Public-sector AI governance: how California's GovOps runs the model

What GovOps is, and why AI governance lives there

The California Government Operations Agency (GovOps) runs the state's back-office functions — procurement, information technology, and human resources — across a set of component departments and offices. Among them are the California Department of Technology (CDT), the Department of General Services (DGS), the Department of Human Resources (CalHR), and the Office of Data and Innovation (ODI). DGS is the state's primary procurement arm; CalHR has been part of GovOps since the agency was created on July 1, 2013.

Generative AI first reached state government as something to be bought, used by employees, and run on state systems — that is, as a procurement, workforce, and technology question all at once. So the natural home for governing it was the agency that already owns all three. Per the GovOps site, the agency leads the state's generative-AI work across the administration. The consequence is structural: California places AI governance inside operations, rather than standing up a separate AI regulator.

Two executive orders set the direction

The state's AI program runs on two Newsom executive orders signed about two and a half years apart. EO N-12-23 (September 6, 2023) launched it; EO N-5-26, titled 'Trusted AI Procurement' (March 30, 2026), extends it and builds explicitly on the first. Neither is a comprehensive AI statute. Both work by directing the executive branch — and, through procurement, the vendors that sell to it.

• Benefits and Risks of Generative AI report (November 2023).
• State of California GenAI Guidelines for Public Sector Procurement, Uses and Training (March 2024).
• Building a GenAI Ready State Workforce (July 2024).
• Choose Your GenAI Journey for State Entities (February 2025).

The 2024 GenAI Guidelines: buy for a need, test for harm

The State of California GenAI Guidelines for Public Sector Procurement, Uses and Training, released March 21, 2024, are the operational core of the program. They were authored jointly by GovOps, CDT, DGS, ODI, and CalHR, and they implement EO N-12-23. They were issued as interim guidance, with final procurement and training policy targeted for 2025 after piloting.

The Guidelines reorder the usual buying reflex: they require a state entity to establish a genuine business need before it procures GenAI, then to assess the risks and impacts of the intended use, test models to reduce bias and errors before deployment, and stand up a GenAI team to keep evaluating the system after it goes live. A planned deployment must complete a Generative AI Risk Assessment — catalogued as SIMM 5305-F — to gauge its risk exposure. The emphasis throughout is on the context of use and the full lifecycle, not on the technology in the abstract.

• Establish the business need before buying — procurement starts from a problem, not a product.
• Assess the risks and impacts of the proposed use.
• Test models to reduce bias and errors before deployment.
• Complete a Generative AI Risk Assessment (SIMM 5305-F).
• Stand up a GenAI team to evaluate the system continuously after go-live.

Trusted AI Procurement: what N-5-26 directs

EO N-5-26 sharpens the procurement lever, but with an important caveat about what it actually does. Most of its substance is a set of directives to state agencies to develop standards within roughly 120 days (reported as a late-July 2026 deadline) — not requirements already in force. The order is forward-looking, applies to vendors nationwide that want to sell AI to California, and works by conditioning the state's purchasing rather than by imposing statutory duties.

The centerpiece is a directive to DGS and CDT to develop vendor certifications under which contractors would attest to, and explain, their policies and safeguards in three areas: preventing illegal content, such as child sexual abuse material and non-consensual intimate imagery; avoiding models that display harmful bias or lack governance to reduce it; and protecting civil rights and civil liberties such as free speech, voting, human autonomy, and protections against unlawful discrimination, detention, and surveillance. As written, this is a certification to be developed and recommended — not yet a live mandate that vendors must already satisfy.

Two related directives round it out. GovOps, working with DGS and CDT, is to propose 'contractor responsibility' reforms aimed at entities judicially determined to have unlawfully undermined privacy or civil liberties. Separately, the CDT State Chief Information Security Officer is to independently review federal supply-chain risk determinations so that state procurement can continue where appropriate — a provision that some analysts note reportedly raises questions about how state and federal AI-procurement regimes fit together.

Workforce enablement: vetted assistants and mandatory training

A recurring theme across both orders is enabling employees, not merely restricting them. EO N-12-23 authorized state employees to experiment with GenAI tools, and the training track became concrete in 2024. Under the GenAI Guidelines, training is mandatory for executive and procurement teams; from March 29, 2024, state purchasing officials could access Procurement GenAI Training through the California Procurement and Contracting Academy (CalPCA). GovOps also published Building a GenAI Ready State Workforce in July 2024.

EO N-5-26 extends the enablement side further. It directs GovOps component agencies to facilitate employee access to vetted GenAI tools with security safeguards and to expand employee AI training — in other words, a sanctioned, safeguarded path to using assistants rather than an ad-hoc or shadow-IT one. Enablement and control are treated as two sides of the same policy, not as competing goals.

Citizen-facing AI: a directed pilot and transparency guidance

The newest direction points the tools outward, toward residents — but with a precise reading. Under EO N-5-26, agencies are directed to develop a pilot GenAI-powered public portal organized around life events, and the order names examples like disaster relief, starting a business, and finding a job. This is a directive to pilot such a service, not evidence of an already-live public tool; treat it as stated intent rather than a shipped capability.

On transparency, CDT and GovOps are directed to issue watermarking guidance for AI-generated or significantly manipulated images and video, referencing California Business and Professions Code sections 22757.2 and 22757.3. The same order directs agencies to publish a data-minimization toolkit and to update the State Digital Strategy — pushing provenance and data restraint into the state's own service delivery, not just its procurement.

How the GovOps model compares to NIST AI RMF and the EU AI Act

The GovOps approach belongs to the same family as the major AI-governance frameworks, even though its lever is different. Where the EU AI Act (Regulation (EU) 2024/1689) is binding regulation and the NIST AI Risk Management Framework 1.0 is voluntary guidance, California governs primarily through what it will buy and how its own employees may use AI. The mechanism differs, but the substance rhymes in ways worth naming precisely.

The Guidelines' business-need-then-risk-assessment sequence mirrors the risk-based posture of both instruments, which judge a system by how it is deployed rather than in the abstract. The standing GenAI team maps onto the NIST framework's Govern, Map, Measure, and Manage functions — governance and measurement as ongoing activities, not a one-time gate. The watermarking guidance parallels the EU AI Act's transparency obligations around AI-generated content. And the certification and accountable-team ideas keep a named human on the hook, consistent with the human-oversight thread that runs across these frameworks.

• Reach: the GovOps model binds only those who sell to, or work for, the state — narrower than a law of general application.
• Maturity: much of EO N-5-26 is still direction to develop standards, so parts of the model are intent rather than enacted rule.
• Form: California leads with procurement and internal policy, where the EU AI Act leads with statutory obligation and NIST with voluntary guidance.

Frequently asked questions