AI in Strategy
Where the machine ends and the strategist begins.
I built a new tool, maybe the most important one for strategy work in the age of AI. No cap.
It’s called AI in Strategy, and it tells you where AI belongs in your strategy work, and where it doesn’t. It’s available as a website you can use right now, and as a Claude skill you can install and run on your own projects. More on how to use both at the end. It exists to answer a question most strategy teams are still working out: when you sit down to do strategy work, what should you actually use AI for, and what should you do yourself?
That question is less obvious than it sounds. The answer depends on the type of strategy being developed, the broader context of the work, and the phase you’re in. In other words, it depends on what kind of thinking the work actually requires. AI is an outstanding thinking companion for some parts of the journey. For others, it’s a liability. Leaning on it in the wrong places can lead to strategic misdirection that takes months to surface and longer to undo. Knowing where the machine ends and the strategist begins is the new meta in strategy work.
The reasoning problem
Charles Sanders Peirce, a 19th century American philosopher who didn’t get nearly enough credit during his lifetime, distinguished three modes of reasoning. Induction, finding patterns in data. Deduction, tracing what follows from a given premise. And abduction, inferring the best explanation when the situation is genuinely uncertain.
Strategy runs on all three. And AI is not equally capable in all three. The core principle here is that LLMs optimize for statistical plausibility instead of epistemic validity. That distinction explains everything that follows.
Induction is where AI shines. It’s trained on next-token prediction across billions of texts, which makes it, at core, a pattern machine, and more specifically, a consensus machine. It aggregates across massive textual environments, smooths inconsistencies, and produces outputs shaped by what’s already culturally stabilized. It’s strongest when signals are redundant across sources, weaker when they’re sparse, emerging, or contradictory. In stable territory, it’s genuinely excellent. At the edges of what’s known, it regresses toward the mean.
Deduction is more complicated. It’s the mode that tests whether things hold together: if our theory of market is right, what follows? If we commit to this positioning, what does it rule out? If this capability is genuinely distinctive, what can we build on top of it? AI can follow these chains when the structure is clear and the problem resembles training data. But it’s imitating the surface form of reasoning, not running a truth-preserving process. It optimizes for plausibility over validity. Errors accumulate. Long chains need human verification, because AI will produce a coherent-looking conclusion even when the logic has already broken down somewhere in the middle.
Abduction is the mode that drives discovery, reframing, and strategic insight. When a strategist looks at market signals and proposes a new premise about what’s actually happening, that’s abduction. It’s the leap that introduces a new way of seeing. Abduction is also where most people get the picture of how LLMs work wrong. The two camps (”AI is great at abduction” and “AI can’t do abduction”) are both right, just about different parts of it. Abduction splits: there’s generation (”what could explain this?”) and selection (”what is the best explanation?”). AI is strong at the first. It generates hypotheses, connects distant domains, produces reframes. It is weak at the second. It has no skin in the game: no consequences, no feedback loop from reality, no cost for being wrong. It produces possibilities. It doesn’t close the loop between hypothesis and reality.
In strategy, the valuable move is usually not generating more options. It’s choosing the right one under uncertainty, with consequences. That’s selective abduction and something that LLMs can’t do.
Where strategy lives
This matters because strategy work isn’t one thing. It runs across a cognitive system that now includes both humans and AI. Different parts of that system are suited to different tasks.
The AI layer is fast, scalable, and pattern-driven. Research synthesis, competitive mapping, scenario construction, generating hypotheses and variants, checking consistency, drafting. Enormous range, no stakes.
Then there’s the gut layer. The read that comes before the argument. The experienced strategist sensing that a direction is wrong before being able to articulate why. Feeling what a client will actually act on, versus what they’ll agree to in the room. This layer is built from years of exposure to how strategies actually play out. It can’t be prompted into existence. N. Katherine Hayles, the cognitive theorist whose work this is partly built on, calls it implicit cognition: the tacit knowledge that accumulates through embodied participation, not text processing.
And then there’s the conscious thinking layer, where decisions get made. Selecting the strategic premise. Committing to the direction. Deciding what to give up. This is abductive commitment: the leap from a field of possibilities to a chosen direction, owned by someone with skin in the game.
The failure mode is when these layers collapse. The team uses AI to generate five positioning territories, picks the one that sounds most compelling, and presents it to the client as a strategic recommendation. A workshop synthesis lands in a deck and gets treated as organizational insight. A competitive analysis shapes a market entry decision nobody examined. The outputs are indistinguishable from the real thing. You don’t know you outsourced the judgment until the work doesn’t hold.
To be fair, the same collapse happens without AI. A strategist who reads markets, culture, or human behavior purely through numbers, or applies linear thinking to genuinely complex situations, or mistakes past success for a map of the future, produces the same results by different means. AI just makes it faster and harder to see.
The skill
AI in Strategy is built on this model. It covers 28 strategy types, from brand and corporate strategy to community, narrative, and talent, and ten clusters of epistemological activity that run through all of them. By that I mean the distinct types of thinking and knowledge production that strategy requires: things like sensemaking, ecosystem understanding, choice architecture, and the actual commitment to a strategic direction. Each cluster has a different AI profile. For any task or challenge, the skill runs four criteria: what reasoning mode is required, how long before you’d know if the output was wrong, whether the task needs inside knowledge that AI doesn’t have, and whether a wrong answer is recoverable.
Any one criterion in the human-critical zone makes the verdict human-critical. They don’t average out.
In practice, it means the skill can tell you: use AI hard in the research and scanning phase, bring your own judgment to the synthesis, don’t let AI make the positioning call, and stress-test the direction you’ve chosen before you build anything on top of it.
How to use it
The website is the quickest way in. No account, no setup. Describe your strategy work and the situation you’re in, and it will run the diagnostic and tell you where to use AI and where to stay in the driver’s seat. Good for one-off questions, specific challenges, or getting oriented before a project starts.
The Claude skill is for heavier use. Install it once into Claude and it becomes a persistent thinking partner across your strategy work. Bring it a research synthesis you’re not sure how to interpret, a competitive analysis you want to pressure-test, a positioning decision you’re about to make. It asks a few questions to understand the situation, runs the four criteria, and gives you a specific read, not a generic framework.
Both are useful for consultants figuring out where to bring AI into client work, for in-house strategy teams building new AI workflows, and for anyone who has started using AI in strategy work and wants to be more deliberate about it.
A final note. This is a framework, not a rulebook. It is based on a theory of the cognitive capabilities of both humans and LLMs, and on definitions of current models’ boundaries. How you use AI in strategy involves judgment that no diagnostic can fully anticipate. Stay accountable for the outputs. Trust your read when something doesn’t feel right. The whole point of the tool is that human judgment matters, here too.
