The OECD.AI Policy Navigator

Our policy navigator is a living repository from more than 80 jurisdictions and organisations. Use the filters to browse initiatives and find what you are looking for.

AI for resilient digital government services (Subnational: Province of Cordoba)


Added by:   OECD analyst
Added on:   21 Sep 2026
Updated by:   OECD analyst
Updated on:   21 Sep 2026

The Province of Córdoba (Argentina) developed an autonomous AI agent system that monitors, diagnoses, and resolves incidents in the digital infrastructure supporting public services. Operating across government cloud platforms serving 3.8 million citizens, it reduces service disruptions from hours to minutes while creating full audit trails for every automated action. https://infografia-aws-gob-cba.vercel.app/

Name in original language

IA para servicios gubernamentales digitales resilientes

Initiative overview

The Province of Córdoba’s citizen‑facing digital infrastructure supports services such as digital identity (Ciudadano Digital – CiDi), online procedures, payments and public information portals for more than 3.8 million residents. Ensuring their continuous availability has become a major governance challenge. To strengthen resilience, the Sub‑secretariat of Technological Infrastructure developed an autonomous AI‑agent ecosystem that monitors, diagnoses and supports incident resolution across the province’s cloud environment. Operating across 66 AWS accounts, the system provides continuous monitoring, faster response times and full auditability of every automated action, directly improving the reliability of essential public services.

The solution is built as an ecosystem of AI agents running on AWS, using Anthropic’s Claude models through Amazon Bedrock so that all inference remains within the province’s cloud environment. Claude Sonnet performs minute‑by‑minute monitoring and first‑line analysis, while Claude Opus conducts deeper diagnostics and produces technical reports. Applications are evaluated across five dimensions—connectivity, HTML structure, content integrity, error detection and external resources—to generate a configurable health score. When the score drops below the alert threshold, a specialized agent initiates a structured seven‑stage workflow covering detection, diagnosis, planning, execution, validation, resolution and closure. All steps are documented in real time, and automated summaries are shared with technical and executive stakeholders.

The architecture was designed with transparency and accountability at its core. All inference takes place within Córdoba’s AWS boundary, with CloudTrail logging and IAM‑based access controls aligning the AI layer with existing infrastructure governance. This ensures that every AI‑driven action is auditable to the same standard as any other infrastructure change, a requirement that shaped the model and platform selection. The broader initiative forms part of Córdoba’s strategy to reinforce resilient digital‑government infrastructure in an increasingly digital society.

The project’s development independently arrived at a “read‑only first” deployment model, prioritizing diagnostic capability before allowing any automated changes. This approach was later validated externally when AWS announced its DevOps Agent in December 2025 with an identical philosophy—confirming Córdoba’s alignment with global best practice in responsible agentic AI. The province is committed to regional knowledge sharing. An interactive infographic documents the workflow, and the team is prepared to share its risk matrix, governance framework and 23‑item documentation standard with peer governments. 

Other relevant details

The system’s deployment follows a phased governance model based on a five‑level risk matrix. The current phase operates strictly in read‑only mode, with the agent functioning as a diagnostic co‑pilot under restricted access conditions enforced through Identity Center and SSM Session Manager. Later phases envision controlled execution of low‑risk actions with explicit human approval, expanding gradually to more complex operations with multi‑operator validation, while destructive actions remain permanently prohibited. The architecture is designed for scalable replication. Adding a new AWS account requires only configuring a read‑only IAM role and loading the relevant workload knowledge base. A standardized documentation framework—covering runbooks, disaster‑recovery plans, Terraform files and other operational artifacts—supports expansion across 26 production accounts. Governance is managed through the organization’s AWS Identity Center, enabling controlled growth while minimizing vendor lock‑in. During MVP trials, the system autonomously handled more than 25 incidents with over 90% success. Detection time dropped from roughly 45 minutes to under two; diagnosis time fell from about an hour to under three minutes; and resolution times decreased from several hours to under twenty minutes. These improvements significantly enhanced the reliability of critical services such as digital identity and online procedures. Following MVP validation, the system entered Phase 1 of production, and performance metrics will continue to be gathered as capabilities expand. Projected impacts include sub‑20‑minute recovery times, 24/7 monitoring, sub‑15‑minute SLAs and a significant reduction in operational burden on IT teams.

About the policy initiative


Category:

  • AI policy initiatives, programmes and projects

Initiative type:

  • AI use cases/projects in the public sector

Status:

  • Active

Start Year:

  • 2024

Target Sectors:


Other relevant urls: