This article has been originally published on Wired Italia on April 7, 2026 (Italian only).
The challenge companies face with AI is not, ultimately, a procurement problem. It is a design problem.
Every few weeks, the artificial intelligence industry generates a new peak of collective attention: a model release, a leap in benchmarks, another wave of speculation over who is leading and who is losing ground. Claude is now part of this cycle, as were GPT, Gemini, Llama, DeepSeek, Qwen, Kimi, and others before it. The pattern is always the same. Media debate narrows around frontier performance, as if the future of AI were to be decided primarily by which company builds the most impressive model first. For the businesses actually trying to use these systems, however, that is merely the most visible layer of the story.
Beneath the spectacle of the model race lies a quieter question, and for the corporate world, it is the one that matters most. Companies do not purchase AI as an abstract demonstration of power. They seek to embed it into sales, marketing, forecasting, customer service, compliance, and decision-making—all areas where performance depends less on raw power than on a system’s ability to work with the data, constraints, and internal routines of the company itself. This is where the terms of competition begin to shift.
Once a system becomes part of a company’s operational fabric, the market structure of its provider can no longer be treated as a secondary concern. The issue is not simply that a single company might become too dependent on a single vendor (which is certainly undesirable). The problem is that the current race for frontier models—fueled by extraordinary capital on the order of hundreds of billions of dollars for OpenAI and Anthropic—risks narrowing the field so drastically that dependence becomes inevitable for everyone. A market with only one dominant model, or even two, would not just reward innovation: it would begin to stifle it, forcing businesses to build their future on a layer of intelligence controlled elsewhere, with little real choice and even less bargaining power. We are still in the early stages, and technologies other than pre-trained transformers could shift the frontier again in the coming years. But the risk of concentration today is anything but theoretical.
Regardless, even when access to technology is available, the competitive question remains unresolved. Industrial history is full of companies that recognized a major breakthrough early on but failed to turn that initial insight into a lasting advantage. Just over 30 years ago, Mercedes prototypes developed under the European PROMETHEUS program traveled at high speeds in real traffic, performed lane changes, and completed long autonomous demonstration drives from Munich to Copenhagen—an achievement that still feels futuristic today. GPS had just hit the market; there was no mobile internet, no hyperscalers like Google, Amazon, Meta, or Tesla; CPUs were limited, GPUs non-existent, and cloud computing had not been invented. Compared to today, it was technological austerity leveraged to the extreme by the superiority of German engineering. Yet, the project was dismantled shortly thereafter. The organization was not ready to redesign itself around that capability or to transform a technical edge into a new business. The failure was structural, not technical: the German automotive industry surrendered a twenty-year lead in autonomous driving to a new generation of Silicon Valley companies. The point is not that the company failed to see the technology. It is that recognizing a new capability is one thing; reorganizing the enterprise around its implications is another. This distinction is decisive again today, as many businesses continue to treat AI as something to be acquired, rather than a force that may require a redesign of how they operate.
This is where the true challenge emerges. What companies face with AI is not, fundamentally, a procurement problem. It is a design problem. The question is not which model to adopt, nor which vendor to place at the center of the stack, but whether the organization is capable of building a system—its own system—that connects models, internal data, operational routines, human judgment, and operational controls into something it can effectively govern, adapt, and evolve over time. AI enters the enterprise as a tool, yet its consequences are potentially far more transformative.
This is also why the conversation about which model is “best” tends to mislead when it migrates from the frontier lab to the enterprise. A model can lead every public benchmark and still be strategically fragile within a company if it cannot be integrated into existing workflows, audited against internal standards, or replaced when conditions change. In most enterprises, the opportunity is more nuanced than what a Large Language Model (LLM) alone can offer. Valuable corporate data is often statistical, visual, operational, transactional, or inconsistently structured; many useful AI applications are not generative at all, even when they draw on the same corporate knowledge base. What matters at the enterprise level is not the isolated performance of a single model, but the quality of the architecture surrounding it—the retrieval layer, data governance, security, updating mechanisms, evaluation routines, and the interfaces that allow human judgment to intervene where necessary. Without that architecture, even an excellent model becomes a costly dependency.
When that surrounding architecture becomes decisive, modularity ceases to be a technical preference and becomes a matter of corporate sovereignty. A company that cannot switch models, reconfigure components, or prevent critical functions from collapsing into a single external dependency is ceding its room for maneuver. The design response is to build with abstraction layers, interoperable components, and the ability for agents and systems to operate across shared interfaces without dismantling or redesigning significant parts of the whole. In this way, when models change, when new capabilities must be incorporated, when costs shift, regulation tightens, or a dominant vendor alters its terms, the enterprise retains its margin of action. The point of modularity is not elegance for its own sake. It is to prevent the business’s intelligence layer from solidifying into something the company depends on but no longer controls. At Datrix, this logic is already visible in work with large financial institutions and highly regulated sectors, where governing hundreds of internal AI applications can require coordinating dozens of distinct LLM models and vendors, used interchangeably based on cost, privacy, speed, memory usage, and accuracy. In that context, data quality and AI governance stop acting as support functions. They begin to determine how the organization is structured and how it operates.
