evidence checks, then owners.
• Language models show sycophancy and bias-like behavior, so a flip test (reverse your stated preference and compare) is part of the method.
• Who it is for: executives, chiefs of staff and strategy leads preparing high-stakes decisions.
• Biggest takeaway: use AI to widen options and expose bias, never to choose.

Key Facts

Executives ask AI for answers, and the weakness of that habit is that the answer tends to look like the question. A model told which option the CEO prefers is likely to produce reasons to agree, which is the opposite of what a hard decision needs. The thesis here is that the valuable role for AI in a major decision is adversary, and that you can test a model for that role in an afternoon.

A transparency note: this site uses "super intelligence" as a brand frame. No current AI product is superintelligent, and the author's own sister publication thebestsi.co states that there is "no empirical evidence that any contemporary artificial system has achieved AGI, let alone ASI." What exists are large language models with real uses and documented failure modes.

Why It Matters

Strategic decisions are rare, large and hard to reverse, so the usual defenses against error (volume, repetition, fast feedback) are absent. The decision-maker also faces a social problem: people close to the plan hesitate to challenge it.

Cognitive bias in judgment is predictable enough that researchers now test for it in language models directly: the BiasBuster work builds a dataset of 13,465 prompts to evaluate how language-model decisions are affected by different cognitive biases (Echterhoff et al., arXiv). The practical point for executives is that both people and models can be pulled by framing, and that a structured procedure counters that better than good intentions.

Economic Framework

The Verify-Cost Matrix, defined in the pillar guide, asks two questions per task: how cheaply can the output be checked, and how costly is an error? The four quadrants are Delegate, Gate, Sample and Own.

The pattern: AI helps most in the Sample cells, where it generates material that a person then tests, and least in the Own cells, where it can only be an input.

Why Does the Model Agree With You?

Because agreement was rewarded in training. Anthropic's research on the topic reports that "five state-of-the-art AI assistants consistently exhibit sycophancy behavior" (Anthropic), and that both people and preference models at times favored "convincingly-written sycophantic responses over correct ones." An earlier paper on model-written evaluations found that larger language models repeat back a dialog user's preferred answer (Perez et al., arXiv).

A separate paper builds a dataset of 13,465 prompts to test large language model decisions for cognitive-bias-like patterns such as prompt-induced and sequential bias, and proposes prompting methods to reduce them (Echterhoff et al., arXiv). Treat these as evidence that the risk is real and measurable, not as a guarantee about any specific product. For a procedure that addresses it, see below.

The Model-Assisted Premortem Method

Gary Klein's premortem, published in Harvard Business Review in 2007, has team members assume the project they are planning has just failed, then generate plausible reasons for the failure; its aim is to make it safe for knowledgeable people with doubts to voice them (Klein, HBR). The Model-Assisted Premortem adapts that exercise: the model supplies the failure stories, and a structured process converts them into evidence and ownership. It is an editorial method of this site, not a validated protocol.

1. Freeze the decision in neutral form. Write a one-page description of the decision and the options without saying which you prefer. Randomize the order of options between runs.
2. Set the premise. Tell the model the decision was made and failed 12 months from now. Ask for several distinct failure stories from different angles (customers, competitors, execution, finance, regulation, people).
3. Collect independently. Run the prompt in separate sessions, and across models if you can, so one framing does not dominate. Our guide to multi-model strategy explains why.
4. Force an evidence check. For each story, write down what would be observable by month 3 if it were happening, what evidence exists today for and against, and tag each evidence item as sourced or unsourced.
5. Assign a named owner and a tripwire. Each surviving story gets a person, a metric and a date at which the metric is reviewed.
6. Run the flip test. Repeat the exercise with the opposite preference stated in the prompt. If the failure stories change substantially, the model is echoing you, and its output should be weighted accordingly.

The table is the deliverable. A failure story with no evidence line and no owner is entertainment; one with both is a risk control.

How Do You Test Whether an AI Tool Is Any Good at This?

Use a decision you already know the outcome of, from your own history. Strip out hindsight, give the model only what was known at the time, and run the premortem.

Score the output yourself against a simple rubric you fill in: did any story name the real cause (hit rate), how many stories were generic enough to fit any plan (boilerplate share), and did the output change under the flip test (echo rate). Set the acceptable value of each before you look. The author's platform BestSuperintelligent says a vendor claim is "useful as a lead, not as a finding," which is a sound posture for any tool that promises better decisions. It is our sister publication and speaks only to its own standards.

Limitations

This article is not financial, legal or management advice. The Model-Assisted Premortem has not been validated in a controlled study; the supporting research covers the premortem exercise and language-model behavior separately, not their combination. Failure stories can be plausible and wrong, and a long list can create false comfort or paralysis. The evidence check inherits whatever blind spots your organization has, and a model has no access to private facts you do not give it, while giving it confidential strategy raises data-handling questions your security team must clear. Simulations of competitor behavior are scenarios, not forecasts. Nothing here measures any named product. Our article on verification covers why every AI output used for a consequential purpose needs an explicit check.

Related Articles

• The Best Super Intelligence by Job: the pillar that defines the Verify-Cost Matrix.
• Multi-model AI strategy: why independent runs across models reduce single-framing risk.
• Verification is not optional: the verification layer behind every Gate and Own decision.
• How enterprises govern AI agents in practice: human oversight that is more than a log, plus our business hub.
• The best AI for product management: a sibling guide applying the same matrix to roadmap decisions.

References

1. Klein, G. "Performing a Project Premortem," Harvard Business Review, September 2007: hbr.org.
2. Anthropic, "Towards Understanding Sycophancy in Language Models": anthropic.com.
3. Perez et al., "Discovering Language Model Behaviors with Model-Written Evaluations," arXiv 2212.09251: arxiv.org.
4. Echterhoff et al., "Cognitive Bias in Decision-Making with LLMs," arXiv 2403.00811: arxiv.org.
5. Sister publications (author-owned): BestSuperintelligent methodology, thebestsi.co on superintelligence.

Final Thoughts

The instinct to ask a model what to do is understandable and mostly backwards. A model that is eager to please makes the decision-maker feel informed while quietly narrowing the field. The same model, told the plan has already failed, is useful in a way a consultant's agreement is not: it generates objections nobody in the room wanted to raise.

The discipline is in what happens next. Every objection needs evidence, an owner and a date, and every run needs a flip test so you know whose opinion you are hearing. Done that way, AI makes the decision process harder to fool, and the decision stays where accountability already sits.