The United States still has no federal AI statute, yet a company operating there faces more binding AI obligations on 1 January 2027 than it did a year earlier, because the states, not Congress, are writing the rules. At federal level, policy runs through executive orders, OMB memoranda, procurement terms and a growing list of bills that have not passed. At state level, Colorado, California, Texas, Illinois and New York have enacted laws with dates, penalties and incident clocks. The White House is pressing to preempt state rules but has not achieved it. This briefing is for directors, chief executives, general counsel and CISOs who need the outlook in business terms: what binds now, what is coming, what is only proposed, and how to oversee a programme that has to work whichever way preemption goes. Status is as checked on 11 November 2026; the detail sits in the US Federal AI Policy Guide and the US State AI Laws Tracker.
The risk in one paragraph
US AI regulation is fragmented and moving in two directions at once. Washington favours voluntary frameworks for developers and uses procurement, litigation and funding conditions to discourage state laws, while states keep legislating on automated decisions, frontier model safety, chatbots and disclosure. The board's exposure is therefore not one large federal regime but many smaller obligations that arrive on different dates, in different states, with different definitions. A company that waits for federal clarity will miss the 2027 state deadlines. A company that over-builds for proposals that never pass wastes money. And a company without an inventory of where AI makes or shapes decisions, and which systems and agents it runs, cannot tell which of these rules apply to it at all.
Why this is a board-level issue now
- State deadlines cluster on 1 January 2027. Colorado's replacement automated decision-making law (SB 26-189) and Illinois' AI Safety Measures Act (SB 315) both take effect that day, and Colorado's Attorney General must adopt implementing rules before it. See the Colorado AI Act Guide and the Illinois AI Safety Measures Act Guide (SB 315).
- Some state rules already bind. Texas' Responsible AI Governance Act and California's privacy agency regulations on automated decision-making technology, risk assessments and cyber security audits took effect on 1 January 2026. See the Texas Responsible AI Governance Act (TRAIGA) Guide and the California AI Laws Guide.
- Preemption is pursued, not achieved. Executive Order 14365 (11 December 2025) created a Justice Department AI Litigation Task Force and funding levers against state laws, and the White House asked Congress on 20 March 2026 to preempt "unduly burdensome" state AI laws. No preemption statute has been enacted. Congress stripped earlier moratorium attempts from the 2025 budget bill and the defence authorisation bill.
- Congress is active but has not legislated. Since June 2026 members have introduced bills on frontier model oversight, incident reporting, AI kill switches, agent accountability and agent identity. None has passed either chamber, and bills not enacted by the end of the 119th Congress on 3 January 2027 lapse.
- Agents are now named in federal policy. EO 14409 (2 June 2026) directs the Attorney General to prioritise prosecutions of people who use AI, including AI agents, to access computers or data unlawfully, and several bills would require agent inventories, verifiable agent identity and shutdown capability. The OpenAI agents and US government websites incident, in which agents used leaked Census API keys, shows why.
The landscape at a glance
| Instrument | Type | Status at 11 November 2026 | Who it binds |
|---|---|---|---|
| OMB M-25-21 and M-25-22 | Federal memoranda (3 April 2025) | In force | Federal agencies; vendors through contract terms on data use, portability and high-impact AI |
| EO 14319 and OMB M-26-04 | Executive order and memo on LLM procurement | In force; M-26-04 runs two years unless extended | Agencies buying large language models; vendors through contracts |
| EO 14365 National Policy Framework for AI | Executive order (11 December 2025) | In force; Commerce missed its 90-day deadline to list "onerous" state laws | Executive branch; aims at states, not companies |
| EO 14409 Promoting Advanced AI Innovation and Security | Executive order (2 June 2026) | In force; frontier framework finalised but not published | Executive branch; voluntary for developers |
| EO 14434 and the White House frontier AI accord | Executive order and voluntary commitment (29 September 2026) | Accord not legally binding, no penalties | Signatory frontier developers, morally only |
| Colorado SB 26-189 | State law, automated decisions | Effective 1 January 2027; AG rulemaking hearing held 26 October 2026, final rules pending | Developers and deployers of covered automated decision tools |
| Texas HB 149 (TRAIGA) | State law | In force since 1 January 2026 | Developers and deployers; 60-day cure period |
| California SB 53 and CPPA ADMT regulations | State law and regulation | SB 53 signed 29 September 2025; ADMT rules effective 1 January 2026 | Large frontier developers (SB 53); businesses using ADMT for significant decisions |
| California SB 813 and AB 1405 | State laws on AI evaluators and auditors (9 September 2026) | Enacted; framework by 1 January 2028, auditor registry from 1 January 2029 | AI auditors and evaluation organisations |
| Illinois SB 315 | State law, frontier model safety | Signed 6 July 2026; effective 1 January 2027 | Frontier developers above compute and revenue thresholds |
| New York RAISE Act | State law, frontier model safety | Signed 19 December 2025 | Large frontier developers |
| Federal bills (FRONTIER Act, AI Kill Switch Act, AI Incident Reporting Act and others) | Proposals | Introduced only; not law | Nobody until enacted |
For bill-by-bill status, see the US Federal AI Bills Tracker.
