The UN Is Rebuilding Global Data for the AI Agent Era
AI can answer almost anything. Getting it to answer with the right number is still surprisingly difficult.
The United Nations is working with Google to rebuild how its enormous collection of global statistics is accessed, with a new system designed not only for people searching databases but also for AI agents.
The UN System Data Commons is built on Google's open-source Data Commons platform and supports the Model Context Protocol, or MCP, allowing AI systems to connect directly to authoritative UN datasets. The platform replaces the traditional UNData portal and launches with data from nearly 20 UN entities.
That sounds like a database upgrade.
It is actually a much bigger shift in how institutions may need to publish information in an AI-first internet.
The problem isn't that AI lacks information
Large language models contain enormous amounts of knowledge.
But factual retrieval remains messy.
UNICEF tested six major AI models across more than 133,000 responses involving global development indicators. The average accuracy score was just 21.2%, according to the agency's chief statistician.
Roughly three in five responses failed to produce a usable number at all. And when the same questions were asked again two days later, models that supplied numbers both times returned the same figure only about half the time. The working paper has not yet completed peer review.
That's an important distinction in the AI reliability debate.
The model may understand the question.
It may know which dataset probably contains the answer.
It still might not retrieve the correct statistic.
The UN is turning data into infrastructure for agents
Traditional websites were designed around humans.
A person opens a webpage.
Clicks through menus.
Downloads a spreadsheet.
Reads a table.
AI agents need something different.
They need structured data they can query programmatically, along with enough source information to understand where each figure came from.
That's where MCP becomes important.
The protocol lets an AI system connect to external data sources instead of relying entirely on information stored inside the model.
For the UN, that means an agent could retrieve an updated mortality rate, population figure or education statistic directly from the authoritative source.
The new platform also preserves provenance, allowing users to trace a statistic back to the UN agency that supplied it.
AI referral traffic is already changing information discovery
The UN isn't redesigning its data systems for a hypothetical future.
Users are already arriving through AI.
UNICEF says its data website receives more than 6 million visits per month, while referrals from ChatGPT increased 67% year over year between January 1 and September 14.
ChatGPT referrals represented 6.4% of sessions this year, and UNICEF estimates AI assistants overall now account for about one in 10 visits.
That creates a strategic problem for every organization that publishes important information.
If users increasingly ask an AI assistant instead of visiting a website directly, the information has to be structured so the AI can retrieve it correctly.
Search-engine optimization may gradually be joined by something else:
agent optimization.
Google gets another infrastructure position
Google.org provided $2 million in capacity-building funding and technical assistance for the platform.
The system runs on a UN-governed instance and is intended eventually to be operated independently by the organization. The UN aims to move 80% of its statistical datasets onto the platform by 2027.
For Google, this strengthens Data Commons as a layer between public information and AI systems.
That is strategically useful.
If AI agents increasingly rely on structured external information, the technology connecting models to trusted datasets becomes almost as important as the model itself.
Authoritative data doesn't guarantee authoritative analysis
There is still a limitation.
Connecting an AI to trustworthy numbers does not guarantee it will interpret those numbers correctly.
A model can pull the correct statistics and still misunderstand the context, confuse correlation with causation or generate an unsupported conclusion.
Google itself cautioned that humans should review model-generated analysis before publishing or citing it.
So the new architecture solves one AI problem.
It doesn't solve all of them.
What happens next?
The UN's move could become a template for governments, research institutions and large public databases.
Instead of expecting AI companies to scrape websites and reconstruct knowledge imperfectly, institutions can expose structured, traceable data directly to agents.
That could significantly improve factual AI systems.
But it also changes how information travels.
The future of the internet may involve fewer users manually searching databases and more AI agents retrieving facts on their behalf.
The next important question for publishers may therefore be less:
Can people find our information?
And more:
Can an AI agent understand where to find it — and prove where it came from?