Connecticut will require human review before artificial intelligence can drive adverse health-plan decisions for more than 270,000 public employees, retirees and participating municipal workers. The safeguards, announced by Comptroller Sean Scanlon on September 16, apply to the State Employee Health Plan and Connecticut Partnership Plan and take effect January 1, 2027, according to the state policy and an independent CT Mirror report.
The central rule is straightforward: no adverse determination may be made solely by an AI system. A person must review it. Insurers also may not use AI or predictive models alone to downcode a claim, reduce a provider payment or change a billing code. The policy does not eliminate automation from routine claims administration; it preserves authorized contractual and automated processes while putting a human checkpoint around decisions that can reduce care or payment.
That distinction matters because the state is regulating its own purchasing relationships rather than banning a technology across Connecticut’s insurance market. The public plans use several administrators: Anthem handles medical benefits for active workers, Cigna administers dental coverage, Aetna manages the retiree Medicare Advantage plan and Caremark is the pharmacy-benefit manager, reporting shows. Scanlon said the administrators agreed to the new requirements after negotiations.
Privacy, disclosure and audits
The policy also bars member data from being used to train, develop or support other AI models. Carriers and providers must tell members when AI materially assists with, directly participates in or recommends a benefit or health service. AI systems must be tested against historical data for accuracy, consistency and fairness, and carriers must disclose their governance and audit procedures to the comptroller. Public-radio coverage from WSHU independently confirmed the human-review, privacy, disclosure and audit requirements.
For patients, the practical effect is a right to human involvement when an algorithm contributes to an unfavorable coverage decision under these state plans. For clinicians, the same principle applies when automated tools would reduce payment or revise coding. The policy does not guarantee that a human reviewer will reverse an initial recommendation, but it prevents an AI output from being the final word. It also creates an oversight trail by requiring administrators to document how their systems are governed and checked.
The reach is substantial but not universal. The rules cover the two plans overseen by the comptroller, not every Connecticut resident. Scanlon said he plans to ask lawmakers in the 2027 session to extend comparable protections to all state-regulated health plans. Roughly 220,000 people obtain coverage through that market, while many large-employer self-funded plans remain under federal rather than state authority, according to CT Mirror.
The Connecticut State Medical Society supported the move, arguing that automated denials and downcoding should not replace clinical judgment. A separate industry report by Becker’s confirmed that the policy covers more than 270,000 members and retains authorized automated claims protocols. The policy therefore functions as a near-term purchasing rule and as a template for a broader legislative debate next year.