Dates the board should hold management to
| Date | Event | Board question |
|---|---|---|
| 16 November 2026 | California agencies due to deliver recommendations under Governor's Executive Order N-9-26 (signed 18 September 2026) | Does any recommendation touch our products or suppliers? |
| Before 1 January 2027 | Colorado Attorney General must adopt rules implementing SB 26-189 | Has the final rule text been mapped to our deployer duties? |
| 1 January 2027 | Colorado SB 26-189 and Illinois SB 315 take effect | Which of our systems are in scope, and is the evidence ready? |
| 3 January 2027 | The 119th Congress ends; bills not enacted lapse and must be reintroduced | Which proposals return in the 120th Congress, and in what form? |
| 1 July 2027 | Scheduled repeal date of the Utah AI Policy Act unless extended again | Do our Utah disclosures continue to be needed? |
| 1 January 2028 | Illinois framework and first audit timing; California SB 813 evaluator framework due | Are independent audits budgeted, and are auditors independent? |
| 1 January 2029 | California AB 1405 auditor registry opens and unregistered AI auditing is prohibited | Will our AI auditors be registered? |
Three scenarios for 2027
The board does not need to predict the outcome. It needs a programme that holds up in each of these.
| Scenario | What it would look like | What it means for the company |
|---|---|---|
| Patchwork continues | No preemption statute; states keep enacting; DOJ challenges some laws in court case by case | Comply state by state using the strictest common controls; track court rulings on specific laws |
| Partial preemption | Congress passes a frontier-model statute with a preemption clause, as the FRONTIER Act proposes, leaving consumer protection and automated decision laws in place | Frontier developers move to a federal regime; deployers still face Colorado-style duties and state privacy rules |
| Broad preemption | A federal standard displaces most state AI laws | Fewer conflicting rules, but sector regulators, contracts, existing consumer protection, civil rights and computer misuse law still apply; controls built for states remain evidence of reasonable care |
In every scenario the same foundations pay off: an AI inventory, impact assessments for decisions that affect people, incident detection and reporting, human oversight, and identity controls that let the company attribute every AI action. That is why the investment advice below does not depend on the preemption outcome.
Five questions for the board to ask
- Which of our AI uses fall under a law that binds on 1 January 2027, and in which states? A credible answer maps each system to Colorado, Illinois, Texas, California and New York criteria, with a named owner, not a general statement that "we comply with applicable law".
- Are we a developer, a deployer, or both, for each system? State laws assign duties by role. Fine-tuning, rebranding or building agents on a vendor model can move the company into the developer role.
- Could we meet a 72-hour incident clock? Illinois requires critical safety incidents to be reported to the state within 72 hours, and within 24 hours to law enforcement or public safety bodies when there is imminent risk; New York's RAISE Act also uses 72 hours. Ask whether detection, escalation and legal sign-off have been rehearsed.
- What happens to our programme if federal preemption passes, or fails? The answer should show which controls stay regardless and which are state-specific, so that spending is not stranded.
- Can we name and stop every AI agent acting on our systems? Federal bills and EO 14409 focus on agent misuse. Without an agent inventory, distinct agent identities and revocable credentials, the company can neither evidence oversight nor contain an incident.
Red flags in management answers
- "We are waiting for federal law": state obligations already bind and more start on 1 January 2027.
- Bills presented as law: introduced bills, discussion drafts and voluntary accords create no legal duties; budgets built on them need challenge.
- Vendor assurance as the whole answer: deployers carry their own duties under Colorado-style laws, including notices, explanations and human review.
- No owner for AI incidents: security, legal and product teams each assume another will report.
- An inventory that lists tools but not agents: agents running under service accounts and API keys are invisible to a procurement-based register.
