As of July 1, 2026, California is rolling out across its entire administration a generative artificial intelligence platform called Poppy. The transition, described by the California Department of Technology (CDT) as taking effect on that date, closes a pilot phase launched on September 29, 2025. More importantly, it offers a model for deploying generative AI in the public sector, where data sovereignty, vendor dependence, and the protection of personal information were treated as non-negotiable requirements from the outset. For a French or European administration looking to equip its staff without exposing its data, the experience is rich in lessons.
Poppy, a single interface for several models
Poppy is presented by the CDT as a generative AI platform vendor-agnostic, meaning it is not tied to a single supplier, and "built by California state employees, for California state employees". In practical terms, a single interface provides access to several language models that agencies can use interchangeably — Claude (Anthropic), Gemini (Google), GPT (OpenAI), or Nova (Amazon) — "without vendor lock-in or contract renegotiation," according to the administration.
The intended uses remain deliberately office-oriented: drafting documents, reports, and clear communications; summarizing and analyzing large volumes of data or complex case files; and searching for rules and public policy guidance using trusted public sources. Poppy is limited to the role of a productivity assistant; automating administrative decision-making is explicitly out of scope. Governor Gavin Newsom made that clear in no uncertain terms in a June 29, 2026 press release: "AI should not replace the human work of government; it should help our staff move faster, solve problems more effectively, and deliver better results for Californians."
Three safeguards designed in from the start
The value of the system lies less in the tool itself than in its trust architecture. The CDT highlights three guarantees that any administration should require before opening such a service to its staff.
Data never leaves the State environment
"Information shared with Poppy never leaves California's trusted environment," says the CDT. The platform relies on State-managed infrastructure rather than a consumer-facing service. That is a direct answer to the main obstacle facing generative AI in the public sector: the risk that sensitive data entered into a prompt could end up feeding a third-party service beyond any real control.
Built-in detection of personal data
Poppy includes "guardrails to detect personally identifiable information (PII) and other sensitive information." The mechanism is designed to identify, at the time of entry, an ID number or protected data before it is sent to the model — a particularly sensitive issue in an administration that handles citizen records on a daily basis.
No training on public data
The CDT sums up a decisive guarantee in a terse phrase: "no model training." Data entered into Poppy is not used to train the underlying models. This clause, often missing from consumer offerings, cuts off the main pathway through which administrative information could indirectly end up in a foundation model's parameters.
A large-scale pilot before general rollout
The rollout did not come out of nowhere. Since September 29, 2025, a pilot phase has involved, according to the CDT, more than 2,800 public employees across 67 State departments, tasked with testing the tool and providing feedback. This approach — broad testing in real-world conditions before deployment — stands in contrast to announcements of mass adoption without prior validation. It also explains the tool's office-oriented positioning: the scope was shaped by the concrete needs of employees rather than by a top-down technological promise.
Do not confuse Poppy with the Anthropic contract
California's recent news has nonetheless brought together two announcements that must be kept distinct. On June 29, 2026, Governor Newsom announced a "first-of-its-kind" partnership with Anthropic, giving State agencies and local governments access to the Claude assistant at a 50% discounted rate, along with training and technical support. That partnership highlights concrete deployments: the DMV for customer service, health services (Medicaid) for internal workflows, and the CDT and CalOES for cyber defense.
Poppy, by contrast, is distinct from the "Anthropic assistant": it is a multi-vendor State platform, with Claude among the available models (even though it was used, among others, in its development). The confusion is easy, but the implications are significant: it is precisely Poppy's agnostic nature that protects California from dependence on a single publisher. An advantageous vendor contract on one side, a neutral and sovereign platform on the other: the two approaches coexist without being conflated.
What French and European administrations can learn
The California case cannot be directly transplanted: the GDPR and the AI Act impose their own constraints. It nevertheless provides a useful lens at a time when French and European administrations are considering how to equip their employees.
Vendor neutrality, first of all: by building a common access layer to several models, California avoids proprietary lock-in and keeps the ability to choose between providers over time. Data control, too: hosting the service in a controlled environment, filtering personal data, and contractually excluding any retraining all answer point by point to the requirements a European data controller should impose. Method, finally: a broad, documented pilot before deployment, and a scope of use clearly framed as an assistant, not a decision-making machine.
None of these principles is theoretically new. California shows that they can be brought together in a service made available to tens of thousands of employees, at the scale of a State with nearly 40 million residents. European administrations that will equip their staff in the coming months now have a documented precedent, backed by nine months of pilot testing, 2,800 testers, and written contractual safeguards.