Government AI Deployment in California: Contracts & Staffing

Government AI deployment in California runs through a framework the state built itself: Governor Newsom's 2023 GenAI executive order, a CDT risk-assessment gate every tool must pass before purchase, five scaled use cases already moving into production, and — since January 2026 — the AB 2013 training-data transparency law that constrains which vendors an agency can defensibly buy.

California is the most structured public-sector AI market in the country, and the most consequential. It was the first state to issue a generative-AI executive order, the first to publish binding GenAI procurement guidelines, the first to require training-data disclosure by statute, and the first to sign production agreements deploying GenAI into live state operations. For agencies and the system integrators serving them, the model is the easy part; the risk assessment, the solicitation rules, the transparency obligations, and the staffing are where programs stall. Gain America closes the last of those gaps — we staff and deploy the forward-deployed engineers, MLOps teams, and public-sector-ready delivery talent that get California AI systems into production without the frontier-lab salary load the state cannot carry.

California's GenAI executive order and the state AI assessment framework

Every California public-sector AI program traces back to one document. On September 6, 2023, Governor Gavin Newsom signed Executive Order N-12-23, making California the first state to formally direct how generative AI is studied and deployed in government. The order did three things that still govern the market: it required state agencies to produce risk-assessment reports on how GenAI could affect their operations, the state economy, and energy usage; it urged the legislature to build new AI policy; and it launched a deliberate, staged process for evaluating and deploying GenAI rather than letting agencies improvise.

That order produced a fall-2023 report cataloging both the risks — GenAI can generate convincing but inaccurate output and automate bias — and a set of promising government use cases. From there, California built the assessment framework that now sits in front of every deployment. The Government Operations Agency (GovOps), the California Department of Technology (CDT), the Department of General Services (DGS), the Office of Data and Innovation (ODI), and CalHR jointly authored the state's GenAI guidelines and toolkit. The centerpiece is the GenAI Risk Assessment, codified as SIMM 5305-F, which forces an agency to classify a proposed tool as Low, Moderate, or High risk — based on the model, the data underneath it, and the intended use — before it can move toward procurement.

This is the same pattern that separates programs that ship from programs that die in pilot. As we cover in why government AI projects fail, the states that treat governance as a design input rather than an afterthought are the ones that reach production. California wrote the governance down first.

In California, the binding constraints are published before you start: an executive order sets the posture, SIMM 5305-F sets the risk gate, and AB 2013 sets the vendor-transparency floor. Architect to all three from day one, or the pilot never clears the assessment.

The five scaled use cases: translation, call centers, traffic, road safety, health inspections

California did not stop at policy. It identified five GenAI use cases with statewide scaling potential and moved them from concept to live agreements — the clearest signal of where public-sector AI demand actually is:

• Language translation and access — Nearly 20% of Californians have limited English proficiency. The California Health and Human Services Agency is testing whether GenAI can produce better translations faster to improve access to benefits and services.
• Call-center productivity — The Department of Tax and Fee Administration (CDTFA) piloted a GenAI assistant to cut average call-handling time. In past crunches, CDTFA reassigned roughly 280 staff to back up its call center; during the trial it sustained service without that reassignment.
• Traffic and congestion analysis — Caltrans is evaluating GenAI to interpret complex transportation data, address bottlenecks, and improve traffic management and incident response.
• Vulnerable-roadway-user (VRU) safety — Caltrans is also testing whether GenAI can identify locations that need infrastructure improvements to protect pedestrians and cyclists.
• Health-care facility inspections — The California Department of Public Health (CDPH) is testing GenAI to make inspection report writing faster and easier for its inspectors.

In April 2025, Newsom announced first-in-the-nation agreements moving the traffic, road-safety, and CDTFA call-center projects into production, involving partners including Microsoft (Azure OpenAI for Caltrans traffic analysis) and delivery vendors implementing an AI assistant for the CDTFA call center. These are not chatbots bolted onto a website — they are high-volume, backlog-heavy, rule-bound workflows where AI accelerates the work while a human stays accountable for the decision. That is exactly the profile that benefits from public-sector agentic AI built around review, escalation, and audit rather than full automation, with human-in-the-loop checkpoints on every consequential output. Most of these use cases also depend on retrieval over authoritative state records, which is why government RAG knowledge assistants — grounded in inspection manuals, tax code, and traffic data — are the dominant architecture rather than open-ended generation.

CDT procurement and California AI transparency laws (AB 2013)

California's procurement rules for GenAI are stricter than its rules for ordinary IT, by design. Two constraints define the buying process.

