WP3 meeting - 19.02.2026

[AmilinaK]

Meeting Overview
The session focused on the coordination and technical planning of Work Package 3 (WP3), aimed at developing technical guidance and illustrative examples for AI-ready dissemination of official statistics. The group is composed of experts in software engineering, metadata, dissemination, and AI, collaborating internationally to align on objectives, scope, milestones, risks, dependencies, and next steps.

Objectives of WP3

• Provide practical technical guidance and illustrative examples on disseminating official statistical data in AI-ready, machine-readable formats.
• Focus on existing standards (e.g., SDMX, DDI, MCP) to support reliable discovery and consumption by AI systems.
• Work primarily on proof-of-concept (POC), non-production environments, not full production systems.
• Ensure interoperability and homogeneity with other work packages (WP1, WP2, WP4).

Scope and Focus

• AI-readiness involves enabling AI systems (such as ChatGPT, Google Bard, Gemini) to consume official statistics effectively.
• The project will not build production-ready dissemination systems but provide technical architectures and reusable guidance.
• Major scope includes defining formats, metadata standards, APIs, and technical patterns for AI consumption.
• There was emphasis on balancing scope breadth to avoid overly broad or impractical guidance.

Dependencies

• Strong alignment with WP2 (metrics definition) and WP1 (technical innovation).
• Use inputs and outputs from WP1 and WP4 to enhance WP3’s implementation quality.
• Coordination required across multiple organizations for contributions and sharing best practices.

Key Risks and Mitigation Strategies

• Scope creep and rapid technology changes: Mitigated by focusing on concrete, simple, and practical patterns and routine review meetings.
• Availability of personnel: Dividing work into milestones and lightweight POCs to manage time constraints.
• Technical capability gaps in some organizations: Use of existing open-source standards and systems rather than inventing new tools.

Important Discussion Points & Insights

Data consumption by AI systems:

• Current AI tools use a mix of training data and live web search to retrieve statistical information.
• Recent AI models (ChatGPT, Bing) prioritize live data search before relying on static training data.
• There is a “black hole” for detailed data available only via APIs or files not incorporated into AI training or web search.
• MCP (Multi-Dimensional Cube Protocol) and similar standards may help fill this gap by structuring detailed data for AI use.

Different categories of AI data consumption:

• Web page content (metadata and textual information) accessible via web crawling.
• Data retrieved from databases or APIs (structured data).
• Training data sets used by AI models.
• These categories require different technical approaches and should be recognized distinctly.

User personas and use cases:

• Different end-users (general public, journalists, researchers) engage with statistical data differently and may require diverse AI interaction methods.
• MCP-based data access is useful for advanced users (researchers) but may not serve casual users.
• Guidance should consider these personas and cater to multiple engagement levels.

Technology and standards mentioned:

• SDMX (Statistical Data and Metadata eXchange)
• DDI (Data Documentation Initiative)
• MCP (Multi-Dimensional Cube Protocol)
• PX format (open-source platform used in Nordic countries)
• ETL processes for integration of standards-based databases

Next Steps and Actions

• Collect baseline information via a survey form distributed to participants regarding existing data formats, metadata, and dissemination approaches.
• Organize bimonthly meetings (approx. 1-hour sessions) to share findings and coordinate ongoing research and development.
• Conduct desktop research on AI data consumption methods and existing literature, including sharing of relevant studies and reports.
• Engage with technical teams (e.g., dot stat SDMX developers) for detailed presentations and collaboration.
• Encourage participant contributions and volunteers for focused tasks within WP3.
• Plan a hybrid sprint event (tentatively in Belgrade, April 2026) for intensive collaboration.

Key Conclusions

• Work Package 3 aims to create practical, standards-based technical guidance to make official statistics AI-ready, focusing on formats and metadata that AI systems can consume reliably.
• The project timeline is structured with concrete milestones, balancing research, drafting, and demonstration phases.
• Close collaboration and data sharing across multiple statistical agencies and work packages are essential for success.
• Understanding AI data consumption mechanisms and user personas is critical to tailoring dissemination approaches.
• Existing open standards and tools (like MCP and SDMX) will form the backbone of the proposed solutions.
• The initiative prioritizes flexible, simple, and maintainable approaches over complex or broad-scope solutions to manage fast-moving AI technology developments effectively.