Can AI deny health insurance claims? 8 states now say not alone

AI health insurance denials are now regulated in eight states. Alabama, Colorado, Georgia, Illinois, Indiana, Iowa, Utah, and Washington each enacted laws in 2026 requiring that a human be involved when an insurer makes an adverse authorisation or coverage decision. An AI agent alone cannot be the final decision-maker in those states.
What the eight state laws actually require
The laws share a common principle: adverse medical authorisation and coverage decisions cannot rest solely on AI or automated systems. Beyond that, the specifics vary. Some states define who must conduct the human review. Others specify what must be disclosed to the patient. Some require that the decision account for the individual patient’s clinical history. The Transparency Coalition’s 2026 State AI Legislation Report summarises the landscape but does not reproduce each bill’s operative text. You must read the individual legislation for each state where your agent operates. This article does not constitute legal advice.
Separately, Colorado, Maine, Rhode Island, Tennessee, and Vermont moved in 2026 to restrict AI chatbots offering mental health therapy, using existing licensing authority. The Transparency Coalition expects 2027 to extend similar restrictions to other licensed professions, which matters for teams building agents that touch clinical recommendations beyond claims.
Why ‘not solely AI’ is harder to satisfy than it sounds
The phrase sounds like a checkbox. It is not. A human reviewer who approves agent output at or near 100% of the time, without documented consideration of the specific case, is not meaningfully reviewing anything. Courts and regulators evaluating these laws are going to look at the substance of the review, not the presence of a signature field.
This is already a measurable problem in adjacent workflows. Optum’s Digital Auth Complete platform, powered by Humata Health, reports a 96% first-pass approval rate and a 45% reduction in manual touches. Those numbers reflect efficiency gains on the approval side, which is fine. The risk emerges on the adverse side: if the volume of denials passed to human reviewers is small and the reviewers are measured on throughput, rubber-stamping becomes the path of least resistance.
Cohere Health’s Unify platform processes around 12 million prior authorisation requests annually with an 88 to 90% auto-approval rate. Again, the efficiency case is clear. But the 10 to 12% that do not auto-approve are exactly where the new state laws apply, and where the human-review step must be genuine.
Between 2024 and 2025, the share of insurers with full-scale AI adoption grew from 8% to 34%. That pace of adoption means compliance infrastructure is being built after the agent is already in production in many organisations. If you are in that position, the design changes below are the starting point.
How to build the human-review step so it holds up
The human review must be inside the product flow, not in a support manual. That means the agent’s adverse recommendation cannot go to an email queue or a PDF report. It must route to a reviewer interface that captures the decision explicitly.
flowchart TD
A[Agent evaluates claim] --> B{Adverse outcome?}
B -- No --> C[Auto-approve, log decision]
B -- Yes --> D[Route to human reviewer queue]
D --> E[Reviewer sees agent recommendation + evidence]
E --> F{Reviewer decision}
F -- Approve --> G[Log: reviewer approved, overrode agent]
F -- Deny --> H[Log: reviewer confirmed denial + rationale]
F -- Modify --> I[Log: reviewer modified, record change]
G & H & I --> J[Per-state disclosure step]
J --> K[Decision delivered to patient/provider]
Three things must be recorded at the point of review: what the agent recommended, what evidence the agent used to reach that recommendation, and what the reviewer decided, including any departure from the agent’s output. If the reviewer’s decision always matches the agent’s recommendation, that pattern is itself a signal worth monitoring.
flowchart TD
A[Adverse decision logged] --> B[Calculate reviewer agreement rate]
B --> C{Agreement rate > 95%?}
C -- No --> D[Normal operations]
C -- Yes --> E[Flag for audit]
E --> F[Review sample of decisions]
F --> G{Genuine consideration?}
G -- Yes --> D
G -- No --> H[Escalate: human step is nominal]
For teams operating across multiple states, the approval path itself must be per-state. What satisfies Washington’s disclosure requirement may differ from what Georgia requires. A single workflow with a state parameter that controls which steps run is preferable to separate deployments, but the per-state logic must be explicit and auditable. For more on how to structure this, see the guidance on AI agent governance and the broader AI governance framework.
The mental health chatbot restriction is a separate track
The five states restricting AI mental health therapy chatbots are using a different legal mechanism: professional licensing authority rather than insurance regulation. This matters for product architecture because the compliance question is not just whether the AI makes a final decision, but whether the AI is practising a licensed profession at all. If you are building agents that provide clinical guidance, triage, or therapeutic conversation, you are in scope regardless of whether a human reviews the output. Check your AI governance in healthcare posture against licensing rules in each state of operation, not just insurance statutes.
What to log and why it matters
Logging is not optional in regulated decision workflows. The record needs to support three use cases: internal audit, regulatory enquiry, and patient or provider appeal. That means storing the agent’s input data, the model version that produced the recommendation, the confidence or scoring output if one exists, the reviewer’s identity and decision timestamp, and any changes the reviewer made.
72% of US adults say it is extremely or very important that their provider tells them when AI is used in their care. Disclosure requirements in the new state laws are likely to require surfacing at least some of this log to patients. Build the log first; the disclosure layer can draw from it.
Tools in this space, including Prefactor, are designed to capture agent reasoning and reviewer decisions in a structured audit trail. Whatever tooling you use, the audit trail must be queryable by case ID, by state, and by reviewer, or it will not serve its purpose when regulators ask for it. For a broader view of what observability needs to cover, see implement agent observability and runtime governance vs pre-deployment review.
Insurer AI adoption is moving faster than compliance infrastructure in most organisations. The 2026 laws are the first wave; the Transparency Coalition’s report signals that 2027 will add more states and extend restrictions to additional professions. Retrofitting human-review workflows into a production agent is significantly more expensive than building them in from the start. For the underlying design principles that apply before you get to compliance, see AI agents workflow and govern agentic AI.
Where to start
Map every state where your claims, prior-authorisation, or utilisation-review agents operate against the eight affected states, then read the individual bills rather than summaries. From there, audit whether your current human-review step meets a genuine-consideration standard or only a procedural one. Take the agent readiness assessment to identify the gaps in your current design before they become compliance findings.