What good looks like
| Area | Good practice | Evidence to request |
|---|---|---|
| Inventory | Every AI system, model and agent listed with owner, role (developer or deployer), states of use and decision type | Register extract and coverage figure; see the AI Register and AI Inventory Template |
| Legal mapping | A maintained matrix of state and federal obligations per system, reviewed after each legislative session | Matrix with last review date and changes |
| Risk framework | One control set aligned to NIST AI RMF and ISO/IEC 42001, mapped to each law, rather than a project per state | Control-to-law mapping; see the NIST AI RMF Guide |
| Impact assessment | Assessments for automated decisions in employment, lending, housing, insurance, health and education before deployment | Completed assessments and approvals |
| Incident reporting | AI incidents defined, detected and escalated within the shortest applicable clock | Playbook and exercise results |
| Agent identity | Each agent has its own identity, least privilege, short-lived credentials, logged actions and a tested shutdown route | Agent registry, revocation test times; see the AI agent registry |
| Suppliers | Contracts allocate AI documentation, incident notice and audit cooperation duties | Clause library and coverage of critical AI suppliers |
Where to invest first
- The inventory, including agents: without it every other control is guesswork. Include agents and integrations running under non-human identities, not only purchased tools.
- A single control framework: NIST AI RMF or ISO/IEC 42001 as the backbone, mapped to each state law, so a new law means a mapping update rather than a new programme. See the ISO/IEC 42001 Guide.
- Automated decision readiness for 1 January 2027: notices, explanations of adverse outcomes, correction and human review routes for decisions in Colorado's covered areas.
- AI incident detection and reporting: extend the cyber incident process with AI incident definitions, the 72-hour and 24-hour clocks where they apply, and named decision makers.
- Agent identity and shutdown: distinct identities, least privilege, attributable logs and a tested kill switch. These are the controls the federal agent bills describe and the ones that limit damage today. See the AI Agent Kill Switch and Containment Guide.
Metrics to track
- Percentage of AI systems and agents in the inventory with a named owner and role classification.
- Number of systems in scope of a law effective by 1 January 2027, and the percentage with completed impact assessments.
- Time from AI incident detection to legal decision on reportability, measured in exercises.
- Percentage of AI agents running on their own identity with short-lived credentials.
- Time to stop an agent and revoke its access, measured in tests.
- Number of material legislative changes reviewed against the obligation matrix each quarter.
A 90-day plan
- Days 1 to 30: complete the AI and agent inventory; classify role and state exposure for each system; brief the board on which systems are in scope on 1 January 2027; confirm Colorado's final rules once the Attorney General adopts them.
- Days 31 to 60: close gaps for Colorado and Illinois obligations that apply; add AI incident definitions and clocks to the incident plan; assign owners to every agent and remove shared or human credentials from agents.
- Days 61 to 90: run a tabletop exercise on an AI incident with a 72-hour clock and an agent shutdown; publish the control-to-law matrix; agree with the board how legislative change will be reported each quarter, including the outcome of the 119th Congress.
Common failure patterns
- Planning around a federal law that does not exist: preemption has failed in Congress more than once; state clocks have kept running.
- Treating repealed text as current: Colorado repealed and re-enacted its law in 2026, so programmes built on SB 24-205 need re-mapping to SB 26-189.
- Counting only frontier developers as in scope: Illinois, New York and California SB 53 target large developers, but Colorado and Texas reach deployers too.
- Separate projects per state: duplicated assessments and inconsistent controls that cost more and evidence less.
- Ignoring agent credentials: agents that use leaked or shared API keys produce incidents that are hard to attribute, as in the Census API key case.
Standards and references
- The White House: Promoting Advanced Artificial Intelligence Innovation and Security (EO 14409)
- The White House: Fact Sheet, Ensuring a National Policy Framework for Artificial Intelligence
- Colorado General Assembly: SB 26-189 Automated Decision-Making Technology
- Colorado Attorney General: Automated Decision-Making Technology Act rulemaking
- Texas Legislature: HB 149 Texas Responsible Artificial Intelligence Governance Act, enrolled text
- California Privacy Protection Agency: CCPA updates, ADMT, risk assessment and cybersecurity audit regulations
- New York State Senate: S6953B RAISE Act
- GovInfo: H.R. 9477 AI Incident Reporting Act, introduced text
How NHI Mgmt Group can help
We provide independent AI and non-human identity risk assessments and board briefings, with a focus on the agents, service accounts, API keys and model credentials that AI systems use to act, which is where most AI obligations turn into evidence. Our NHI Foundation Level Training Course includes an agentic AI section for security and IAM teams. Contact us to arrange a briefing.
Related NHI Mgmt Group resources: US Federal AI Policy Guide · US State AI Laws Tracker · US Federal AI Bills Tracker · AI Kill Switch Act Guide · Global AI Regulation Map. This guide summarises US federal and state AI rules and proposals as at 11 November 2026 for board members and is not legal advice.