State AI laws in 2026: 115 new measures across 30 states

State AI laws in 2026: what the numbers mean for agent teams
State AI laws in 2026 now cover 30 states, with 115 AI-related measures passed or enacted between 1 January and 15 October 2026, according to the Transparency Coalition 2026 State AI Legislation Report. That count includes a small number of measures approved by legislatures but still awaiting a governor’s signature, so it is not a tally of laws currently in force. The total is up nearly 58% from 73 laws across 27 states in the 2025 report.
There is no comprehensive federal AI statute. That gap means the rules your agents must follow depend on where your users are, not on a single national policy document.
What categories the 2026 laws cover
The 115 measures fall into several distinct areas. Understanding which categories apply to your agent’s behaviour is more useful than tracking the raw count.
Chatbot safety and disclosure
Fourteen states advanced chatbot safety measures. California, Oregon, and Washington have the most extensive protections among those 14; the remaining states are not named individually in the report. Separately, disclosure requirements are spreading across categories: several states now mandate that agents identify themselves as AI in advertising, customer service, campaigns, and chatbot interactions. Washington adds a further requirement for provenance data in AI-generated or altered media.
For teams shipping AI agents for customer service, this is the most immediately operational category. An agent that handles support conversations may need to identify itself, log the interaction, and in Washington’s case, carry metadata about any generated or altered content it produces.
Addictive design and minors
Seven states, California, Colorado, Connecticut, Idaho, Illinois, Minnesota, and South Carolina, enacted measures targeting addictive design patterns aimed at children and teenagers. If your agent surfaces in a consumer product with a younger user base, these rules govern how the interface is allowed to engage that audience.
Health insurer AI decisions
Alabama, Colorado, Georgia, Illinois, Indiana, Iowa, Utah, and Washington have limited health insurers from relying solely on AI for adverse authorisation or coverage decisions. This category does not prohibit AI involvement; it requires that a human can be in the loop before a denial stands. Teams building AI agents for workflow automation inside health insurance operations should treat human-review checkpoints as a design requirement, not an optional feature.
AI therapy and mental health
Colorado, Maine, Rhode Island, Tennessee, and Vermont have placed limits on AI therapy applications. The measures address what an agent is permitted to say in a therapeutic context and when a human professional must be involved. If your agent operates in a mental health adjacent setting, these states impose category-specific constraints that a general disclosure policy will not satisfy.
Deepfakes
California, Hawaii, Kansas, Louisiana, Maryland, Tennessee, Utah, and Washington all enacted deepfake measures. These primarily address synthetic media of real people, with particular attention to political content. Agents that generate images, audio, or video of identifiable individuals need to account for these obligations at the output layer.
Education
California, Illinois, Maryland, Oklahoma, Tennessee, Utah, and Virginia enacted education-related AI measures. The obligations vary but generally address disclosure, permitted use cases, and in some states, data handling for student-facing systems.
Frontier models and auditing
Illinois stands out here. The Illinois AI Safety Measures Act requires frontier model developers to operate a risk-management framework and undergo annual independent third-party audits. California enacted standards and infrastructure for independent AI auditing organisations. Both states are moving toward a regulated auditing market; California’s work on auditor infrastructure suggests that certification of auditors, flagged as a likely 2027 development in the Transparency Coalition report, is already being scaffolded.
flowchart TD
A[Agent interaction occurs] --> B{What does the agent do?}
B --> C[Customer service / chatbot]
B --> D[Health insurance decision]
B --> E[Therapy or mental health]
B --> F[Generates synthetic media]
C --> G[Check disclosure rules\nby state]
D --> H[Human review required\nbefore adverse decision]
E --> I[Check therapy limits\nin CO, ME, RI, TN, VT]
F --> J[Check deepfake rules\nin 8 states]
G --> K[Log interaction,\nprovide AI disclosure]
H --> K
I --> K
J --> K
K --> L[State-level compliance record]
Why this creates a per-state configuration problem
The patchwork is not an accident of slow federal action. It reflects genuine disagreement between states about where AI risk sits and who bears it. The practical consequence for agent teams is that a single policy document does not cover your obligations. You need per-state configuration of what your agent says, which actions trigger a human review, and what records you keep.
This is not a future concern. 97% of companies report deploying AI agents in the past year, yet only 11% are running them in production at scale. Part of that gap is governance: teams that built agents for internal pilots did not build the state-level configuration and audit logging that production deployment requires.
The gap is visible in compliance functions too. 83% of compliance leaders report using AI tools, but only 25% have implemented a strong governance framework. That mismatch is precisely where state-level obligations land hardest, because the obligations are behavioural and ongoing, not one-time assessments.
flowchart TD
A[Map user geography] --> B[Identify applicable\nstate categories]
B --> C{Categories present?}
C --> D[Chatbot / disclosure]
C --> E[Health insurer decisions]
C --> F[Deepfakes / synthetic media]
C --> G[Minors / addictive design]
D --> H[Configure per-state\ndisclosure behaviour]
E --> I[Build human-review\ncheckpoint]
F --> J[Add provenance metadata\nat output layer]
G --> K[Restrict engagement\npatterns for minors]
H --> L[Log agent behaviour\nby state]
I --> L
J --> L
K --> L
L --> M[Audit-ready records]
What the 2027 outlook adds to your planning horizon
The Transparency Coalition report identifies the most likely directions for 2027: product liability for AI, stronger chatbot safeguards, certification and licensing of third-party auditors, procurement standards for AI vendors, restrictions extended to more licensed professions, and testing and monitoring requirements for advanced models.
For teams building on frameworks like LangGraph or CrewAI, the testing and monitoring direction is worth flagging now. If monitoring requirements for advanced models become statutory in additional states, teams that have already instrumented their agents for observability will have a shorter path to compliance than those who have not.
Third-party auditing is also moving from frontier-model-only to a broader category. California’s work on auditor infrastructure and the Illinois audit requirement together suggest that independent review will extend down-market over the next two to three years.
How to think about what this means operationally
The obligations that emerged in 2026 share a structural pattern: they land on behaviour in production. What the agent says, whether a human reviewed a consequential decision, and whether a record exists of how the agent actually behaved. A legal review of your terms of service does not satisfy a disclosure requirement. A policy document does not satisfy an audit requirement.
Organisations like KPMG, which deployed AI agents to 276,000 employees across 138 countries in May 2026, and PwC, which deployed AI agents across finance, tax, and compliance workflows for 30,000 US professionals, operate in exactly the regulated categories this report covers. Both organisations had to build governance infrastructure alongside the agent deployment itself.
For teams earlier in that journey, the AI governance framework you build needs to include per-state configuration, interaction logging, and human-review routing as first-class components, not retrofits. Tools in the runtime governance category, including platforms like Prefactor, can help you manage state-level rule sets at the agent layer rather than hard-coding them into application logic.
The governance index at /governance-index/ collects the relevant frameworks and guidance in one place if you want a structured starting point.
This article is not legal advice. The Transparency Coalition report is the primary source; verify current status of specific measures with qualified legal counsel before making compliance decisions.
Where to start
If your agents are in production or approaching it, the most useful next step is to map which state categories apply to your deployment, then assess whether your current logging and human-review infrastructure would satisfy an audit. Take the agent readiness assessment to get a structured view of where your governance gaps are before the 2027 legislative cycle adds further obligations.