AI strategy advisors love trendy buzzwords — can real industry details actually snap them out of it?

This explores whether feeding an LLM rich, industry-specific context can pull it away from its habit of recommending whatever strategic choice sounds most current. The short answer from the corpus is: not much. The most direct test ran 15,000 simulations across six LLMs and found that each model recommended the same side of every strategic tension it was given, such as exploring new markets versus exploiting existing ones. Changing the industry context shifted that bias by only about 11%. Simply reordering the answer options shifted it by 19% Do LLMs consistently favor the same strategic choices regardless of context?. In other words, the order in which the choices were listed mattered more than the facts about the business. The authors' explanation is that models recombine trend-coded vocabulary instead of reasoning about the situation in front of them.

This isn't only a problem with older models, and newer ones haven't fixed it. A strategy simulation benchmark found that mid-to-late 2025 frontier models scored below earlier models and below MBA students. They consistently grabbed short-term profit rather than investing in uncertain growth Do newer frontier LLMs actually make better strategic decisions?. The pattern looks less like a lack of information and more like a built-in leaning. That matches what other parts of the collection find about LLM behavior in general. Models develop coherent, stable preferences that grow stronger with scale Do large language models develop coherent value systems?. Their ethical stances act as fixed defaults set during training rather than judgments adapted to context Can language models balance competing ethical norms in context?. Strategic bias seems to be one more example of the same thing: a default that context barely moves.

The more hopeful thread comes from a different direction: structure, not context. When LLMs play strategic games, they drift further from rational play as the games get more complex. But when a step-by-step game-theory workflow guides them, they get close to optimal Do language models make rational strategic decisions in games?. The same pattern shows up in forecasting. Models do much better when the workflow handles the numbers and the contextual reasoning as separate steps rather than putting everything into one prompt Can LLMs actually forecast time series better than we think?. It also shows up in judging research novelty, where a three-stage pipeline (extract the paper's claims, find related work, compare) reached 86% agreement with human reviewers Can structured pipelines make LLM novelty assessment reliable?. Across these cases, the lesson is that you don't fix a model's defaults by giving it more to read. You fix them by changing how it is made to reason. For strategy work, that might mean making the model argue both sides of a tension before it recommends one, or scoring the options against explicit criteria for the industry.

There is also a human-side warning. In a 348-person experiment, people using an LLM to help evaluate strategic options considered a wider range of factors but made predictions that were no more accurate. They also felt more overloaded and less ownership of their decisions Does using LLMs actually improve strategic decision making?. So even when an LLM's advice seems richer and more tailored, that richness doesn't guarantee better judgment.

The takeaway you may not have expected is this: if listing the options in a different order moves an LLM's strategic advice more than describing your industry does, then context is the wrong lever. The levers that work are the structure of the reasoning process and checks on the answer framing itself.