California Built Its Own AI Assistant Just for Government Workers

California has created an AI assistant called Poppy that only works for state government employees. The system runs on the state's own secure computer networks, not on the public internet. The California Department of Technology built it to help government workers do their jobs better while keeping their information safe and private.

Think of it this way: most AI assistants like ChatGPT can search the entire internet to find answers. Poppy works differently. It only looks at information from official California government websites. This design choice solves a major problem that government agencies face when trying to use AI: how to get help from this new technology without accidentally exposing sensitive internal documents or letting information leak out.

The system is already being used. More than 2,600 people across 66 different state departments are using Poppy right now. The name comes from California's official state flower, following a tradition of the state naming its technology projects after local symbols.

How Poppy Stays Secure

By only pulling information from California government websites, Poppy avoids the biggest risks that come with general AI systems. It cannot accidentally share internal documents. It will not give answers based on unreliable information from random websites. It also cannot be tricked by someone trying to inject harmful prompts through third-party content.

There is a trade-off. Poppy cannot access academic papers, news articles, or research from the broader internet, so it would not be useful for someone trying to do research on cutting-edge topics. But for what government workers actually need—figuring out internal procedures, understanding regulations, and finding official policy documents—the California government website content covers what employees need to know.

Another benefit: because all the conversations stay on California's own servers, the state meets strict security and privacy rules that government agencies have to follow. The information never gets sent to an outside company's servers, which means the state keeps full control over where its data goes.

Who Is Using It and How Fast It Is Spreading

About 2,600 users across 66 departments tells us something. This is not an experiment anymore. Government technology projects often get stuck in the testing phase because of security reviews and the slow pace of change in large organizations, but Poppy has moved past that stage into actual daily use.

The fact that departments average about 39 users each suggests this is not something forced on people from above. Instead, word is spreading between departments as people see it actually helps. This organic spread is how government technology usually grows when it works.

Why California Built It Rather Than Bought It

California decided to build Poppy in-house rather than pay a vendor for an off-the-shelf government AI product. This choice tells us something about how the state thinks about these new tools. Building it themselves gives California more control over how the system works, the ability to customize it for their specific needs, and better control over long-term costs. These are all major concerns for government agencies spending taxpayer money.

The broader context here is that California's approach mirrors something we saw happen before with cloud computing. When companies and government agencies first started moving their data to the cloud, they began with low-risk projects first—not their most important systems. Once they felt confident the cloud was secure and reliable, they gradually moved more work over. Poppy works the same way: California is starting with a limited, contained version to build up experience and confidence before possibly expanding it.

What This Means for Other Governments

California's Poppy offers a practical middle ground for government agencies nervous about AI. It provides AI's benefits while keeping the tight security and data control that government agencies must maintain. Other government agencies watching this might see a template they can follow.

There is also a broader lesson worth considering. Many people assume AI systems need access to the whole internet to be useful. Poppy shows that is not true. An AI system can do meaningful work within strict boundaries. This also sidesteps some thorny problems that affect general-purpose AI systems—unreliable information from random websites, bias in training data, and questions about accuracy. When you only feed the AI official government documents, those problems shrink significantly.

California's timing matters too. As more attention focuses on how government should use AI—with questions about whether these systems are fair, transparent, and accountable—this deployment shows one way to build AI systems that address those concerns through the actual technical design rather than just policies on paper.