In September 2026, the New Jersey Legislature’s published text of Senate Bill 1029 placed human medical accountability at the centre of a proposed framework for health claims and automated utilisation management.

The bill would require every claim submitted to a payer to be reviewed by at least one physician or medical director employed by, or contracted with, the payer. It would also require annual public reporting on approvals, denials, appeal outcomes and review times, together with disclosure of whether an automated utilisation management system is used and how many claims it reviews. The Department of Banking and Insurance would receive authority to audit the system and the data generated through it.

At this stage, the proposal remains legislation rather than an enacted rule, and its text may change. Even so, it illustrates how political scrutiny of artificial intelligence in healthcare is moving from broad principles toward operational controls. The central question now becomes, who remains accountable when automated recommendations affect access to care.

Human Review Must Be More Than a Formal Signature

A requirement for medical-director review can protect members only if the reviewer has enough time, information and authority to change the result. A process in which a clinician merely confirms a model output would satisfy the appearance of oversight without delivering a genuine independent assessment. Effective review requires access to the clinical record, the policy wording, the criteria applied by the system and an explanation of the factors that drove the recommendation.

The bill addresses this concern indirectly through transparency. Denial notices would include the reviewer’s denial rate and average review time, as well as instructions for accessing state consumer assistance. Payers would publish annual rejection statistics, including decisions made by physicians or medical directors. These measures could reveal unusually fast reviews or persistent denial patterns, although raw rates would still need adjustment for case mix and product design.

The proposed definition of automated utilisation management is broad. It covers systems that recommend or determine whether a service should be reimbursed or provided, including prior authorisation, continued-stay review, discharge planning and retrospective review. That breadth reflects the reality that consequential automation often appears inside ordinary workflow software rather than in a system marketed explicitly as AI.

Employers Need Their Own Assurance Model

Employers purchasing insured health coverage should not wait for regulators to define every control. Requests for proposal and renewal reviews can ask carriers to describe where automation is used, which decisions require clinical confirmation, how models are validated and how adverse outcomes are monitored. The employer should also request denial and appeal data relevant to its population, while respecting privacy rules.

For self-funded plans, the governance challenge is more direct. ERISA pre-emption may limit the reach of some state insurance rules, but it does not remove the sponsor’s duty to monitor administrators. Contractual standards should specify that automated tools cannot issue final adverse determinations without appropriate human review. Audit rights should cover algorithms, clinical criteria, data quality, overrides, turnaround times and member communications.

Appeal design is another important control. Members should be able to reach a qualified reviewer quickly, submit additional clinical evidence and receive an explanation that addresses the facts of their case. Employers should monitor reversal rates and repeated reasons for appeal. A high reversal rate may indicate that initial automation is too aggressive, that the underlying data are incomplete or that clinical rules are not being applied consistently.

Multinational benefits teams can apply the same principle across countries: the more consequential the decision, the clearer the human accountability must be. A useful control framework distinguishes administrative automation, clinical recommendation and final benefit determination. It then assigns an accountable person, evidence requirement and escalation route to each level.

New Jersey’s proposal may evolve before adoption, but it offers a practical benchmark. Human oversight should be observable, documented and capable of changing the outcome. Without those features, a human in the process may provide little more than legal decoration.