This article was originally published on AI4Business.
Behind the choice of model, several control points are emerging: from models to post-training, from inference to routing, up to the rights over the information and meta-information generated during operations. Understanding who can observe, store and use them – and for what purpose – is becoming a new strategic decision
We are used to approaching the relationship between businesses and artificial intelligence by focusing mainly on one question: which model or platform should we choose? It is an understandable question, of course. But could it be the wrong one?
Adopting AI is not simply an upgrade to a company’s digital toolbox, but rather a deeper change: the way knowledge is produced, work is coordinated, decisions are made and human judgment is preserved.
The strategic question therefore becomes not only which AI we use, but how we want to design the company when AI becomes part of its operating architecture. And it is a particularly important question today, because we are still in the first moments after what we might call the “Big Bang” of generative artificial intelligence. The supply ecosystem, regulation, geopolitics and the competitive structure are all evolving at the same time. The fundamental components and, above all, the control points of the new market have not yet settled into lasting combinations.
For this reason, a company should not design its future by betting on which technological configuration will prevail; it should rather preserve its ability to change.
Behind the choice of an AI lie at least six decisions
What apparently looks like a single choice — which model or platform to place at the center of the AI architecture? — actually contains at least six different and partly interdependent decisions.
The first concerns which models and capabilities to use: proprietary APIs, open-weight models, specialized models or a portfolio of different models.
The second concerns who retains control of the model once it has been adapted to the company’s needs, and whether the weights and checkpoints resulting from fine-tuning or post-training remain tied to the provider or can be controlled and transferred by the customer.
The third concerns who provides the post-training infrastructure.
The fourth concerns where inference is run: at the provider, through a hyperscaler, via a specialized operator, in a private cloud or on-premises.
Even the first decision does not necessarily mean choosing a single model destined to become the company standard. Models differ in capability, cost, latency and failure modes. A model may be excellent at one task and less effective at another. For this reason, one possible evolution is the mixture of models: instead of electing one model as the universal default, different tasks are routed to different backends.
The fifth decision concerns routing: who decides which model should receive a given request? Routing can be controlled by the company, entrusted to a third-party service or left to the provider. Whoever manages routing also holds a particular informational position: they can observe which models work best on different tasks, where they fail and when an alternative is needed.
The sixth decision, perhaps less obvious but strategically fundamental, concerns who can observe, store and use the information and meta-information generated during AI operations, and for what purpose. And it is precisely on this last point that one of the most interesting issues for businesses is opening up.
Where is my data?
Every interaction with an artificial intelligence system obviously produces an output, but not only that: it can also generate information about what happens afterwards, when the workflow makes these outcomes observable. A response can be accepted, modified, rejected or repeated. The model may need to be asked again to perform the task, an error may be corrected by a person, a result may be escalated or turned into an action, and the professional may ultimately make a different decision from the one suggested by the system: each of these steps can generate further information when it is observed or recorded.
In generative workflows and, even more so, in agentic systems, this trace becomes particularly rich because it can show where the system failed, how the result was corrected and which subsequent behavior produced a better outcome.
The question, therefore, is no longer just “where is my data?”, but becomes: who can observe what happens while my organization uses AI? Who can store it? Who can use it? And to improve what?
The boundary between use, observation and reuse
This distinction matters because the contractual boundary is far more granular than a generic right to use the information generated during use. Observing information, storing it and using it to improve a model or a service are in fact different permissions. Agentic systems add further complexity because they generate an intermediate state that is carried from one call to the next, and part of this state may remain opaque or tied to the provider. If that state is not portable, the ability to route the next call to another model may also be reduced: freedom of routing therefore also depends on who controls the intermediate state.
Here a fundamental difference emerges compared with other technology decisions: a model can be replaced, inference can be moved, routing can be redesigned, and a post-training infrastructure can, at a cost, be recreated elsewhere. Information already absorbed by another organization, however, cannot be called back in the same way. For this reason, the rights over the information generated during AI operations should not be considered a mere appendix to technology choices.
The AI company must be designed to change
This does not mean that every company should break down its AI architecture to the maximum. Integration has value; especially for smaller companies, choosing a provider able to absorb technological and operational complexity can be a perfectly rational decision. The point is different: the company must know which decisions it is bundling in the hands of a single provider, which control points it is giving up and what freedom it might lose in the future.
A very simple test is to ask what would happen tomorrow if it were necessary to change model. Would it be enough to replace one component, or would it be necessary to renegotiate the rights over the use of information, repeat the post-training, rebuild parts of the workflow or give up accumulated knowledge? This is where the concept of the “engineered enterprise” takes on meaning.
The engineered enterprise is not designed around the AI ecosystem that exists today. It is designed so that it can adapt as external conditions change, keeping the complexity of the architecture invisible to the people and workflows that do not need to manage it directly.
Data as an asset
Finally, there is one last element. We are used to thinking of data as an asset with a relatively stable value. But the value of information also depends on the technological architecture able to use it. Current LLM-centered systems place great value on text, code, preference signals and corrections. Other technological trajectories could place greater value on state, actions, trajectories, physical interaction and simulation. An information stock can therefore remain accurate, proprietary and legally usable, but become more or less valuable if the architecture able to exploit it changes.
The same applies to the knowledge a company accumulates about its vendors: benchmarks, evaluations, routing rules and error profiles can quickly become obsolete when models and interfaces change. We are still at an early stage of the transformation. This is precisely why AI strategy cannot consist of taking a snapshot of today’s ecosystem and building the company around that snapshot.
Because in the age of artificial intelligence, competitive advantage will also depend on the company’s ability to understand who can use the information generated while these systems work with it, what knowledge remains under its control and what freedoms are preserved for the future.
