Healthcare AI evaluation imagery representing clinician review of medical AI evidence.

Healthcare · · 8 min read

What the PATIENT AI Act means for health AI data

Acumen Editorial · Updated November 22, 2026

Most health AI policy news is about who regulates what. This draft is also about evidence: how often you produce it, who reviews it and what happens after launch. For heads of clinical AI, regulatory leads at health AI companies and evaluation teams, that makes it a data question as much as a legal one.

It is a draft, not law. It may change or stall. But it shows where federal thinking is heading, and the comment window is open now.

What is the PATIENT AI Act?

The Protecting Americans Through Intelligent, Ethical, and Transparent Artificial Intelligence in Health Care Act is a discussion draft from the chair of the Senate HELP Committee. Per Akin Gump, it was released October 29, 2026, and comments go to [email protected] by December 1, 2026.

Law firm summaries from Hogan Lovells and Akin Gump group the draft into four areas:

  • HHS activities. HIPAA guidance for AI use, internal use cases, workforce training, a public process on AI liability and datasets for AI-based health research.
  • FDA review. New flexibility for software changes, third-party review for lower-risk software and a postmarket surveillance pilot.
  • Health IT. An expanded ONC "Certified Health IT" definition that would reach AI applications that are not FDA-regulated devices.
  • Medicare and Medicaid. A coverage evaluation framework, payment methodologies, HCPCS Level II codes for AI-based services and a CMS Innovation Center payment model for AI that delivers care autonomously.

Liability gets lighter treatment. Becker's reports the draft would apply existing federal preemption of state product liability laws to integrated and software-based devices, and Akin Gump notes HHS would work with a nonprofit to hold a public meeting on AI liability and publish recommendations. Hogan Lovells observes that the draft does not set out a comprehensive framework for allocating liability when AI contributes to patient harm. Until it does, your evidence file is your best defense.

What does the draft change at FDA?

The FDA title has four parts that matter for evidence. The FDA Law Blog walks through each:

SectionWhat it would doWhat it means for your data
201 — Component reviewVoluntary preliminary review of software components of integrated devices. The full device still needs a 510(k), De Novo or PMA.Component-level test sets need clean documentation so a later submission can rely on them.
202 — Risk-based review and PCCPsSponsors could submit predetermined change control plans for changes made while a 510(k) or PMA is pending.Every in-review change needs a rerun on a frozen, versioned evaluation set.
203 — Third-party reviewAccredited third parties could review low- to moderate-risk software devices, including De Novo requests. FDA would have 30 days to accept or reject a recommendation.Outside reviewers will read your clinical data cold. Provenance and grading rules must stand on their own.
204 — Postmarket pilotA voluntary pilot to modernize postmarket surveillance for integrated and software devices, including real-world data and CMS coordination.Postmarket monitoring needs a sampling plan and clinician adjudication, not just dashboards.

The same authors are skeptical. They argue that "Title II risks adding procedures rather than improving the path to market," and that the draft does not change manufacturers' core obligations. They also flag that PCCP flexibility during review could create version-control confusion without clear timing rules.

That critique is the practical point. If change control gets more flexible, the burden moves to the sponsor's ability to show exactly which model version passed which test.

Does the draft create public datasets for health AI?

Yes, in part. The HIPAA Journal reports the draft instructs the HHS Chief Information Officer to "build datasets to be used to support health care research" using AI-based applications. Mintz describes the provision as standardized datasets plus guidance on their curation, validation and use.

Public datasets help research and comparison. They will not replace intended-use-specific evaluation data. A shared federal set cannot know your patient population, your specialty mix, your workflow or your failure modes. Teams should expect regulators and reviewers to ask how their own test data was curated and validated, using whatever standard HHS writes as a reference point.

Why does the autonomous AI payment model matter?

Because payment raises the evidence bar. Becker's and Hogan Lovells report that the draft would have the CMS Innovation Center test a payment model for AI applications that deliver care autonomously, alongside a broader coverage evaluation framework.

Payers do not pay for benchmark scores. They pay for documented performance in a defined use, with a clear account of when the system defers to a clinician. Any team aiming at autonomous or near-autonomous functions should assume it will need specialist-graded evidence on real case mixes, plus proof of safe escalation.

