Ask a general chatbot to explain the difference between a TAM and a SAM and it will give you a clean, correct answer. Ask it to build the market-size chapter of a strategy deck and it will eventually, confidently, hand you a number with no source behind it. Both answers come from the same system. The difference isn't intelligence: it's the type of work.
Strategy is the sharpest example of where general AI hits its limits, because it combines every failure mode at once: exact claims that must be sourced, multiple documents that must reconcile, a storyline that must survive from chapter to chapter, and consequences for being wrong. Here's what's actually happening.
The number problem
Language models don't calculate. They predict the next word. When a model "sizes a market," it's not measuring anything: it's generating the most probable figure given everything it has seen. For a well-known industry, that sometimes lands in the right ballpark, because the pattern is common. For a niche market, a new geography, or a segment nobody has written neatly about, the probability is nowhere near good enough. The model will produce a plausible-looking figure that is simply wrong, and it will do it with the same confidence it used for the correct framework explanation two messages earlier.
This is the difference between a chatbot and an analysis platform. A proper tool computes: it takes your inputs, applies the model structure, and produces a number you can trace to the assumptions. A chatbot guesses the shape of a correct answer. For a client deck, guessing isn't good enough.
The multi-source problem
Strategy work is synthesis. The client's data, the market research, the competitor analysis, the analyst reports, the industry filings: the whole job is making them agree into one storyline. Cross-checks are the point; a market size that contradicts the client's own revenue numbers is exactly what a good engagement catches.
A chatbot has to hold all those sources in one context window to synthesise them, and as covered in why AI loses the plot on real projects, that window is finite. Compress the sources to fit and you lose the detail where the contradiction hides. Keep the detail and the window overflows and earlier sources get dropped. Either way, the synthesis suffers. Real analysis platforms don't paste documents into a window: they keep the data structured, and "the model" only ever looks at a scoped slice for a specific task.
The storyline problem
Strategy isn't a list of facts. It's an argument: a storyline where every chapter builds on the ones before it. The market chapter sets up the opportunity; the competitive chapter sets up the wedge; the financial model sets up the recommendation. If the model loses the thread between chapters, the deck stops hanging together.
A chat model is stateless. It has no persistent understanding of "this engagement" beyond what's in the current window. Close the conversation and the storyline is gone. Even within one long conversation, the cumulative build-up of context degrades accuracy: the model starts repeating itself and agreeing with the last thing said. You cannot run a multi-chapter strategy engagement on that. The storyline has to live outside the model, in a structure that keeps each chapter connected while the model works on the piece in front of it.
The provenance problem
Strategy isn't just getting the argument right. It's being able to show where every claim came from. A market size needs a source or a stated assumption; a recommendation needs evidence; a number a client challenges needs a trail back to something they can open.
A chatbot produces answers, not trails. It can tell you what it thinks the market is, but it can't construct the evidence chain underneath it. The moment a client asks "where did this number come from," the chat answer stops being useful. This connects directly to the hallucination problem: a model that invents a citation for a court case will equally invent a supporting source for a market figure. When the output has to stand up to a partner review, generation without provenance is a liability, not a tool.
Where it goes wrong, concretely
It's worth being specific, because "strategy is hard" is too vague to act on. A general chatbot on a real engagement will trip on all of these:
- It will invent a market number to fill the size cell, because the real data was too much to hold in the window.
- It will mix up the client's own data. Ask it to build the financial model and it will sometimes use figures from the wrong year, because the dates blurred in the compressed context.
- It will fabricate a competitor fact (a funding round, a product launch, a price point) that sounds right and isn't.
- It will agree with the last thing you said. Long chat threads drift toward the latest prompt; if you suggest a number is wrong, the model will often agree, even when the original was right.
None of these are exotic. They're the ordinary behaviour of a prediction system doing strategy-shaped work. Each one individually is catchable by a careful human; together, across a full engagement, they become a game of whack-a-mole.
What strategy AI actually looks like
The useful version of AI in strategy isn't a chatbot that "builds the deck." It's a set of narrow, reliable systems that work on structured data, each one small and verifiable, with the model used where it's genuinely good:
- Scoping and structuring. Turning a vague request into a clear workplan with chapters, deliverables and a timeline. The model is excellent at this because it's pattern recognition with a defined output.
- Drafting and synthesis. Writing the chapter, explaining the logic, connecting the evidence to the storyline. This is where general AI genuinely shines.
- Research and flagging. Gathering source material, surfacing relevant facts, spotting anomalies in the client's data. The model surfaces the candidate; a human decides what survives.
- Not the numbers. The market sizes, the financial models, the figures that drive the recommendation: those stay in software that computes and traces, because that's what they require.
The rule that captures all of it: let the AI do the structuring, the writing, and the researching: never the measuring. The moment a model is asked to produce a number it should be computing, it starts guessing, and guessing doesn't survive a partner review.
That's the design philosophy behind a strategy platform: structured data that computes correctly, a storyline that lives outside the model, tasks that are scoped small, and a human in the loop where judgement and liability live. The model is a force-multiplier for the parts it does well, not a replacement for the parts it can't.
Want to see what that decomposition looks like? The interactive demo shows a strategy engagement broken into scoped, structured steps instead of one giant prompt, and the scoping tool is the fastest way to feel how a big engagement becomes small, manageable pieces.
Get new posts and free tools
Join the list: one email when we publish. No spam.