First, a written solicitation is required every time a state entity purchases GenAI. Acquisition shortcuts that normally speed low-dollar buys — most notably the Fair and Reasonable Acquisition Method — are prohibited for GenAI. There is no quiet path to a sole-source AI tool. Second, before any purchase, the agency must complete the SIMM 5305-F risk assessment, and every state entity must inventory all GenAI uses — intentional and incidental — and report that inventory to CDT. An "incidental" use, such as a GenAI feature embedded in a SaaS product an agency already owns, still has to be surfaced. For a full walkthrough of how these vehicles and gates work across states, see our government AI procurement guide.

The transparency layer arrived on top of procurement. AB 2013, the Generative Artificial Intelligence Training Data Transparency Act, was signed September 28, 2024 and took effect January 1, 2026. It requires developers of generative AI systems made available to Californians to publish high-level documentation of the data used to train the system, covering roughly twelve categories — the sources or owners of the datasets, whether the data includes copyrighted or trademarked material, whether it contains personal information as defined under the CCPA, and how the data serves the system's purpose. The obligation reaches back to any covered GenAI system offered to Californians since January 1, 2022, and disclosures must be updated on substantial modification.

AB 2013 changes procurement in practice, not just in theory. An agency buying a foundation model or a GenAI-powered application now has to consider whether the vendor has met its transparency obligations — because a vendor that cannot document its training data is a compliance and legal-exposure risk the agency inherits. California's broader posture, including model-transparency and safety measures aimed at the largest developers, reinforces the direction. Public-sector buyers now shortlist vendors partly on how defensibly they can answer "what is this model trained on?" — a discipline that mirrors the documentation demands of the EU AI Act compliance regime taking effect the same year.

StateRAMP and the cloud-authorization path for California agency AI

AB 2013 and SIMM 5305-F govern what an agency can buy; cloud-security authorization governs how it can run. Most California agency AI workloads touch constituent data — tax records, benefits eligibility, health-inspection findings, roadway sensor feeds — which means the underlying cloud has to clear a recognized security baseline before an authorization to operate is realistic.

StateRAMP — rebranded to GovRAMP in 2025 — is the standard path for demonstrating cloud security to state and local government buyers, the state-and-local analogue to the federal FedRAMP program. California has not imposed a single blanket StateRAMP mandate the way some states have, but the framework is the recognized route, and agencies and localities increasingly require or prefer it in solicitations. For a GenAI system on constituent data, the practical route to production is a GovRAMP-authorized cloud baseline plus alignment to the NIST AI Risk Management Framework and the SIMM 5305-F risk classification. We break down how these overlap in our guide to StateRAMP and GovRAMP AI compliance, and where the higher federal bar applies in FedRAMP AI compliance.

The compliance stack determines the deployment architecture. A High-risk classification under SIMM 5305-F, or workloads involving especially sensitive data, can push an agency toward isolated or sovereign AI deployment models and dedicated AI data centers for government workloads rather than a shared commercial cloud tenancy. Those are not model decisions — they are infrastructure and staffing decisions, and they are where most agency timelines slip.

Staffing AI delivery for California state and local government

California has the policy, the use cases, the procurement discipline, and the transparency law. What it does not have — and cannot buy on the open market at state pay scales — is enough of the specialized engineering talent to build and run these systems. This is the real bottleneck, and it is the same one we document in the broader enterprise AI talent gap: the people who can take a GenAI pilot through a SIMM 5305-F assessment, a written solicitation, a GovRAMP boundary, and into production are scarce and expensive, and civil-service compensation cannot compete for them.

The answer most agencies and their integrators land on is not permanent hiring. It is forward-deployed engineers and staff augmentation embedded directly into agency or prime-contractor delivery teams. A forward-deployed engineer sits inside the program, learns the agency's data and constraints, and ships against them — which is why the model has become the default for forward-deployed engineers in government. It also changes the cost structure: the agency buys delivery capacity at an engagement rate rather than absorbing a permanent salary load, an advantage we quantify in AI staff augmentation vs. hiring.

This is Gain America's role in California. We staff and deploy public-sector-ready forward-deployed engineers, MLOps and AgentOps specialists, and data-center teams into California state and local agencies and the system integrators holding the contract — including the primes and small-business subcontractors that DGS procurement pathways favor. Our engineers are matched to the exact stage a program is stuck at: standing up the retrieval architecture behind a translation or inspection assistant, building the human-in-the-loop review layer a High-risk classification demands, instrumenting evaluation and observability so the CDTFA-style call-center assistant stays accurate under load, or sizing the compute for an isolated deployment. California wrote the most complete public-sector AI framework in the country. Gain America supplies the delivery capacity that turns that framework into systems in production.