How does this fit with FDA's generative AI discussion paper?

It points the same way. Mintz reports that FDA's discussion paper on generative AI-enabled devices drew about 540 comments before its October 19, 2026 close. Medical professionals asked for clinical confirmation scaled to risk, prospective trials for high-risk or autonomous functions and postmarket monitoring that supplements premarket evidence rather than replacing it. Developers asked for intended-use-specific performance metrics.

Put the two together and a pattern appears. Evidence is becoming continuous, intended-use-specific and reviewable by people outside your team.

What should health AI teams do now?

Use this checklist whether or not the draft becomes law.

  • Freeze and version your evaluation set. Tie every score to a dataset version, a model version and a date.
  • Write grading rules a stranger can apply. Third-party reviewers need rubrics, error taxonomies and severity scales that work without your team in the room.
  • Use specialists who match the intended use. A general reviewer cannot grade a cardiology or oncology output with the depth a reviewer will expect.
  • Measure agreement. Double-grade a share of items and report inter-rater agreement and how disagreements were adjudicated.
  • Document provenance. Record who wrote and graded each item, their credentials and how they were verified.
  • Plan reruns for every change. Model updates, prompt changes and retrieval changes should all trigger the same frozen test.
  • Build a postmarket sampling plan. Decide now how many real-world outputs clinicians will review each month and what triggers escalation.
  • Test deferral, not just accuracy. For autonomous functions, show when the system hands off to a clinician and whether it does so safely.
  • File a comment. If your organization has a view on PCCPs, third-party review or dataset standards, the deadline is December 1.

How Acumen helps

We build the evaluation data this draft assumes you already have. Our verified clinicians and specialists write test items for your intended use, grade outputs against documented rubrics and adjudicate disagreements. Every dataset ships versioned, with grader credentials and agreement statistics, so you can rerun it on each change and hand it to an outside reviewer.

See how it works on our medical model evaluation page, or read what FDA competency-based evaluation means for generative AI devices.

Request a sample dataset Schedule a conversation

FAQ

What is the PATIENT AI Act?

The PATIENT AI Act is a Senate discussion draft from Sen. Bill Cassidy, chair of the HELP Committee, released October 29, 2026. It covers HHS guidance and datasets, FDA software review, ONC health IT certification and Medicare payment for AI-based services. It is not introduced legislation, and stakeholder comments are open until December 1, 2026.

When are comments on the PATIENT AI Act due?

Comments are due December 1, 2026, and go by email to [email protected], according to Akin Gump and the College of American Pathologists. Because it is a discussion draft, the text may change based on feedback. Health AI developers, health systems and data providers with views on review or datasets should comment before that date.

How would the PATIENT AI Act change FDA review of AI software?

According to the FDA Law Blog, the draft would allow voluntary preliminary review of software components, let sponsors submit predetermined change control plans for changes made during a pending 510(k) or PMA, expand accredited third-party review to low- to moderate-risk software and create a voluntary postmarket surveillance pilot using real-world data.

Does the PATIENT AI Act create datasets for AI research?

Yes. The HIPAA Journal reports that the draft instructs the HHS Chief Information Officer to build datasets to support health care research using AI-based applications. Mintz describes the provision as standardized datasets with guidance on curation, validation and use. These would support research, not replace intended-use-specific evaluation data.

Would Medicare pay for autonomous AI under the draft?

The draft would have the CMS Innovation Center test a payment model for AI applications that deliver care autonomously, per Becker's and Hogan Lovells. It would also create a coverage evaluation framework and a process for HCPCS Level II codes. Any payment would depend on the final law and on evidence of performance in a defined use.

What evaluation data should health AI teams prepare?

Prepare a frozen, versioned test set built and graded by specialists who match your intended use. Document grading rules, grader credentials and inter-rater agreement. Rerun the same set on every model, prompt or retrieval change, and plan clinician review of a sample of real-world outputs after launch. That evidence holds up whatever the final law says.

From the Acumen archive. Explore expert data and Healthcare & Pharma.

← Back to Insights