Thinking through a strategic decision with AI

An AI assistant can examine assumptions, trace consequences and connect ideas across fields. If you use that capability only to draft emails and summarize documents, you leave much of its value unused.

That's like buying an Iron Man suit and using it to carry groceries.

In the year leading up to March 2026, I directed AI agents through sixty-plus projects spanning product strategy, executive communication, organizational design, software architecture and investment analysis. The most useful shift in that work was deciding which kind of thinking a problem needed before asking for a recommendation. Choosing the mode of reasoning matters more than polishing the prompt.

A competitive analysis, for example, can be detailed and still begin from the wrong explanation for lost customers. Asking an assistant for more detail may produce a longer answer to the same mistaken question. I want it to help examine that explanation: what supports it, what would contradict it, and what I should investigate next.

I use five approaches: first principles, systems thinking, complexity thinking, analogy and design thinking. Each directs attention to a different part of a decision. The questions work in a meeting or on a whiteboard, too. With an AI assistant, I ask for a result I can examine and use in the next decision.

First principles: What is actually true here?

Imagine a company losing market share. Its leadership commissions a competitive analysis, then considers matching a rival's prices, adding three features and expanding into two channels. Those may be sensible responses. They also share an assumption: customers are choosing competitors because of price, features or access.

Before funding the response, I'd want to know whether departing customers are buying another product at all. Some might be leaving the category. Others might be unable to finish onboarding. Cancellation records, exit interviews and observed customer behavior could distinguish those explanations; a feature comparison cannot do so on its own.

First principles strips away borrowed assumptions and rebuilds from verified facts. In my experience, it is the most valuable of these thinking modes. An unexamined assumption can send a competent team toward the wrong goal, and an assistant can elaborate that assumption into a convincing plan. I want it to examine the foundation before building on it.

I'd give an assistant the available records and ask:

Before suggesting solutions, what assumptions are embedded in how I described this problem? For each one, identify the evidence we have and the observation that could contradict it. Mark missing evidence explicitly.

“Customers are price-sensitive” is an unhelpful answer if it merely repeats the premise. “The cancellation form records cost complaints but never asks whether the customer bought an alternative” identifies a gap we can investigate. That second answer is useful only if the supplied form actually has that limitation.

In a strategy session, spend fifteen minutes writing the assumptions on a board. Star the ones nobody has verified. Start with the starred assumption whose failure would most change the proposed investment. An assistant can help organize the list; someone still has to inspect the records or speak with customers.

Systems thinking: Who is affected, and what might they do?

Systems thinking follows connections. A decision changes the conditions under which other people act, and their responses can change the result we expected. I want to understand what a decision causes as well as what it immediately does.

Names are architecture decisions. When a company announces that it is using AI to improve operations, executives may hear efficiency, employees may hear layoffs, investors may hear growth, regulators may hear risk and customers may wonder whether their data will be safe. One announcement can set several feedback loops in motion. A useful stakeholder map directs us toward the people whose interpretations could change the outcome.

The Dartmouth proposal used “artificial intelligence” in its title on August 31, 1955, proposing research for summer 1956. That name invites people to compare a machine's abilities with human intelligence. I see naming as part of the system we design: the words shape expectations, and those expectations influence how people use, fund and govern the technology.

There is specific evidence about names in driving automation. Eric Teoh's study published in January 2020 reports a 2018 telephone survey of 2,005 drivers. Respondents were asked about two randomly selected names from five Level 2 systems. Across the behaviors examined, “Autopilot” was associated with the strongest belief that a secondary activity would be safe. These were reported beliefs, not measured driving behavior or a comparison of Tesla and Waymo regulation and public trust.

September 2026 product documentation: Tesla describes FSD (Supervised) as driver assistance requiring a fully attentive driver. Waymo describes its Driver as fully autonomous from pickup to destination, with areas mapped before operation. Those are different responsibilities for the person using the service. A name should help that person understand the responsibility they retain.

For a decision closer to home, I'd ask:

Map every stakeholder affected by this decision, including ones I haven't mentioned. Separate known concerns from predicted reactions. What should we ask each group before acting?

For the market-share example, a broad discount might reassure a departing customer while making an existing customer question what they paid. Sales incentives might favor signing discounted accounts that support teams struggle to serve. I would want the assistant to trace those possible connections and identify the people who could confirm or challenge them.

Complexity thinking: What will we need to learn as people respond?

A merger combines reporting lines and financial commitments with norms, incentives and informal networks. Financial assumptions can be wrong, and people also adjust to what they think the change means. A new approval process might centralize authority on paper while decisions continue through old relationships.

Complex systems exhibit emergence: people adapting to one another produce behavior that no individual part explains. Combining two cultures can create new alliances, resistance and working habits. Complicated work can often be decomposed into parts with reasonably stable relationships; in complex work, those relationships change as people respond. A leader who treats a merger solely as a logistics plan will miss much of what determines its outcome.

I would build fast feedback into the work from the start. Run pilots before rollouts, make surprises visible and decide how to revise the plan before the first surprise arrives.

For a proposed organizational change, I would design a thirty-day pilot with an explicit review point. Suppose two teams are testing a shared approval process. Track the time to reach a decision, work sent back for clarification, and decisions made outside the process. Interview the people handling the exceptions. Faster approvals would be a poor result if difficult decisions simply disappeared from the record.

Plan three possible adjustments before starting: clarify who can decide, reduce an unnecessary approval step, or pause expansion while a recurring exception is resolved. The actual thresholds should reflect the work and its risks.

The question for the assistant is:

What about this situation is genuinely unpredictable? Where might people's responses change the result, and what early observations would justify changing the plan?

I would reject a generic instruction to “monitor adoption.” I would keep a suggestion that identifies a particular behavior, how to observe it and which decision it could change. That gives the team something to do when the original plan meets an unexpected response.

