Four Orders, Three States, Three Models of Obligation
The four orders differ in structure, but each one transfers work to state technology, procurement or governance offices that wasn't there before.
California EO N-5-26 (signed March 30, 2026) directed the Department of General Services and the California Department of Technology to submit recommendations for new AI vendor certification requirements within 120 days. The deadline was July 28, 2026. The order asked for recommendations on new certifications that may be incorporated into state contracting, allowing vendors to attest to and explain their safeguards in at least three areas: exploitation or distribution of illegal content such as CSAM and non-consensual intimate imagery; models that display harmful bias or lack governance to reduce it; and violations of civil rights and civil liberties. The order also directed GovOps to propose reforms to contractor responsibility provisions, including suspension and ineligibility authorities for entities judicially determined to have unlawfully undermined privacy or civil liberties. Separately, per the same order, GovOps, including CDT, ODI, DGS and CalHR, was directed to facilitate employee access to vetted generative AI tools and publish a data minimization toolkit. California's primary control surface is the contract: governance would flow through what the state asks vendors to certify before an award.
California EO N-9-26 (signed September 18, 2026) builds on SB 53 and California's new independent-verification-organization law (it cites N-5-26 among prior actions). It directs GovOps to complete the independent-verification-organization requirements of Government Code Section 8898.1 by May 1, 2027, and, in consultation with Cal OES, to submit recommendations by November 16, 2026 on amendments to state AI safety law, including requiring a "kill switch" for frontier models whose efficacy is verified on an ongoing basis by an independent verification organization. It regulates frontier developers through statute rather than through state procurement, but California vendors will see both regimes.
Oregon EO 26-26 (signed September 23, 2026) placed obligations on the state's Chief Information Officer directly. The CIO has 90 days to propose a working definition of "frontier AI" for state government purposes, establish criteria for what constitutes adequate third-party safety review of those systems, and assess the viability of requiring kill-switch capability on frontier AI systems under state contracts. Governor Kotek cited the absence of adequate federal oversight as the reason for state action. The order remains in effect until the Governor terminates it, and the Governor must reassess its necessity every three months, so the framework could be retained, modified or supplemented depending on what federal action materializes. Its core policy is that the State "shall seek to purchase, contract for, or otherwise utilize frontier artificial intelligence models that have received independent, third-party review for artificial intelligence safety" (order text, as published via the Governor's newsroom).
Illinois EO 2026-07 (signed September 22, 2026) established the Illinois Artificial Intelligence Cabinet, an advisory body of academic and expert partners joined by senior leaders from the Department of Innovation and Technology, IEMA–Office of Homeland Security, the Department of Financial and Professional Regulation, the Commerce Commission, the State Police, the Department of Public Health and the Illinois EPA. Governor Pritzker is to appoint members within 30 days, and DoIT provides administrative support. The Cabinet is charged with assessing AI-related risks, advising on procurement safeguards, and developing guidance on AI incident response. Illinois chose a deliberative model rather than a deadline model: it convened the governance structure first, with substantive recommendations to follow.
These are not equivalent levels of urgency. Oregon's CIO faces a December 2026 deadline to produce something specific. California's N-5-26 deadlines have passed; its follow-on N-9-26 sets a November 16, 2026 recommendation deadline, and the question for state agencies is now whether any recommendations submitted in July have been incorporated into procurement templates. Illinois is still standing up its cabinet.
The Definition Problem Oregon Gave Its CIO
Oregon's most technically demanding assignment is the definition of "frontier AI." The term is generally used for the most capable general-purpose models; Oregon's own FAQ describes them as the most powerful models and the ones with significant potential risk. For a state CIO, the definition must be operationally actionable — specific enough that a procurement officer can determine whether a given contract falls inside it.
