Most arguments about AI and strategy pick one of two sides. One treats large language models as general strategic reasoners whose continued scaling will absorb most knowledge work. The other points to conspicuous errors and concludes that LLMs have little place in serious decision-making. Both ignore the structure of the work.
Strategy is not one cognitive task but a collection of activities: searching and summarizing existing knowledge, constructing causal explanations, imagining unprecedented possibilities, committing resources, and acting under uncertainty. An honest account of AI’s value begins by separating these activities.
That task-based approach is the center of my paper Mean Articulation Machines.
Where LLMs provide genuine leverage
Large language models are at their best when the task rewards broad association with documented knowledge. They can:
- search and summarize large bodies of material;
- compare established frameworks and options;
- translate among technical, managerial, and public vocabularies;
- produce clear first drafts in a specified professional form;
- generate variations on known approaches;
- expose assumptions by presenting alternative formulations;
- organize background research for further investigation; and
- act as patient, inexpensive interlocutors during early analysis.
Strategic work contains substantial amounts of articulation, synthesis, comparison, and preparation. Compressing that work changes the economics and speed of analysis.
Even when an LLM cannot originate a breakthrough hypothesis, it can be useful after a person proposes one. It can summarize relevant technologies, identify analogous cases, structure a test plan, draft communications, and reveal the conventional objections the proposal will encounter.
Dismissing such systems as “mere autocomplete” misses what association across an immense corpus, expressed in flexible language, can do.
Where fluent output becomes dangerous
The same architecture that produces impressive breadth also shapes the boundary of reliable use.
LLMs learn from patterns in their training data and generate responses that continue those patterns. They are consequently strongest where the relevant knowledge is documented, repeatedly represented, and close enough to established combinations. Their fluent output can conceal difficulty when a task requires something else.
Strategic problems become less well matched as they depend more heavily on:
- causal explanations rather than linguistic association;
- counter-consensus reasoning;
- recognizing that a widely repeated belief is wrong;
- constructing a genuinely new theory;
- sustained deductive consistency;
- long-horizon planning under changing conditions;
- confidential, local, embodied, or tacit knowledge; and
- responsibility for a commitment whose consequences unfold in the world.
An LLM can state that conventional wisdom may be wrong. That is different from identifying which established belief must be rejected, why it must be rejected, and what unknown theory should replace it.
The problem of the mean
The phrase “mean articulation machine” identifies a central tendency. LLMs are trained to produce likely continuations over distributions of human expression. They can articulate the center of documented knowledge with remarkable range and polish.
Strategy, however, often depends on ideas that are valuable precisely because they are rare. A sustainable advantage cannot be built entirely from the same high-frequency recommendations available to every competitor using similar models.
This does not mean that every contrarian idea is good. Most are not. It means that an architecture optimized around established patterns should not be expected to supply the epistemic vigilance, causal understanding, and courage required to depart from those patterns responsibly.
There is also an organizational risk. When many firms ask similar systems for best practices, those systems may accelerate convergence on the same language, frameworks, and candidate actions. The output can improve the average quality of routine analysis while narrowing the space of strategic difference.
Exploration and exploitation
The familiar distinction between exploitation and exploration marks the same boundary.
Exploitation refines what is already known: applying established routines, comparing familiar options, improving efficiency, and extending successful patterns. Exploration searches beyond current routines and may require seeing the problem differently, rejecting a shared assumption, or acting before reliable evidence exists.
LLMs are naturally strong instruments of exploitation through interpolation. They make documented knowledge easier to access, combine, and express. They can also support exploration by expanding the set of examples a person considers or by reducing the cost of testing an articulated hypothesis.
But supporting exploration is not the same as originating the rare conceptual leap on which exploration may depend.
A better division of cognitive labor
The organizational question is not whether to use AI but which parts of a strategic workflow can be delegated, which require verification, and which must remain under human judgment.
A useful sequence is:
- Locate the task. Is the work primarily search, synthesis, comparison, explanation, prediction, theory construction, or commitment?
- Identify the missing knowledge. Is the relevant information documented and available, or local, confidential, tacit, and still unknown?
- Use AI where the fit is strong. Let models perform the broad search, drafting, formatting, and conventional analysis at which they excel.
- Test the output. Verify sources, inspect causal claims, look for omitted alternatives, and compare model responses where failure matters.
- Preserve human ownership of judgment. People remain responsible for the interpretation, contrarian hypothesis, commitment, and action.
This is neither an anti-AI position nor a defense of every existing professional role. The point is to match a powerful tool to the cognitive requirements of the task.
The practical conclusion
LLMs can map much of the terrain of what is already known. They can make the traversal of that terrain dramatically faster. For most organizations, failing to use them for appropriate tasks will become a competitive disadvantage.
The error is to assume that because a system can articulate the terrain, it can also determine which undocumented path should be created next.
At the boundary of novelty, causal judgment, tacit knowledge, and counter-consensus action, human expertise remains indispensable. The strongest use of AI in strategy is a deliberate partnership: machines expand access to documented knowledge, and people stay accountable for the judgments that go beyond it.