Analogical thinking: Where has this been solved before?

Analogical thinking looks for solutions in fields that share a problem's structure. To try it, describe the problem without its usual nouns: variable inputs arrive unevenly, pass through several stages, and can create a queue when one stage slows down.

That description could fit an airport checkpoint or an emergency department. The comparison prompts useful questions about arrival patterns, staffing and where work accumulates. The purposes of the two systems determine which ideas can travel between them.

Clinical urgency sets the hospital priority. The historical 2012 Emergency Severity Index handbook describes assessing acuity first and then, for patients outside the high-acuity categories, considering expected resource needs. Preparation for an airport screening does not transfer into a rule for prioritizing patients. The shared queue structure helps us examine where work waits; the clinical purpose governs whom to treat first.

I've used cross-domain comparisons to solve problems in my own work. CPU cache architecture gave me a way to organize AI project context into material needed immediately, deeper reference and history. Database transaction theory helped me solve a code integration challenge. In each case, the useful structure came from outside the problem's home domain.

Specialists go deep. Generalists go wide. The leaders who create real breakthroughs go diagonal: they find structural parallels between fields that have never talked to each other.

Two problems share structure when the relationships between their components are similar, even if the components look nothing alike. A CPU cache and an information management system can share structure. So can a labor negotiation and a product design decision. The engineer who reads labor history, the hospital administrator who studies logistics and the CEO who watches how standup comedians structure a set have more shapes to recognize when a difficult problem arrives.

For a strategic decision, I'd ask:

What problems in completely different fields have the same structure as this one? Explain the shared relationships and the condition that would make each analogy fail.

In the customer-loss example, an assistant might compare onboarding to a queue with a bottleneck. That could be worth examining if customers are waiting at a shared approval step. It would be a poor explanation for customers who have decided the product no longer serves their needs. The analogy should help choose an observation, not decide the answer before it is made.

The technical details of the context example are in Keeping AI project context useful as it grows.

Design thinking: What does the person living with this decision need?

I learned this through a system migration. We had mapped the APIs, verified data integrity and exceeded the performance benchmarks. The technical work succeeded. We had neglected the people who would have to work in the new system.

The team had years of muscle memory around the old system's quirks. The new system required them to relearn workflows they performed dozens of times a day. Adoption was slow. Workarounds proliferated. The team quietly rebuilt the old system's quirks inside the new one.

The architecture was right. The design was wrong. It optimized for the system instead of the people.

I should have examined those working habits as part of the design. Which apparent quirk carried information that a user needed? Which shortcut let someone recover from an interruption? A migration can improve a benchmark while making a frequent task harder. Watching someone perform that task would give me evidence that another architecture diagram would not.

The same responsibility belongs in leadership decisions. An org chart can look efficient while making daily work miserable. A strategy that makes sense in the boardroom can leave the people executing it unable to explain their next decision. Technology can improve speed and integrity while disrupting the mental models people use to get their work done. Those are design failures.

Design thinking brings that person's work into the decision. An assistant can help prepare interview questions, compare observations or suggest a prototype. The affected person's experience belongs in the work itself.

If you were the person most affected by this decision, what would you need? Treat your answer as a hypothesis. What should we observe or ask that person to check it?

For the market-share problem, I might ask a departing customer to walk through the last task they tried to complete. If we only ask which feature they want, we have already narrowed what they can tell us. A walkthrough could expose an awkward handoff, a changed need or a problem the product was never meant to solve.

Taking one decision through the questions

The decision worksheet includes a blank version and a completed hypothetical market-share example. The exercise starts with a proposed price cut and ends with evidence to gather before deciding whether to test one.

QuestionUseful next step in the hypothetical case
Which assumption matters most?Check whether departing customers bought a competitor, stopped buying the category or failed to complete onboarding.
Who else is affected?Ask sales and support how discount rules would change their work; include existing customers in the discussion.
What response could change the plan?In a bounded pilot, track retention alongside support effort and exception requests; decide in advance when to adjust or stop.
Where might an analogy help?Examine a shared onboarding bottleneck, while checking whether the queue explanation fits the departing customers at all.
What does the affected person need?Watch a customer attempt the task that motivated the purchase, with their permission.

For each answer, I want a specific assumption, observation or decision. “Understand customers better” does not meet that standard. A recommendation to inspect a particular cancellation field can, provided that field exists and would help distinguish the explanations. Keep an unknown marked as unknown when the needed record is missing.

A routine operational choice may need only the first two questions. A new strategy with costly consequences may justify all five and several rounds of investigation. I choose the depth according to the stakes and the unresolved questions; completing five headings is not a measure of decision quality.

Using the framework with AI

I've packaged all five modes in the free, open-source synthesis skills, including cross-domain reasoning and guidance on matching the depth of analysis to the stakes. Reusable instructions make the method easier to bring into daily work. The skills introduction explains the collection.

September 2026 documentation: An Agent Skill is a package of reusable instructions and supporting resources. The Agent Skills specification describes loading those resources as needed. Current OpenAI documentation distinguishes explicit invocation from a client's decision to select a skill implicitly. I can invoke the method deliberately and then judge the resulting analysis using the same five questions.

The questions also work when typed into an ordinary AI conversation, written on a card or used to guide a meeting. Choose the form that makes it easy to examine a real decision. The method earns its place when it reveals an assumption, a connection or a human need that changes what you do.

For engineering examples and executable checks, the companion is Five checks for AI-assisted engineering decisions. The broader practices are described in synthesis engineering, AI-native project management and synthesis coding.

Bring an important decision to your next AI conversation. Choose the kind of thinking it needs, ask the corresponding question and use the answer to decide what to examine next. That is how I turn an assistant into a thinking partner.