Abstract

National Statistical Offices (NSOs) face mounting pressure to modernize data production and dissemination amid severe human, technical, and financial constraints. Rapid advances in artificial intelligence (AI) and large language models (LLMs) present transformative opportunities to automate complex statistical tasks—provided governance, transparency, standards compliance, and intelligent monitoring are upheld. This paper proposes a governance-aware multi-agent AI architecture that automates SDMX compliance while incorporating adaptive intelligence for resource-constrained environments.

The proposed architecture leverages specialized LLM agents to emulate statistical roles: analyzing semi-structured Excel sources, designing SDMX Data Structure Definitions (DSDs), generating artefacts, and performing formal validation. Retrieval-augmented generation (RAG) integrates authoritative SDMX documentation, best practices, and institutional conventions to ensure conservative, standards-compliant outputs compatible with ODP 2.0 platforms. A dedicated SDMX Validation Agent enforces structural and semantic rules, delivering machine-assisted quality assurance aligned with official statistics governance.

Beyond automation and governance awareness, the architecture embeds intelligent capabilities through monitoring that tracks key metrics and powers self-optimization loops. Agents dynamically adapt behaviors, detect anomalies, and generate actionable insights for human oversight—enhancing reliability, scalability, and maintainability in low-resource NSO settings.

This approach directly addresses SDMX expertise gaps, fragmented institutional knowledge, and limited IT infrastructure by deploying lightweight, open-source LLMs on modest hardware. In NSO contexts with constrained resources, it augments human statisticians by reducing technical burdens, ensuring consistency, strengthening data domain interoperability, and enabling proactive system intelligence.

Aligning with IAOS 2026 themes of responsible AI use, capacity building, and statistical modernization, this work demonstrates how well-governed, intelligent multi-agent systems can drive sustainable transformation when embedded within international standards and institutional oversight.