California just ordered its agencies to accelerate independent AI oversight and advance an emergency shutoff for frontier models. Call it a kill switch if you want. The Sep 18 order does not flip the switch today. It tells experts to recommend how to require one, how to put independent auditors inside labs, and how to expand incident reporting when models go sideways.
Washington already drew a different map. In December 2025 the White House framed a national AI policy and stood up tools aimed at state laws that conflict with that frame, including litigation and a path to federal preemption.
If you buy technology for a state, a city, a county, a transit agency, or a campus, you now sit between those two clocks.
One clock says: build coercive stop authority at the state line. The other says: those state rules may not survive. Your RFP, your MSA, and your incident playbook cannot assume a single rulebook.
This is not abstract. Kill switch language without preemption clarity is theater. Preemption without operational stop authority is also theater. SLED buyers need both answered in writing: who can order a stop, how fast, against which systems, and what happens if a federal claim overrides the state clause you negotiated.
Start at Zero still holds. Needs first. Controls before slogans. Write the stop hierarchy into the deal before you score the demo.
What California ordered
On Sep 18, 2026, Governor Newsom issued an executive order to accelerate independent oversight of AI and to advance the creation of an emergency shutoff for frontier models. The order is not itself a statutory kill switch that fires today. It sets a clock: expert recommendations on how to require a shutoff, how to put independent auditors onsite at labs, and how to widen incident reporting when systems misbehave. Roughly two months for that advice to land.
For a SLED buyer, the useful read is not the culture-war headline. It is the implication for contracts you are signing now. If your state is building stop authority, vendors will face new duties. Your agency may inherit new notification and testing expectations. None of that is wait-and-see if you are mid-selection.
What Washington already set in motion
In December 2025, the federal executive order on a national AI policy framework pointed a different direction. It framed a national approach and created tools aimed at state AI laws that conflict with that frame, including litigation capacity and a legislative path toward preemption.
That is the second clock. A clause that assumes California (or any state) gets the last word on stop authority may not be the clause that survives. A clause that assumes federal uniformity may not match the operational reality of a state that is designing its own shutoff.
The procurement collision
Two sovereign claims can land on the same stack: the model, the agent, the SaaS feature, the on-prem appliance, the logs, and the human approval path.
Your RFP assumed one AI rulebook. You no longer have one.
The risk is not “too much regulation.” The risk is buying as if only one capital gets to write the rules. When stop language and preemption language both stay vague, the vendor owns the ambiguity and you own the outage, the public records request, and the council hearing.
Procurement sitting between those clocks has to put the controls in the contract. The AI contract clause has to be answered in writing before anyone scores a demo.
SLED buyer scorecard
Paste these into RFP language and due diligence before you score demos. AI in local government still has to survive the same list.
- Stop authority. Who can disable which component, in what SLA, without vendor permission?
- Scope. Is the stop aimed at the model, the agent, a SaaS feature, or an on-prem appliance?
- Evidence. How is stop efficacy tested and logged, and can the buyer see it?
- Preemption contingency. What does the vendor owe if state stop rules are enjoined or preempted?
- Incident reporting. What must reach the buyer vs the state vs the feds, and on what clock?
- Indemnity and continuity. Who pays when compliance paths conflict, and how do you keep the system of record running while the fight happens?
If a vendor cannot answer these in writing, you are not buying a control. You are buying a press release.
Close
The danger for public buyers is not picking a tribe. It is treating stop authority and preemption as someone else’s problem until go-live week.
Write the hierarchy into the deal now: who can force a shutoff, how you prove it works, and what happens when Sacramento and Washington disagree.
Needs first. Controls before slogans. Then score the demo. That order is the same one in Going to Market Is a Decision, Not a Diagnostic. If the clock in the letter is an end-of-support date rather than a kill switch, the companion essay is Manufactured Urgency.
If this was useful, I write for SLED buyers who still have to live with the contract after the headline fades. Start at Zero. Advisory path: Government AI Navigator.
If you are mid-selection and the stop clause is still a slogan, write to us. A sentence is enough. If we cannot help, we will say so. Services are buyer-side and vendor-neutral.
AV, Avero Advisors