Agentic artificial intelligence has moved beyond the laboratory as a promising concept and is now an operational reality in businesses and public administrations across Europe. Autonomous agents, capable of planning, executing, and making decisions without constant supervision , are transforming processes ranging from customer service to the management of critical infrastructure. However, this new capability comes with a host of legal, ethical, and security challenges that organizations must address with a clear strategy.
In Spain, the business ecosystem is already starting to stir. Microsoft has presented its new roadmap for the partner channel, aiming to boost the deployment of intelligent agents in more than 12.000 companies and public organizations. According to IDC data, by 2029 there will be more than 1.000 billion AI agents deployed worldwide, a growth that will generate increasing demand for design, implementation, and security services . This movement is complemented by the vision of consulting firms like Gartner, which identifies agentic AI as the most influential trend in the modernization of the public sector.
From pilot to deployment: why companies launch
The difference between reactive and agentic artificial intelligence is crucial. While traditional assistants answer questions or generate content, agentic AI works with a defined objective, uses external tools, and delegates tasks to sub-agents to complete entire processes. The leap from experimentation to production is the major challenge today , as Iván Martínez, Microsoft Spain's partner director, acknowledged during a meeting with integrators in Madrid.
An example of this progress is SAGE's initiative, which already offers a platform for its clients to design agents that automate back-office tasks. Similarly, the Valencian startup Maisa is implementing agents in financial and insurance environments, but with a key difference: they require the agents to work using deterministic code to guarantee traceability and prevent glitches. This approach contrasts with that of other companies that rely on large-scale language models for more flexible tasks.
The public sector is also getting on board. Gartner warns that government CIOs must modernize their operations and strengthen cybersecurity to govern AI at scale. Many governments are already working on governance frameworks and digital identities that will allow for the secure deployment of AI agents, recognizing that the transformation is not only technological but also organizational.
Risks and regulation: the fine print of autonomy
The new European Artificial Intelligence Regulation and the directive on liability for defective products are the cornerstones of a regulatory framework that is sure to have repercussions. The final phase of the EU General Safety Regulation, applicable from July 2026 , completes a regulatory framework that obliges manufacturers to assume responsibility for the behavior of their systems.
The courts have already issued a warning with real-world cases. In 2024, Air Canada attempted to distance itself from the errors of its chatbot, but the court ruled that the company is responsible for the information it provides on its website, regardless of its source. The agents that are emerging now are far more autonomous, and the law is preparing to define what constitutes a “defect” in a system that learns on its own. The Product Liability Directive includes software and AI , and stipulates that continuous learning and cybersecurity factors are relevant when assessing a defect.
On the competition front, the European Commission has imposed precautionary measures on Meta to restore access to WhatsApp Business for rival assistants. The battle for control of the channels through which agents enter the platform is another dimension of the same problem: delegating pricing or access to a single provider can lead to anti-competitive practices , as the US Department of Justice has already observed in the case of rental pricing software.
Cybersecurity: When agents become weapons
Cybercriminals have also seen an opportunity in agentic AI. The Google Threat Intelligence Group's AI Threat Tracker report reveals that nation-state actors are using language models and multi-agent architectures to automate entire phases of their attacks. One of the most worrying findings is a campaign that compromised thousands of credentials in six hours using hijacked infrastructure.
Model poisoning has become a common tactic: the TeamPCP group deploys malware that infects code assistant tools, causing them to recommend malicious dependencies. Furthermore, attackers steal AI-related intellectual property and abuse compromised cloud accounts to host local models and evade monitoring. The Spanish Data Protection Agency (AEPD) proposes a simple rule: an agent should not have simultaneous access to sensitive data, untrusted content, and the ability to act without authorization.
The local PC: the alternative for sensitive data
In this context, privacy-focused solutions are gaining traction. Perplexity and NVIDIA have expanded their Portable Computer platform to PCs with GeForce RTX GPUs, enabling agents to process sensitive documents without sending information to the cloud. Local processing not only protects data but also reduces remote computing costs . However, the agent can request permission to use external resources when the task requires it.
This commitment to data sovereignty aligns with the European trend of regulating the processing of personal information. The combination of local agents and cloud services is emerging as the most balanced model for companies handling critical data, although the hardware requirements are high: a GPU with at least 24 GB of VRAM is needed.
The impact of agentic AI isn't limited to large corporations. In Spain, startups like Supersonik are applying agents to optimize sales processes, while Orbio is automating the employee lifecycle in HR. The key lies in finding the balance between autonomy and human control , as experts pointed out at the Revista Emprendedores event. It's not about replacing workers, but about freeing them from repetitive tasks so they can focus on strategic decisions.
Ultimately, the question isn't whether agents will arrive, but rather how much operational capacity we are willing to delegate and with what safeguards. The European legal framework is moving towards holding those who deploy AI accountable , but organizations must recognize that the governance of these systems is just as important as their technical implementation. Agent intelligence offers enormous potential, but it will only materialize if accompanied by a vision that integrates innovation, ethics, and regulatory compliance from day one.