A workable framework for state purposes might categorize AI systems along two dimensions: consequence level and reversibility. A model that scores unemployment claims, ranks child welfare referrals, or generates real-time recommended actions for emergency dispatch is both high-consequence and low-reversibility — an error cannot easily be undone, and the affected person may not have a meaningful review mechanism before harm occurs. These systems warrant the oversight Oregon envisions regardless of their technical sophistication. A model that summarizes public comments on a proposed regulation is high-volume but low-consequence; errors are caught in the editorial process.
A capability-based threshold may also apply, particularly for general-purpose models procured on a subscription basis rather than purpose-built systems. But for state government, the operational consequence framework captures the relevant risk better than capability metrics alone: a highly capable general-purpose model used for scheduling has a different risk profile than a narrower model used for risk-scoring children.
Oregon's CIO will need to inventory current state AI systems against whatever definition emerges. Building that inventory from scratch after the proposal is drafted will compress the subsequent procurement revision work into an impossible window. The inventory should be underway now, in parallel with the definitional work.
Kill-Switch Requirements: What the Assessment Must Address
Among the orders discussed here, only Oregon's EO 26-26 frames a kill switch as a procurement standard for state contracts. California's EO N-9-26 (September 18, 2026) separately directs GovOps to recommend, by November 16, 2026, on requiring frontier developers to build a "kill switch" whose efficacy is verified by an independent verification organization. Oregon's CIO must assess viability, not mandate implementation — but the assessment requires understanding what implementation would actually entail, since a kill switch that can't be exercised in practice is not a governance control.
Viable kill-switch capability on a frontier AI system under a state contract would require at least four things:
Unilateral suspension authority in the contract. Current state IT master services agreements were not designed with this in mind. A kill-switch provision needs to specify: who in the state organization has authority to invoke it, what notice (if any) the vendor receives, the timeframe within which the system must be suspended, and what the vendor's obligations are during suspension. These are novel terms. Procurement offices need to draft them before vendors have leverage in negotiations over a specific system.
An operational continuity plan without the AI. A kill switch is only a governance control if the underlying function can continue when the AI is suspended. For AI systems embedded in eligibility determinations, case management, real-time dispatch, or infrastructure monitoring, this requires a documented manual or automated fallback. If no fallback exists, then the choice is not "use the AI or don't" — it's "use the AI or suspend the function." That's a different risk calculus, and it means the kill switch is a crisis measure rather than a routine governance tool. Identifying which programs have this gap is part of the assessment.
Audit trail and incident documentation requirements. The decision to invoke a kill switch is a significant event. Contracts need to specify what records must be maintained, what the handoff process looks like, and what post-suspension review the vendor owes. Without these terms, the kill switch has no accountability structure.
A jurisdictional scope question. If the frontier AI system is a general-purpose model procured from a major vendor on a subscription basis, the "kill switch" question is really a contract termination question — the state can stop using the service. If it's a custom model integrated into state systems, suspension has technical dependencies that must be planned for. The assessment Oregon requires will need to distinguish these cases.
None of this is impossible to build. But it requires procurement templates, legal review, and technical architecture decisions that take months. Oregon's CIO has until late December. Starting with the inventory of what systems would fall under the requirement is the right first move.
How These Mandates Should Change Procurement Templates
Three of the four orders (California N-5-26, Oregon EO 26-26 and Illinois EO 2026-07) converge on procurement contracts as the implementation vehicle; California N-9-26 works through statute instead. This is logical: most AI systems in state government arrive through a vendor agreement, and the contract is the most durable governance instrument a state has over vendor behavior after award.
Taken together, the procurement-focused orders point toward a set of contract provisions that state procurement offices should begin developing:
- Vendor attestation requirements: attestations on training data governance, bias testing protocols, and performance monitoring before award (a Spartan X recommendation, not a requirement of any of these orders). California's N-5-26 asked for recommended certifications covering illegal content, harmful bias and civil-rights and civil-liberties safeguards; Oregon's implementation proposal is expected to add independent third-party safety review for frontier models.
