UN turns to Google to make its global data ready for AI agents

The United Nations announced a new data platform developed with Google to make the UN system’s statistics easier for AI tools to access and use. Called the UN System Data Commons, the service is built on Google’s open-source Data Commons framework and replaces the older UNData portal with a natural-language search interface and features that let AI agents query statistics directly.

A key technical addition is support for the Model Context Protocol (MCP), a standard that enables AI systems to connect to external data sources and retrieve figures and their provenance. The UN said the platform records the origin of each statistic so that results generated by connected AI systems can be traced back to the original UN source. At launch, the system includes data from nearly 20 UN entities, and 26 UN organizations have committed to the project; the UN aims to bring 80% of its statistical datasets into the Commons by 2027.

The initiative responds in part to evidence that current large language models often struggle to return reliable numeric answers for development indicators. UNICEF tested six models across more than 133,000 responses and reported an average accuracy of 21.2%; roughly 60% of responses lacked a usable number, and when models did provide numbers, identical answers repeated only about half the time when the same queries were re-run two days later. UNICEF described the work as a working paper being prepared for journal submission and said it will publish methodology, code, and data alongside the paper.

Google.org contributed $2 million in funding and technical support for the platform’s core infrastructure. Google and UN staff said the system is hosted on a UN-governed instance and is intended to be operated and scaled by the UN in time. Demonstrations showed that an AI agent using MCP to query the Data Commons can automatically combine multiple indicators to generate dashboards, charts, and written summaries, though Google staff emphasized that human review remains necessary because models can misinterpret nuance.