Highlights
- In the agentic era, governments could delegate processes to AI agents, focusing on oversight and decision-making.
- Beyond the technology, preparing for agentic AI requires governance, process redesign and leadership.
- Latin America and Caribbean countries must engage now in shaping the standards for agentic AI in government.
By Lenin Fabricio Rodriguez Yanez
Artificial intelligence is already changing how governments work. So far, the most visible uses have been at the interface layer: chatbots that answer questions or tools that help public officials draft documents.
These advances are still in the early stages of implementation, and they share a common premise: a human being remains the user of these tools, making decisions, initiating actions, and carrying responsibility for outcomes. But that operating model may not last for long.
The transformation that was already underway
A 2024 report on digital transformation in Latin America and the Caribbean from the Inter-American Development Bank (IDB) highlighted both progress and persistent constraints, including fragmented legacy systems, weak interoperability, and the gap between digitisation and genuine transformation.
Since its publication, a new shift is now emerging on top of those challenges: the transition from AI as a tool that supports work to AI as an operational layer through which government functions may increasingly be executed.
From assistants to operators: The shift to agentic AI
This is the core distinction behind agentic AI. Traditional generative AI tools respond to prompts and support human users. AI agents, by contrast, can pursue a defined objective across several steps: gathering information, querying systems, and applying rules or context. The human role shifts from executing each step to designing the process, supervising outcomes, and intervening when exceptions arise.
Estonia’s Bürokratt initiative offers an early glimpse of this transition. Rather than sending citizens to navigate separate agencies, it connects interoperable AI assistants across more than 18 government organisations.
A citizen states a need; the system helps identify the relevant steps and services. While legal frameworks for fully autonomous public-sector agents are still evolving, the infrastructure is already being built, and governments beyond Estonia are beginning to follow suit.
As NVIDIA CEO Jensen Huang noted at NVIDIA GTC 2026, every IT and software-as-a-service (SaaS) company will eventually become an Agentic-as-a-Service (AaaS) company. A recent MindStudio article on AaaS captures the implications well: in the SaaS model, a human operates the tools; in the AaaS model, AI agents can do the work autonomously, continuously, and often simultaneously across multiple systems.
That distinction matters in public administration. An AI agent handling a business-license renewal, for example, could retrieve registry data, review inspection records, identify inconsistencies, and escalate only the cases that require human judgment. In that model, the question is no longer only how officials use AI, but how governments govern AI systems that participate in execution.
Interoperability reimagined
This shift also matters for interoperability. In many countries, citizens still act as messengers between public institutions, moving documents from one agency to another because systems do not communicate effectively. Agentic architectures could reduce some of that friction when they rely on shared standards that allow agents to access and coordinate across systems more consistently.
The emerging standard at the center of this shift is the Model Context Protocol (MCP), introduced in 2024 and now adopted by several major technology providers. Before MCP, connecting AI systems to external data sources often required custom integrations. MCP offers a shared interface for agents to interact with tools and data sources, which could reduce the cost and complexity of connecting systems.
For governments, relevance is practical. If agents can work through these protocols, they may be able to coordinate across public systems even where full structural integration has not yet been achieved.
For example, the US Government Publishing Office offered an early preview of this logic in January 2026, launching a public MCP server that allows AI agents to access the official federal information repository directly, describing it as a universal API for agents.
Agentic architectures complement existing interoperability efforts and don’t eliminate the need for data governance, integration strategies, or interoperability policy. They do, however, suggest a more flexible operating model for environments where full integration remains incomplete.
The cost of doing nothing
For Latin America and the Caribbean, this matters now. If governments do not begin building the technical and institutional conditions required for agentic operations, they will not simply miss an efficiency opportunity. They will continue to operate government-as-usual while the world around them is increasingly shaped by systems that can operate at a speed, scale, and level of coordination that human bureaucracies cannot match.
The standards and operating models that will shape agentic AI are defined in real time. If governments in the Global South do not engage early, those standards may evolve without reflecting their institutional realities, service-delivery constraints, or public-policy priorities.
Preparing for agentic AI, therefore, means more than experimenting with new tools. It means redesigning processes for agent readiness, investing in infrastructure, and building governance arrangements that can operate on a scale.
The governance gap is real
Still, the promise of agentic AI is only as strong as the governance built around it. Existing frameworks were not designed for systems that can retrieve data from multiple sources, apply probabilistic reasoning, and take actions within defined parameters. When a human official acts, accountability can usually be traced through established procedures. When an agent acts, responsibility, auditability, and escalation become more complex.
Recent research has started to address that challenge. One proposed model, Governance-as-a-Service, introduces an external enforcement layer that evaluates and constrains agent actions at runtime without changing the internal model logic. Because monitoring agent activity at scale becomes increasingly difficult for humans, the framework introduces automated mechanisms to evaluate, monitor, and escalate actions in real time. In experiments, GaaS successfully blocked high-risk actions while preserving operational results.
The broader point is important: in the agentic era, governance cannot remain only a policy statement. It must also become operational, technical, and testable in practice.
Public administrations also face real implementation risks. Early enterprise deployments have highlighted cybersecurity exposure, the risk of unauthorised actions, reasoning errors that can scale quickly, and vendor lock-in that may reduce long-term flexibility. Most governments do not yet have the institutional, legal, and technical capacity needed to manage these risks on scale.
Not a new chapter — a different book
Agentic AI is not a guaranteed outcome, but a transition that demands deliberate preparation. Generative AI opened a new chapter in how people interact with software and information; agentic AI may represent a different book altogether, one in which systems do not simply assist human decision-making but increasingly participate in the execution and coordination of government operations. The central question is no longer whether this transition deserves attention; it is whether governments will build the capacity to shape it before it shapes them.
The cost of delays will be a wider digital divide between countries already building the enabling conditions for this transition and those that have not yet begun.
Recognising the importance of this emerging agenda, the IDB has been leading regional discussions on the governance and operational implications of Agentic AI, including the 2025 Regional AI Policy Dialogue held in San José, Costa Rica. A new regional dialogue planned for 2026 will dive deeper into how governments can shape this emerging operational layer of the State in ways that strengthen public institutions and deliver greater value in the region.