- Third-party review terms: once Oregon's criteria are published, contracts for systems meeting the frontier definition will need to require an independent review before the system is deployed in production, with findings disclosed to the state.
- Suspension authority clauses: unilateral suspension rights with defined timelines, notice requirements, and handoff procedures — distinct from termination-for-cause provisions, which in our view are a poor fit for a precautionary pause because they generally depend on a vendor default.
- Incident response obligations: what constitutes a reportable AI incident, how quickly the vendor must notify the state, what documentation must follow, and what remediation is required.
- Performance benchmarking: measurable performance floors tied to the specific operational context (eligibility accuracy, decision consistency, appellate reversal rate) rather than generic model benchmarks.
None of these terms is standard in current state IT procurement. The states with published AI EOs will need to develop model contract language through their procurement agencies; the states without them will need to develop it themselves or wait for a peer state's template to emerge. California's process, being the most advanced, is the most likely source of a practical template that other states can adapt.
A Practical Sequence for CIO Offices
For CIOs responding to an existing order, or preparing for one:
1. Inventory current AI systems against a working frontier-AI definition. Don't wait for the official definition. Categorize deployed and contracted AI systems by consequence level and reversibility. Systems making or recommending irreversible decisions affecting residents — benefits, custody, criminal justice, public safety dispatch — are the priority tier regardless of how the official definition resolves.
2. Audit existing contracts for suspension and termination provisions. Review AI vendor contracts for unilateral suspension authority. We expect most will not. Identify which contracts are up for renewal within 18 months — those are the near-term revision opportunities.
3. Identify which programs lack an operational fallback. Map the gap between "we could suspend this AI system" and "we could continue operations if we did." Where the fallback is missing or undocumented, that's a program risk independent of the procurement requirement.
4. Begin building model contract language now. Third-party review criteria, kill-switch provisions, incident response obligations, and performance benchmarking are all drafting tasks that can start before an official standard is published. Using the NIST AI Risk Management Framework (NIST AI 100-1) and IEEE Standard 3119-2025 as scaffolding gives procurement offices a principled starting point that can be refined as state guidance matures.
5. Track California's vendor certification framework. California's 120-day process was due to deliver recommendations to the Governor by July 28, 2026. If and when they are published, they will be among the most detailed state AI procurement standards available to adapt, and adapting them to your state's context is likely faster and more defensible than drafting from scratch.
The pattern established in 2026 — multiple states acting independently, in rapid succession, on AI governance — is not going to reverse. State CIOs who invest in the procurement infrastructure now will be managing known risk. Those who wait will be managing the compressed timeline that comes with an EO they didn't see coming.
Spartan X advises state technology leadership on AI strategy, architecture and governance frameworks, including [work with Colorado's Governor's office, OIT and CDLE](/services/ai). The practical question these orders raise is where a governance control actually lives: in the contract, in the operating procedure or in the system architecture. Inventory and fallback planning are the work that tells you.
Sources and further reading
- Oregon EO 26-26 (September 23, 2026): Governor Kotek Issues Executive Order to Advance AI Safety and Oversight
- Illinois EO 2026-07 (September 22, 2026): Gov. Pritzker Establishes Illinois Artificial Intelligence Cabinet
- California EO N-5-26 (March 30, 2026): 3.30 FINAL Trusted AI Procurement EO N-5-26
- California EO N-9-26 (September 18, 2026): FINAL N-9-26 AI EO
- Illinois EO 2026-07, signed text: Executive Order 2026-07
- Oregon frontier AI FAQ: Oregon Governor's office FAQ
- Oregon EO 26-26, order text: EO 26-26
- NIST AI Risk Management Framework (NIST AI 100-1): nist.gov/itl/ai-risk-management-framework
- IEEE Standard 3119-2025, Standard for the Procurement of Artificial Intelligence and Automated Decision Systems: standards.ieee.org
