Aquí tienes una traducción profesional, fluida y adaptada al lenguaje corporativo e institucional en inglés. Se han mantenido los nombres propios como AI Act (la Ley de IA de la UE) y se ha adaptado la terminología técnica para que suene natural en el ecosistema tecnológico actual de 2026.

Both positions evade the question that truly matters: what kind of AI can a Spanish public administration responsibly deploy in 2026?

The problem is not technological. It is a matter of fit.

Large language models exist, they work, and, in many contexts, they offer genuine value. The problem is not their technical capability. The problem is that they are designed for a business model that is structurally incompatible with the conditions of public administration.

Pay-per-use billing, provider-hosted cloud infrastructure, opaque models without audited traceability, fine-tuning starting at several million euros, and dependence on corporate decisions entirely detached from any public governance framework. None of these elements is a minor detail. They are design features of a product built for a completely different type of client.

After more than two decades accompanying the technological transformation of Spanish public bodies, we at AXPE have observed a consistent pattern: initiatives that fail do not do so due to a lack of technological ambition, but due to a lack of fit between the proposed solution and the actual conditions of the agency that must operate it. The exact same thing is happening with AI, only under greater media pressure and with less margin for error.

A public administration is neither a banking client nor a tech company. It manages data that does not belong to it—it belongs to the citizens—, operates under regulatory frameworks that are not optional, and answers to oversight mechanisms that demand an explanation for every relevant decision. Using generic, commercial AI in this context is not just a technical risk. It is a problem of institutional legitimacy.

European regulation is not an obstacle. It is an opportunity.

Read from the perspective of public administration, the AI Act is not a set of restrictions hindering AI adoption. It is, in reality, a rather precise description of what any public body should demand from any artificial intelligence system it deploys: traceability, human oversight, explainability, bias management, and continuous auditing.

What is striking is that many agencies are waiting for regulation to “clear up” before taking action, when the regulation is precisely what provides them with the strongest argument to reject models that fail to meet these requirements and to demand alternatives that do.

The regulatory framework does not block AI in the public sector. It blocks a specific way of doing AI—the kind that relies on hyperscalers, lacks explainability, and outsources data governance.

Small models, real value

One of the most widespread misunderstandings in the sector is equating AI capability with model size. The race among major labs to increase parameters has created a distorted perception: bigger equals better, and any alternative is necessarily inferior.

Empirical evidence points in a different direction. Specialized models of between 4 and 7 billion parameters, trained on domain-specific data, match and often exceed the performance of generalist models a hundred times larger in the specific tasks they were designed for. This comes with a substantial difference in inference costs, computational footprint, and the ability to deploy on existing infrastructure without requiring GPU investments that no public body could justify.

For public administration, this distinction is not a technical nuance. It is the difference between a financially viable model and one that is not.

What is already working

We are not talking about future forecasting. In projects where we have directly participated alongside Quant AI Lab, specialized models with sovereign architecture have been operating for some time in sectors with regulatory demands and data sensitivity comparable to public administration: real-time payment system fraud detection, a drastic reduction in false positives in anti-money laundering (AML) compliance—from 96.7% to 3.7% in an IBEX 35 company—, energy prediction in urban transport infrastructure, and document analysis with source traceability in regulated financial institutions.

Translating these capabilities to the public sphere does not require reinventing the wheel. It requires adapting what already exists to the procurement, technological integration, and governance frameworks unique to each agency. This is precisely where accumulated experience in public administration makes the difference between a pilot that fails to scale and an implementation that generates sustained value.

The true cost of not deciding

There is a third position, less visible than the two mentioned at the beginning, but probably the most widespread: indefinite waiting. These are agencies that recognize the potential of AI, identify specific use cases, and have read the AI Act, yet do not move forward because the vendor ecosystem fails to offer a solution that fits their actual conditions.

This waiting carries a cost. Not only in terms of deferred operational efficiency, but also in institutional capacity: every year spent without developing internal AI skills is another year of future dependency on external vendors, likely under less favorable conditions.

The relevant question is not whether AI will reach Spanish public administration. It is whether it will arrive under conditions that reinforce institutional autonomy or erode it.