When everyone has the same AI, what makes your company smarter?

After years of observing corporate AI implementations, it remains puzzling that most companies start by discussing the specific model they plan to use. Companies like Microsoft, Google, and Chinese firms all have popular models. However, the choice of model may not be as critical as it seems.

When considering a corporate AI implementation, the model is like a computer's microprocessor - important, but only one piece of the puzzle. The model is crucial, but larger isn't always better. In fact, the choice of model is becoming more about cost optimization rather than ideological commitment. Models can be tailored to specific queries, with simpler questions using cheaper models and more complex ones using more sophisticated ones.

Orchestrators like RouteLLM, developed by Berkeley, can route queries to different models based on their complexity, saving money while maintaining the quality of answers. Chinese AI companies, like Deepseek, are leading the way in making models easier to substitute, making them almost commoditized. This commoditization means that value can naturally shift to higher layers in the AI stack, not just the model itself.

Microsoft is a prime example of a company adapting to this trend. They're positioning small language models as the best option for domain-specific tasks, not just for their prestige. These small models, like those in Microsoft's Phi family, can provide strong performance with lower computational resources and better control over data.

True corporate AI differentiation comes from the second and third layers - the company's context (its objects, documents, rules, ontology, relationships, and history) and the loops of consequences, evaluations, and feedback. These layers allow for true optimization and competitive advantage. For instance, a university could leverage its vast amount of documents and evaluations to create a highly specialized context corpus, providing strategic advantages over other institutions.

This focus on context engineering - managing the surrounding state of tools, instructions, external information, and history - can lead to significant competitive advantages.