Why do AI-generated strategies all sound the same?
The short answer
AI-generated strategies resemble one another because a language model reproduces the average of what has already been written about a question. The range of results depends more on the procedure than on the model: feeding in company data and several creative methods produces different proposals than asking an open question.
Three models, the same question, almost the same answer
The test takes ten minutes. Ask three language models the same strategy question about your company, put the answers next to one another and mark what differs.
You will find the same structure: market analysis, differentiation, core capabilities, roadmap, metrics. What you will not find is a decision.
That is not an operator error. A language model reproduces what has already been written about a question, weighted by frequency. For an open strategy question, that is the average of all strategies ever published. An average is the opposite of differentiation.
One model lifts the individual idea and lowers collective diversity
In a preregistered experiment, 300 people wrote short stories, some with access to AI-generated ideas. 600 independent raters judged the AI-supported results as more novel and useful, especially for previously less creative writers. At the same time, the texts became measurably more similar to one another: about eleven percent more overlap than without AI (Doshi & Hauser, Science Advances, 2024).
The authors call this a social dilemma. For the individual, using AI is rational because the result improves. For the group, the differences from which competition emerges shrink.
The same pattern appears in product ideas. In a comparison between students and GPT-4, thirty-five of the forty highest-rated ideas came from the model, while human ideas were more novel (Girotra et al., Wharton, 2023).
Idea variance is a procedural question
The same research group tested 35 formulations of the same task and measured the spread of generated ideas. With the best wording, the model almost reached the range of human groups; with the worst, it was far below (Meincke, Mollick & Terwiesch, Wharton, 2024).
The practical finding is that the range is decided less by model choice than by the procedure used to query it.
For mid-sized companies, a second effect matters. The question itself is often already average. It uses the market’s shared language: the same segments, cost drivers and slogans from the same industry publications. The model then answers generally to a general question with general material.
The consequence is concrete: ask an open question and you get the average. Feed in company data, the existing strategy and several creative methods in one run, and the proposals relate to the company’s own starting point.
AI fails in strategy simulation where strategy begins
Thirty-four language models played a Harvard strategy simulation in which a company must balance exploitation of the existing business with investment in a new technology under delayed and noisy feedback (Allen & McDonald, Strategy Science, 2026).
Reasoning models from the turn of 2024/2025 outperformed historical MBA cohorts. Frontier models from the second half of 2025 remained below both earlier models and students.
More important than the ranking is the documented weakness. The models struggled with strategic uncertainty and systematically tended to exploit the core business instead of investing in future growth. That is the trade-off for which a company has an executive team.
What follows for strategy work
Models are strong where completeness matters: pre-structuring a question, producing counterarguments, and checking whether a list of measures matches the goals. These tasks are tedious and often the first to disappear in practice.
The average also remains in the room after the tool is closed. People can inherit AI biases and repeat them later without the model (Vicente & Matute, 2023). A model therefore acts as an amplifier of the assumption with which someone entered.
BizzPlAI starts one step later as a Strategy Execution System. It generates initiatives from the company’s data, existing strategy and multiple creative methods in one run. The industry average is not the input; the company’s numbers, customer structure and constraints are.
Run the test anyway. If three models suggest the same strategy three times, you have learned something about your industry and nothing about your company. The interesting part begins with the question why the obvious answer may be wrong for you.
Common follow-up questions
Can better prompts avoid the similarity of AI answers?
Partly. In a test of 35 prompt variants, the best almost reached the idea variance of human groups, while the worst was far below it. The idea space grows, but it is not unlimited. Own data and several creative methods in one run are more effective than wording alone.
Does this make artificial intelligence useless for strategy work?
No. It is strong in pre-structuring, counterarguments and consistency checks, where completeness matters more than originality. It becomes weak at decisions under uncertainty and trade-offs between today's earnings and future growth.
How does idea generation with company data differ from an open prompt?
An open prompt retrieves the average of what has been published on a question. A procedure that feeds in company data and the existing strategy produces proposals related to that starting point. The difference lies in what the model sees while answering.
Sources
- Generative AI enhances individual creativity but reduces the collective diversity of novel contentScience Advances (Anil Doshi, Oliver Hauser), 2024
- Ideas are Dimes a Dozen: Large Language Models for Idea Generation in InnovationMack Institute, Wharton (Karan Girotra, Lennart Meincke, Christian Terwiesch, Karl Ulrich), 2023
- Prompting Diverse Ideas: Increasing AI Idea VarianceMack Institute, Wharton (Lennart Meincke, Ethan Mollick, Christian Terwiesch), 2024
- How Well Can AI Do Strategy? Empirical Benchmarking Using Strategy SimulationsStrategy Science (Ryan Allen, Rory McDonald), 2026
- Humans inherit artificial intelligence biasesScientific Reports (Lucia Vicente, Helena Matute), 2023
See how strategy arrives.
In 30 minutes, we show you how to steer your initiatives instead of merely managing them.