On July 1, 2026, Delaware became the first state to write into its medical practice code that an AI agent cannot be a doctor, a nurse, or a physician assistant, no matter what it is capable of doing. House Bill 191, signed April 23 and passed 38 to 0 in the House and 20 to 0 in the Senate, bars any nonhuman entity from being licensed or credentialed as a physician, physician assistant, or nurse in any of Delaware's four nursing tiers, and bars that entity from using any of those titles. Six months earlier, on January 1, 2026, the federal government's own Interoperability and Prior Authorization Final Rule had already reached full effect across Medicare Advantage, Medicaid managed care, and exchange plans, requiring that a clinician review any prior authorization denial produced through an automated process before it becomes final. Two governments, working independently, answered the same question the same way inside a single calendar year. Neither answered the harder one: once an AI agent is barred from the license and barred from the final call, who inside the health system owns the agent itself.
The Sector Signal
HB 191 amends Title 24 of the Delaware Code to state plainly that a nonhuman entity, including an agent powered by artificial intelligence, may not be licensed or certified to practice as a professional nurse, an advanced practice registered nurse, a practical nurse, a physician, or a physician assistant, and may not use any of those titles. Delaware was not first. California's AB 489, in effect since January 1, 2026, already lets the state enforce title-protection law against any AI system or developer that implies a license the system does not hold. Washington's HB 2155 restricts the titles RN, ARNP, and LPN to human persons. Oregon's HB 2748 does the same for APRN, CRNA, CNS, LPN, RN, NP, CMA, and CNA. Four states, inside seven months, closed the same door from four different statehouses.
CMS closed a second door, not the same one. Under its Interoperability and Prior Authorization Final Rule, when a payer denies an item or service through an automated process, a clinician has to review the determination before it stands. CMS paired the rule with the WISeR model, a pilot that began January 1, 2026 in a set of states, testing AI-assisted review against seventeen outpatient services already flagged for overuse or fraud risk, with a human clinician retained in the loop by design. The federal government did not ban AI from prior authorization. It banned AI from being the last signature on the file.
GOVERNING EVIDENCE
Delaware's HB 191, signed April 23, 2026 and effective July 1, 2026, bars nonhuman entities from holding a physician, physician assistant, or nursing license or title, clearing the House 38 to 0 and the Senate 20 to 0. California's AB 489, Washington's HB 2155, and Oregon's HB 2748 impose comparable title bars on AI systems and nursing credentials, all in effect during 2026. CMS's Interoperability and Prior Authorization Final Rule, fully effective January 1, 2026, requires clinician review of any algorithmic denial across Medicare Advantage, Medicaid managed care, and exchange plans. The WISeR model, launched the same day, pilots AI-assisted review with mandatory human clinical oversight across seventeen high-risk outpatient services.
What Health Systems Already Know
The Delaware Healthcare Association did not oppose HB 191. It supported the bill in committee, and its membership includes ChristianaCare and Nemours Children's Health, two of the state's largest health systems. That is a separate signal from the legislative one: the organizations that would have to build compliance around a licensing bar asked their legislature for the bar themselves. A bright line on what an agent cannot be is easier to build against than an unresolved question of what an agent may do unsupervised.
The harder half of the problem shows up nowhere in statute. Black Book Research's 2026 Health System and Hospital AI Governance Resource Guide, released November 2025, found that AI pilots built with a named owner and a tested kill switch are twice as likely to scale system-wide within a year, and that health systems operating a standing AI Governance Council are twice as likely to reach measurable ROI within twelve months. The systems pulling ahead are not the ones with the cleanest compliance file against the new title laws. They already know, in writing, who can pull the plug.
The Governance Implication
This is the Governance Boundary Principle applied to a hospital board rather than a corporate one: the board's job is to own oversight of the AI Governance Council's mandate, not to manage the clinical workflows an agent touches. A quality committee that reviews an AI vendor contract once a year has satisfied the license question CMS and Delaware just answered. It has not touched the ownership question those same regulators left alone, because a licensing bar and a clinician sign-off requirement were never designed to answer it. That gap is also the Accountability Contract Model at the bedside: a named executive stating what a clinical agent is authorized to do, what authority has been granted to override it, what evidence proves it stayed inside that authority, and who has standing to shut it down before the next accreditation survey asks. A licensing bar tells the agent what it cannot be. It says nothing about who answers for what the agent is still permitted to do every day it operates inside a HIPAA-covered workflow.
Three Proportionate Responses
First, every health system running an agent inside a prior authorization, triage, or clinical documentation workflow should confirm that a single named executive, not a committee, holds documented authority to pause or decommission that system, filed with the same rigor as a credentialing record.
Second, systems in a state without a title-protection law should not wait for one. Delaware, California, Washington, and Oregon did not act because a system failed. They acted while the ownership question was still open. Building the AI Governance Council now costs less than building it after a legislature or a CMS enforcement action forces the question on a shorter timeline.
Third, boards should ask their quality committee one question at the next cycle: if an agent's action inside a clinical workflow were reviewed by a state licensing board tomorrow, whose name is on the file as the accountable owner. If the honest answer is a committee rather than a person, the governance architecture is not finished.
Delaware answered what an AI agent cannot be. CMS answered who has to sign the file. Neither regulator was ever going to answer who owns the system in between, because that was never a question a regulator could answer for a board. It is the one board members have to answer for themselves, before an examiner, a plaintiff's attorney, or a state licensing board asks it for them.