None of this works, however, if a company’s knowledge remains locked in silos, buried in legacy systems, scattered in incompatible formats, or simply held by individuals rather than recorded in databases. Proprietary data is often treated as a strategic asset in the abstract, but it only becomes one in practice when it can be organized, retrieved, governed, and brought to the point where a decision is actually made. This includes not only documents and structured records, but also operational traces, images, signals, spreadsheets, tacit routines, and forms of knowledge that have not yet been digitized. The competitive resource is not stored information. It is usable organizational memory—the ability to surface the right knowledge, in the right structure, at the moment a workflow requires it.
AI does not enter the enterprise as a neutral layer gently laid over existing routines. It forces choices about where tasks begin and end, which decisions can be delegated, where human review remains indispensable, how exceptions are handled, and what standards of reliability are acceptable before a system is considered production-ready. This is why “adoption” is the wrong word for what many companies are facing. The question is not whether AI can be inserted into the workflow, but how the workflow itself must be reorganized around new forms of assistance, prediction, and control.
Once work has been reorganized in this way, governance can no longer sit alongside the system in a policy document. It must be built into the machine itself: in permissions, escalation rules, retrieval boundaries, evaluation routines, audit trails, and the moments where human judgment must remain present. Governance becomes part of enterprise engineering, not a commentary on it. Companies that treat it as a compliance layer applied after deployment will find that the gap between what their systems do and what they can account for widens faster than any review committee can bridge.
There is, however, a risk deeper than any failure of architecture or governance. When a company relies heavily on AI-generated outputs—summaries, recommendations, forecasts, drafts—the people within it may gradually stop performing the cognitive work the system seems to handle for them. Errors matter, but the more significant change is quieter: fluent and plausible output begins to substitute for the organization’s own understanding. There is a 2026 neologism that Treccani introduced to describe this phenomenon at the individual level: Epistemia, defined as “the comfortable illusion of knowledge produced by interaction with the generative AI of large language models (LLMs), where the simulative plausibility of fluent discourse and narrative coherence replace cognitive efficiency and data reliability.”
Mutatis-mutandis, moving from the individual to the socio-organizational level, a business may begin to mistake generated output for authentic understanding. A model may appear informed while remaining disconnected from the company’s evolving experience, its institutional memory, and its judgment. Over time, the danger is that judgment itself begins to atrophy—not because people disappear from the process, but because their role is progressively reduced to accepting, forwarding, or lightly tweaking conclusions they no longer interrogate deeply.
This is not an abstract hypothesis. In a qualitative study published in March 2026 by Anthropic, a lawyer confessed: “I use AI to review contracts, to save time… and at the same time, I wonder: am I losing the ability to read for myself? Thinking was the last frontier.” The fact that he said this to a chatbot is itself part of the diagnosis.
In Europe in particular—where the AI Act and the Digital Operational Resilience Act (DORA) are already imposing requirements on transparency, auditability, and third-party dependence, and where the dominant AI infrastructure is largely provided from outside the continent—AI cannot be treated as a procurement choice delegated to vendors or technical teams. The issue goes further: it touches on sovereignty, accountability, and the enterprise’s ability to maintain control over how judgment is formed and exercised. This leaves managers facing a task many were never trained for. Their responsibility is no longer limited to approving the adoption of a technology; it extends to shaping the conditions under which that technology enters the organization without hollowing out the organization’s own capacity to think, decide, and act.
What emerges is not a set of tools to be installed, but a different type of company. Let’s call it the Engineered Enterprise: an organization that does not merely consume AI like a functional illiterate, but deliberately designs the system through which models, data, workflows, controls, and human judgment interact. The companies that navigate this transition most successfully will not necessarily be those with access to the single most powerful model. They will be those that learn to build around intelligence as an internal capability—without surrendering the autonomy to change, adapt, and govern what they use, while protecting the only asset no external system can replace: the organization’s capacity to understand what it is doing and to imagine what does not yet exist.
It is interesting, then, to shift perspective. In a recent qualitative study conducted by Anthropic of over 80,000 Claude users across 159 countries—a study to be taken with significant methodological caution, given these are self-selected users interviewed by the company’s own bot—a geographic divide clearly emerges. While concerns over governance, loss of autonomy, and cognitive risk dominate in Europe and North America, in sub-Saharan Africa, South Asia, and Latin America, AI is described primarily as a lever for learning, entrepreneurship, and overcoming structural disadvantages. An entrepreneur in Uganda uses the same AI model to build a business from scratch that no investor would have funded. In Cameroon, someone reaches a professional level in cybersecurity, marketing, and design in just a few months—skills that otherwise would have required years and inaccessible resources. For them, AI is not an architectural problem. It is a lever for emancipation. The Engineered Enterprise is the answer for those who already have a solid present: they possess structure, capital, and complexity to govern. But in other parts of the world, something different is being born: the augmented entrepreneurship of the individual, powered by a model trained with far more knowledge than is locally available and a single connection. The future of AI, like beauty, is in the eye of the beholder.
This capability also depends on the competent and inquiring curiosity of its people—the disciplined habit of asking better questions instead of accepting easy answers. Ultimately, the companies that will matter—as was true yesterday and remains true today—will not be those that simply face the future as it arrives, but those still capable of inventing and taking ownership of a future that no AI system trained on the past could have ever parrotted back to them.
