After a few years of monitoring and studying corporate AI implementations, I’m still puzzled by one thing: most of them begin with a discussion of the model being used. Should we go ahead with Copilot, considering that Microsoft is so firmly consolidated in our company that it has become a sort of lingua franca for everything? Should we try GPT, since they were pioneers? Or Gemini, that has all the Google power behind it? Or Claude, that seems so fashionable now? How about Grok? Or, as is happening in many American companies, dare to explore Chinese models such as Deepseek and Qwen.
But what if this decision was not so crucial? After all, when we consider a corporate implementation, the model looks a bit like the microprocessor in a computer: important, yes. The larger the better? Maybe. But just one of the pieces, and not necessarily the most important or strategic one.
In fact, what we are already witnessing in the American corporate landscape is precisely that: the choice of a model is becoming a matter of economic optimization, instead of some sort of ideological commitment. Architectures are becoming very different from the initial “this company runs on GPT,” and there are many reasons for that (besides the cost per token).
First of all, a company does not need the same powerful, frontier model for each one of their queries, and using one is often overkill and can become extremely expensive. Simple queries can be routed to cheaper models when a good enough model is sufficient, while other, more complex questions or tasks can be escalated to the more sophisticated ones. Orchestrators such as the RouteLLM project from Berkeley hints precisely at that, and can save lots of money while preserving the integrity of the answers, and a reasonable cost structure.
Deepseek is an interesting case of a company that has positioned itself in a clear way to take advantage of that: extremely competitive token economics, to reinforce the idea that, even for an American company, models can be extremely substitutable, almost commoditized. And when models become easy to substitute at the API layer, value starts to naturally migrate to higher layers in the stack.
The goal of putting “the biggest model available” at the fingertips of your employees is becoming less and less important, and concepts such as the dreaded tokenmaxxing are now being seen as patently absurd. And the company that seems to be interpreting this trend better is no less than Microsoft, the undisputed king of corporate IT (as they used to say about IBM long time ago, “no CIO or CTO ever has been fired for buying Microsoft!”) The company is explicitly positioning small language models as the best option for domain-specific, highly focused tasks or environments, in which they can be appropriate to yield a strong performance with not too stringent computational resources and a high control over the data. They have also produced models adapted to specific industries using their Phi family, not trying to beat large models with small ones, but proving that model size should be a function of task complexity, instead of a matter of corporate prestige.
Let’s try, then, to approach corporate AI as something that starts with general intelligence, follows with institutional context, and ends in institutional learning. Trying to produce the first one seems not only impossible, but also completely anti-economic and out-of-scope for anyone who’s not an AI company. But the second layer consists of things such as a company’s objects, documents, rules, ontology, relationships and operating history. And the third one is even more interesting, since it is made of what actually worked: consequences, evaluations and feedback, the so-called loops. These two latter layers, not the first one, are where companies can really obtain and compound true differentiation and optimization.
Anthropic specifically mentions the improvements companies can achieve by focusing on context engineering, on managing the surrounding state from tools to instructions, external information or history, instead of just becoming obsessed with improving the prompt or the model.
Imagine my case: I work at a big university. My professors and even my carefully selected students are producing an incredible amount of documents for every course, many of them with the corresponding evaluation associated as feedback, be that grades, peer reviews, etc. Couldn’t that become a significant part of a specific context corpus with which we could make strategic decisions, and even derive a competitive advantage from that differentiates us from other universities? This idea goes along with what Satya Nadella said on companies owning their own learning and loops, instead of just buying a big, fat LLM and using it pretty much in the same way as other, non-related companies in other, non-related industries are using it.
If you think about it this way, the real asset is not the model, but the loop. Imagine two universities using the same model: will they become equally smart institutions? What if one of them brings decades of accumulated decisions, faculty expertise, pedagogical experimentation, student outcomes, organizational culture and feedback? When you are able to capitalize on all these assets, you can start with the same commodity model, but you will probably end up with a totally different institutional intelligence. If you are not able to do that, you will be, essentially, renting the same brain. But once you are able to build that architecture, the LLM itself becomes merely one component inside it. And a replaceable one. The question, therefore, is not “does your company have access to the latest model,” but more like “if you were to change your model tomorrow, how much of your institutional learning will stick with you and how much will you lose? If you think you will be losing a lot of that valuable information, then the company that sold you the model owns way too much of your institutional intelligence.
It is, in essence, a matter of institutional sovereignty. What’s the value of that?