Autonomous AI agents in healthcare: the Q4 2026 regulatory frontier

Three weeks in October 2026 did more to shape the regulatory future of autonomous AI in healthcare than the previous two years of agency activity combined. The FDA cleared the first AI-enabled device that generates a diagnostic recommendation without a required clinician concurrence step. HHS opened a Request for Information on autonomous agents in clinical workflows that reads like the preamble to a rulemaking docket. ONC published an HTI-2 enforcement framework that creates the first federal registry of in-use clinical decision support algorithms. These are not three independent events. They are the three legs of a federal posture on autonomous AI that will, if it holds, define the vendor economics of the category for the rest of the decade.

The posture has three loosely-coupled pillars. The FDA pillar governs devices — AI-enabled products that produce a diagnostic or therapeutic output that a clinician would otherwise produce. The HHS pillar will govern autonomous agents more broadly — the triage bots, outreach systems, prior-auth submission agents, and agentic workflow orchestrators that increasingly sit outside the FDA’s device jurisdiction because they do not themselves make a diagnosis. The ONC pillar governs the data layer — what algorithms are deployed, where, with what inputs, and with what disclosures to the clinician and the patient. Together these three pillars will constitute the operating framework that vendors and health systems will spend 2027 and 2028 learning to live inside.

The autonomy carve-out as a template

The FDA’s de novo authorization for SKOUT Autonomous is the most structurally interesting of the three because it establishes a template rather than a one-off decision. The template has four elements: narrow indication, defined threshold, mandatory registry, named stopping rules. Each of those elements is likely to show up in every subsequent autonomous-mode AI authorization.

The narrow indication is the operational core of the template. An autonomy carve-out is not a general permission for the AI to operate without clinician oversight; it is a permission for the AI to operate without real-time concurrence for a specifically-defined finding category, in a specifically-defined setting, when operating within specifically-defined input conditions. Vendors that want to pursue autonomy clearances in cardiology, ophthalmology, dermatology, or any other specialty will have to specify their carve-outs with the same level of granularity.

The defined threshold is the operational boundary. SKOUT Autonomous can characterize polyps below a size-and-morphology threshold; above the threshold, the clinician concurrence step reactivates. Every autonomy carve-out will need a threshold specification of this kind, and the vendor’s responsibility is to make the threshold mechanically enforceable — not a clinician judgment call, but a product-controlled gating condition.

The mandatory registry is the post-market condition. Every center using autonomous mode must contribute per-case data to a registry that the vendor cannot unilaterally control. This is a significant operational cost — registry infrastructure, data stewardship, independent oversight committee, patient-level consent considerations — and vendors should expect the registry cost to be a material component of the business case for pursuing an autonomy clearance.

The named stopping rules are the pre-market commitment that the authorization will be withdrawn if real-world performance deviates from pre-market validation beyond specified bounds. This is new. Historically, FDA post-market surveillance has been a reactive mechanism. Named stopping rules make it a pre-committed mechanism, with the vendor and the agency aligned in advance on what performance drift will trigger withdrawal of the autonomy indication.

What the HHS RFI tells us about the coming rule

The HHS RFI is deliberately open-ended, but the structure of its question clusters tells the industry what the eventual rule will look like. Four areas will almost certainly end up in rulemaking.

Scope-of-practice interactions will be the most contested. State professional licensure boards have historically guarded the scope of practice for licensed professions — nursing, pharmacy, social work, medicine. An autonomous agent performing clinical-adjacent work (triage, outreach, medication reconciliation) is doing work that in a human workforce would be assigned to a specific profession. The question the rule will have to answer is whether the agent operates under state licensure (configured to the scope of the deploying clinician) or under a federal preemption (operating within a federally-defined scope regardless of state law). Vendors need to track this closely. A federal preemption would consolidate the operating model nationally and reduce the configuration burden on vendors. A state-licensure model would preserve state variation and raise the compliance cost of national deployment.

Documentation accountability is the question that will shape what agent-authored content looks like inside the EHR. The current field is divided: some vendors insert agent-authored notes flagged as AI; others treat agent output as a draft requiring clinician sign-off; others log agent actions in a parallel audit trail. The eventual rule will likely require a specific approach — the leading candidates are a prominent AI-authorship flag in the clinical record plus a parallel audit trail with per-decision provenance.

The human-in-the-loop trigger cluster is where the rule will most directly constrain vendor product design. The agency will specify conditions under which an autonomous workflow must escalate to human review — clinical escalation signals, scope-boundary breaches, model-uncertainty thresholds, patient preference signals. Vendors should expect the final specification to be stricter than their current product behavior, and should begin the engineering work now to add the escalation hooks and configuration surface.

