Why LLM’s are really poor at strategy

1. 1. The bias is already there

Ask an AI model to generate an image of a watch and there is a good chance it will show roughly ten past ten. Claude, Gemini and ChatGPT all reproduce the same familiar pattern because it is deeply represented in their training data. Something similar happens when you ask these models for strategic advice.  

In their Harvard Business Review article, Angelo Romasanta, Llewellyn D.W. Thomas and Natalia Levina describe more than 15,000 trials in which LLMs were tested on seven fundamental strategic trade-offs. The models consistently leaned toward certain choices, such as differentiation over cost leadership, augmentation over automation, long-term over short-term thinking and collaboration over competition. 

The researchers call this strategy trendslop: advice that sounds modern and sophisticated, but is not necessarily right for the specific company or situation. 

Even when the researchers provided richer organizational and industry context, these biases remained remarkably persistent. 

When the models were not forced to choose, another problem appeared. They often recommended both sides of the strategic trade-off. The researchers call this the hybrid trap. That may sound balanced, but strategy requires choices.¹

The implication is important: Don’t rely on LLM’s when it comes to Strategy. 

2. Context helps, but it is not enough 

The obvious solution is to give AI more company information. That helps, but the research shows that more context alone does not remove the problem. The reason is simple. Giving a model documents does not mean it understands the business.

To answer a question such as “What happens if this supplier can no longer deliver?” the system needs to understand the relationships between supplier, material, product, plant, customer, inventory and contract.²

It also needs to understand the logic behind those relationships. Can the supplier be replaced? How long does qualification take? Which customers have priority? What happens when inventory drops below a certain level? This is where business understanding becomes more important than simply adding more text to a prompt.

3. Intelligence first 

At verticallm.ai, we therefore start with three building blocks.

The first is industry context. Every industry has its own terminology, structures, regulations, processes and relationships. This creates the foundation the AI needs to understand the domain.

The second is company context. This includes contracts, policies, procedures, strategies, supplier agreements, customer requirements and internal terminology. It explains how your company operates within that industry. The third is structured operational data. Products, suppliers, customers, orders, inventory, assets, locations and transactions show what is actually happening inside the business.³

The most important part is how these three layers are connected through business logic and relationships. AI should not only know what information exists. It should understand how the business works, what depends on what, which rules apply and what happens when something changes.

That is the difference between generic AI and Embedded Intelligence.

A general-purpose model knows a lot about the world. It does not automatically know your company. For business-critical questions, the priority should therefore be your own data, your own language, your own relationships and your own logic, with AI running on top of that intelligence.