On July 1, California became one of the first governments in the world to roll out a generative AI assistant to every state agency. The platform is called Poppy. It runs 10 models. It pulls information only from official CA.gov sources. And state employees built it.
That last part is the detail most people will skip past. It is the only one that matters.
The Pilot That Actually Worked
Poppy started as a pilot in September 2025. The California Department of Technology put it in the hands of 2,800 employees across 67 state departments. That broad test put the same platform in front of people doing different kinds of government work, including DMV clerks, environmental regulators, and procurement officers.
The pilot ran for nine months. Employees gave feedback. The CDT iterated. And on July 1, 2026, it went statewide.
Compare that to the average enterprise AI pilot. Most companies pick one team, run a 90-day test, produce a report, and then spend the next year debating whether to expand. California tested across 67 departments simultaneously and shipped to production in under a year.
The platform itself is boring on purpose
Poppy consolidates 10 commercially available AI models into a single interface, configured for the specific needs of state government work.
Employees can draft documents, analyze data sets, and research policy. They can compare outputs from multiple models side by side in a single view. The interface includes pre-built query templates for common state business tasks so that someone who has never touched AI before can get useful results on day one.
Here is what Poppy does not do: it does not send data outside California’s trusted environment. It does not use employee inputs to train the underlying models. The CDT’s guarantee is four words: “no model training.” Every interaction stays inside the state network.
The design keeps familiar tasks and explicit privacy boundaries close at hand.
The Design Decision That Explains Everything
Adoption can stall even when a model can do the work. Employees may distrust it, struggle to use it, or leave after a bad first result. That makes usability an implementation concern alongside capability.
California designed around that exact failure mode. Pre-built templates mean nobody has to learn prompting. Trusted CA.gov sources mean employees do not have to evaluate whether the output is pulling from reliable data. The “no model training” guarantee means the privacy conversation is settled before it starts. And vendor-agnostic model access means no one is locked into a single provider’s interface or pricing.
Each technical decision also makes the tool simpler for employees to use.
So what would yours look like
Before choosing a model, business leaders should ask what their own version of Poppy would look like.
Start with the work your people actually do every day: document drafts, data pulls, policy lookups, and the repetitive analysis that eats hours every week. Evaluate vendor use cases against those tasks.
Then build the simplest possible interface that connects AI to those tasks. Pre-built templates. Trusted data sources. Clear privacy boundaries. Something a new hire can use in their first week without asking IT for help.
California took 10 existing models, wrapped them in an interface its employees could use, restricted the data environment to sources it already trusted, and shipped it. That gave the state a practical way to deploy the capability available at the time.
2,800 pilot users across 67 departments. Statewide rollout in under a year. Built by the people who will use it.
Meanwhile, most large companies are still circling. New vendor eval. New pilot cohort. New steering committee. A state government — the DMV people — shipped production AI to every agency while the Fortune 500 was still scheduling alignment meetings.
Making the tool simple enough for people to use was central to the rollout. California got that right. If your organization hasn’t, the uncomfortable truth is that a government bureaucracy just lapped you.