Liability allocation is the question with the largest financial consequence. The RFI explicitly asks whether existing frameworks are adequate or whether a new allocation (across the vendor, the deploying institution, and the supervising clinician) is needed. A shared-liability model will raise vendor insurance costs and change the economics of indemnification clauses in vendor-health-system contracts. A clinician-primary model will push health systems to tighter supervisory requirements, which will slow adoption.

The data layer under HTI-2

The ONC HTI-2 enforcement framework will interact with the FDA and HHS pillars in ways that are not yet fully understood by the field. The core mechanism is that every DSI deployed inside a certified EHR must have published source-attribute metadata that includes the inputs, the output, the intended use, and the uncertainty characteristics. The quarterly attestation cadence and the random audit mechanism mean that this registry will be substantially more detailed and current than any public inventory of clinical AI that has existed to date.

For autonomous AI specifically, the HTI-2 registry will create an unavoidable public artifact. If an autonomous agent is deployed inside a certified EHR environment and makes a decision that influences care, it almost certainly meets the DSI definition and will have to be registered. That means autonomy decisions that vendors and health systems have made quietly become public knowledge — competitors, researchers, journalists, and plaintiffs’ attorneys will all have access to the inventory of autonomous AI in use.

The interaction effect with the FDA pillar is that the FDA’s narrow autonomy carve-outs will be visible in the HTI-2 registry. Deviations from the carved-out indication — autonomous mode used in a setting the FDA did not authorize, or on a patient population outside the authorized indication — will be inspectable from the registry entry. Health systems that treat the HTI-2 attestation as a documentation exercise rather than an operational truth will find themselves exposed to enforcement from both ONC and the FDA.

The interaction effect with the HHS pillar is that the eventual autonomous-agent rule will almost certainly reference the HTI-2 registry as the official inventory of autonomous systems in use. The agency will not need to build parallel reporting infrastructure; it can leverage ONC’s registry as the authoritative source of what is deployed where.

What vendors should do in Q4 2026

Four actions are now time-sensitive for vendors building in this space.

First, build the registry-ready evidence package now. Whatever autonomy carve-out a vendor eventually pursues, the per-case data capture, independent registry infrastructure, and pre-committed stopping rules will take 9-12 months to stand up. Vendors that start now will have the infrastructure ready when the regulatory path becomes clear; vendors that wait will be 9-12 months behind competitors.

Second, respond substantively to the HHS RFI. The rule will be written using the vocabulary that the RFI responders provide. Vendors that submit substantive comments — real deployment data, specific proposed definitions, named failure modes — will shape the final rule. Vendors that stay silent will be regulated by the frameworks their more-engaged competitors propose.

Third, build the HTI-2 metadata layer into product now rather than retrofitting it. The DSI source-attribute metadata will be auditable inside live customer deployments beginning in 2027. Vendors whose products emit the metadata natively will absorb the enforcement workload inside their normal release cycle. Vendors that treat HTI-2 as a documentation project will face expensive remediation inside customer installations.

Fourth, re-examine indemnification and liability allocation in current customer contracts. The eventual HHS liability framework will not be retroactive, but customer procurement teams will begin asking for forward-looking liability clauses the moment the RFI responses are public. Vendors with coherent, defensible positions will negotiate from strength; vendors that have not thought through the question will be forced to accept the terms the health-system procurement team proposes.

What health systems should do

For health systems, the near-term priorities are different but equally time-sensitive.

Build the AI inventory. The HTI-2 enforcement framework will require this inventory to be published and auditable in 2027. Health systems that have been operating with a loose or implicit inventory need to formalize it now.

Build the AI governance capability. The HIMSS 2026 benchmark data showing governance maturity lagging execution maturity by roughly 18 months is a precise warning. The federal rulemaking cadence is going to close that gap on its own schedule; health systems that lead rather than follow will be in materially better regulatory posture.

Audit autonomous AI exposure. Any vendor product that performs an action without clinician concurrence — triage, outreach, prior-auth submission, documentation finalization — is going to come under the HHS rule when it finalizes. Health systems should inventory those deployments now and begin preparing the escalation configurations, documentation flows, and clinician supervisory structures that the rule will require.

Outlook

The next 12 months will be formative. The RFI comment period closes in January 2027; the proposed rule will likely follow in the second half of 2027; the final rule will likely take effect in 2028. HTI-2 enforcement begins in January 2027. The FDA will likely clear two or three additional autonomy carve-outs during 2027. By the end of 2027, the operating framework for autonomous AI in healthcare will be substantially more defined than it is today, and the vendor-and-health-system commitments made during this transition period will shape the competitive landscape for the rest of the decade.