When companies all ask AI for strategy advice, do they end up copying each other into one big industry-wide herd?

This explores whether many firms asking LLMs for strategy advice could end up following the same playbook. The corpus has strong evidence for the first half of that worry: the models do converge. It has much less on the second half, whether industries actually become more alike as a result. In the clearest study, six different LLMs across 15,000 simulations picked the same side of every strategic tension tested Do LLMs consistently favor the same strategic choices regardless of context?. The surprising number is that industry context changed the recommendation only 11% of the time, while simply swapping the order the options were listed in changed it 19% of the time. So the advice responds more to how the question is worded than to the business it is about. The authors' reading is that the models are recombining fashionable strategy vocabulary rather than analyzing the situation.

The convergence also leans in a particular direction. A strategy simulation benchmark found that frontier models from mid-to-late 2025 scored below earlier models and below MBA students, because they kept choosing immediate profit over uncertain investments in growth Do newer frontier LLMs actually make better strategic decisions?. Taken together, these two findings describe a shared bias toward short-term exploitation, not just a shared answer. Newer models don't fix it, and the evidence suggests the bias may be getting stronger. If many competitors consult similar models, the risk isn't only that they look alike. It's that they all underinvest in the same way at the same time.

Recommender systems research offers a useful lateral view of how convergence spreads. Studies of product recommendation networks show that whether opinions converge or diverge depends on the type of recommender. 'Frequently bought together' and 'co-viewed' links pull in different audiences with different expectations, which produces different rating patterns Do different recommender types shape opinion convergence differently?. The lesson carries over: homogenization isn't automatic. It depends on how the advice reaches people and who is receiving it. A firm that uses an LLM to enrich its own analysis is in a different position from one that takes the model's recommendation as the decision. In recommendation tasks, LLMs do better as content enrichers feeding a specialized system than as direct recommenders Does LLM input augmentation beat direct LLM recommendation?.

There is also a reason people might not notice the convergence. Users rate AI responses higher when they include more citations, even when those citations are irrelevant Do users trust citations more when there are simply more of them?. Advice that sounds well supported can feel tailored even when it is generic. The prompt-sensitivity finding Do prompt techniques work the same across all LLM tiers? suggests one partial lever: how you phrase the question changes what comes back. But phrasing is a fragile kind of diversity, and it isn't the same as the model actually reasoning about your situation.

The gap is that the corpus shows models converging in controlled simulations. It doesn't yet contain field evidence of real industries becoming more alike because of LLM advice. The mechanism is well documented. Whether it plays out in actual markets is still an open question.