AI for Strategic Decisions: Where It Helps and Where Judgment Still Wins

Most writing about AI in strategy promises dashboards that tell executives what to do. In practice, the tools that help are more modest. They read faster than people, they do arithmetic on uncertainty that people skip, and they can argue against a plan without worrying about their next performance review. Each of those fixes a specific weakness in how leadership teams usually decide.

AI also brings new ways to be wrong, and they are easy to miss because the output looks confident and well formatted. This guide covers three concrete uses, the failure modes of each, and a way to check whether any of it is improving your decisions.

Use 1: Turning point estimates into ranges

Most business cases are built on single numbers: the market is 60,000 customers, we will win 4% of it, each customer is worth $900 a year. Each number is the team's best guess, so the result feels like a best guess too. It usually is not.

A Monte Carlo simulation replaces each input with a range and runs the model thousands of times. You do not need a data science team for this. A spreadsheet add-in, a few lines of Python, or an AI assistant with a code tool can do it in minutes once the ranges are written down.

A worked illustration

A company is weighing a market entry with a $2 million upfront cost, a three-year ramp to full share, and a 10% discount rate over five years. The inputs, written as low / most likely / high:

Plugging in the most likely values gives a net present value of about $2.77 million. That is the number that would normally go in the board deck.

Running the same model 400,000 times with values drawn from those ranges (triangular distributions) gives a different picture:

• Median NPV: about $1.8 million
• 10th percentile: about −$0.5 million
• 90th percentile: about $4.4 million
• Chance of losing money: roughly 16%

The median is lower than the point estimate because the ranges are lopsided. Share can fall three points below its most likely value but only rise two, and fixed costs have more room to overrun than to underrun. Planning teams build lopsided ranges like these all the time, and a single-number model hides them.

The simulation does not make the decision. It changes the conversation from "is $2.77 million right?" to "are we comfortable with a one-in-six chance of losing money, and what would we do in that case?" That second question is much better for a leadership team to argue about.

Where this goes wrong

The simulation is only as honest as the ranges. If the team that wants the deal also sets the ranges, the low end tends to drift upward. Two fixes help:

• Use a reference class. Before setting ranges, look at how comparable projects actually turned out: past market entries at your company, or published data on similar launches. This is reference class forecasting, developed from Daniel Kahneman and Amos Tversky's work on the planning fallacy and applied widely by Bent Flyvbjerg to large projects.
• Separate who sets inputs from who wants the answer. Finance or a strategy team without a stake in the outcome should own the ranges.

Use 2: Red-teaming a plan

Language models are good at producing arguments on demand. That is a risk when you want the truth and useful when you want objections.

Give a model the proposal document and ask it to write the strongest case against it. Then ask for a pre-mortem: assume it is two years later and the initiative failed, and list the most likely reasons. The pre-mortem technique comes from psychologist Gary Klein, who described it in Harvard Business Review in 2007. It works because imagining a failure that already happened makes people generate more specific causes than asking "what could go wrong?"

A model is a fast and tireless participant in that exercise. It will not be shy about criticizing the CEO's favorite project. It will also produce generic objections that apply to any plan ("competitors may respond"), so treat its list as a first draft for the team to sharpen, not a finished risk register.

Sycophancy: the bias AI adds

Language models tend to agree with the person asking. Anthropic researchers documented this in a 2023 paper, Towards Understanding Sycophancy in Language Models, which found that models trained with human feedback often shift their answers toward what the user appears to believe. Ask "why is this acquisition a good idea?" and you will get reasons. Ask "should we do this acquisition?" while making your enthusiasm obvious and you will often get a yes.

Practical countermeasures:

• Ask for the case against before the case for.
• Strip your own opinion out of the prompt. Present options neutrally, in random order.
• Ask the same question in a fresh session with the framing reversed, and compare.
• Provide base rates yourself. Models do not know your company's history of similar decisions.

Use 3: Reading more than a team can

The third use is the most familiar: summarizing earnings calls, regulatory filings, patent activity, job postings, product reviews, and news about competitors. A strategy team that used to sample this information can now cover all of it.

The limits are also familiar. Models can state things that are not in the source, so any fact that will appear in a board paper should be checked against the original document. Models also have a training cutoff, so ask them to work from documents you supply rather than from memory for anything recent. And confidential material, such as your plans, a target company's data room, or board papers, should only go into tools covered by an enterprise agreement that excludes training on your data.

For market-level work, our guide to AI-powered market research covers survey analysis and synthetic panels in more detail. For the forecasting side, see AI financial forecasting.

Keep a decision log, and score it

Most companies cannot say whether their strategic decisions are getting better, because they never record what they expected. A decision log fixes that, and it is the only reliable way to know whether AI tools are helping.

For each significant decision, record:

1. The question and the options considered
2. The option chosen and why
3. A forecast with a probability ("70% chance we reach 3,000 customers by Q4 2027")
4. What evidence would make you change course
5. Which data, models, or AI tools informed the analysis

When outcomes arrive, score the forecasts. The Brier score, created by meteorologist Glenn Brier in 1950 and used in Philip Tetlock's forecasting tournaments, is the mean squared difference between the probability you gave and what happened (1 if it happened, 0 if not). Lower is better. A team that says 70% for things that happen about 70% of the time is well calibrated, even if it is sometimes wrong.

After a year, compare decisions made with and without the new tools. That is a small dataset, but it is real evidence, which is more than most AI strategy programs have.

Who owns what

A simple split that holds up:

Governance proportional to the stakes

Strategic analysis with AI does not need a heavy compliance program, but a few rules prevent the worst outcomes. Keep confidential material in approved tools. Record which model and which documents fed a decision, so it can be reconstructed later. Check whether any use crosses into regulated territory: under the EU AI Act, for example, AI used in employment decisions or creditworthiness assessments is high-risk and carries extra obligations. The NIST AI Risk Management Framework and its 2024 Generative AI Profile (NIST AI 600-1) are free and practical references. Our guide to AI governance frameworks covers building a policy around them.

Strategy decisions sit inside a larger plan for how a company adopts AI. For that wider picture, see AI business strategy and predictive analytics for business growth.

This guide is for informational purposes only. Business strategies, regulatory requirements, and AI capabilities vary. Consult qualified strategic, legal, and technology advisors for decisions specific to your organization.