HTI-5 Final Rule Ends AI Model Card Rules: What Practices Should Do
Executive Summary
- The HTI-5 final rule (91 FR 67916, published October 28, 2026) removes the decision support transparency requirements often called AI “model cards” from ONC certification on December 28, 2026. Certified EHRs will no longer have to support the 31 predictive source attributes HTI-1 specified, and developers will no longer have to apply the related intervention risk management practices. The rule does not require vendors to delete information they already hold.
- HTI-5 changes what EHR developers must certify. It does not amend Section 1557 and creates no new provider deadline. ONC used the final rule itself to remind providers of 45 CFR 92.210, the Section 1557 duty to identify and mitigate discrimination risk in patient care decision support tools, and HHS confirmed in June 2026 that it continues to enforce the protections unaffected by the gender identity vacatur.
- Whether that information stays available will depend on the product, other law and your contract. ONC’s own answer to commenters who objected was to encourage users to ask their developers for it, and California’s SB 503 separately requires clinical AI developers to provide related documentation for systems with a known or foreseeable risk of biased impacts starting January 1, 2027.
- The CMS Clinical Super Intelligence Roadmap released November 13, 2026 is a nonbinding eight page policy statement. It signals national standards and possible FDA actions that could trigger existing statutory preemption for certain devices, but it creates no practice level mandate or deadline.
- Our recommendation, not a legal deadline: before December 28, inventory the AI tools in your EHR, ask each vendor in writing for the source attribute information it currently maintains, and document a 92.210 review for each tool used in clinical decisions. A vendor request supports that work but is not, by itself, proof of compliance.
In this article
- What does the HTI-5 final rule change for AI in your EHR?
- What exactly were the AI model cards?
- When does the HTI-5 final rule take effect?
- Why ONC removed model cards over most commenters’ objections
- What does HTI-5 mean for medical practices that use EHR AI tools?
- Section 1557 and 45 CFR 92.210: the duty HTI-5 did not touch
- What to ask your EHR vendor before December 28
- Where the CMS clinical AI roadmap fits
- What HTI-5 left unresolved
- A practical 40 day plan for a small practice
1. What does the HTI-5 final rule change for AI in your EHR?
The HTI-5 final rule removes the federal certification requirement that EHR software support and give users access to standardized source attribute information, often called an AI “model card,” for the decision support tools a developer supplies, and it removes the developer’s duty to apply intervention risk management to predictive tools. The Office of the National Coordinator for Health Information Technology (ONC) published the rule in the Federal Register on October 28, 2026 as 91 FR 67916, and the change to the decision support interventions criterion takes effect December 28, 2026.
For a five physician family practice or a two location specialty clinic, the practical effect is simple. For developer supplied interventions in the certified module, HTI-1 required a standardized set of source attribute fields and user access to plain language descriptions of them, although certain predictive fields such as external validation and local performance could be marked unavailable. After December 28, certification will no longer require those capabilities. Whether the information stays available will depend on the product, other applicable law and your contract. HTI-5 also does not show that every vendor previously held complete validation and fairness evidence.
The HTI-5 final rule is an ONC regulation that narrows the federal Health IT Certification Program. For AI, it removes the source attribute (“model card” is shorthand) and intervention risk management requirements for decision support interventions in certified EHRs, effective December 28, 2026. It does not remove a medical practice’s own duties under Section 1557, HIPAA or state law.
The rule is a deregulatory package with a much wider scope than AI. ONC proposed on December 29, 2025 to remove 34 of the 60 certification criteria then in effect and revise seven. The final rule removed 14, revised seven and left 39 unchanged, according to ONC’s own table of removals and revisions. ONC estimates a present value of $1.55 billion in cost savings from the whole final rule, measured in 2024 dollars at a 7 percent discount rate beginning in 2027 and in perpetuity. That figure covers every change in the rule, not the AI transparency removal alone.
The repeal may increase the evidence gathering burden on the practices that buy and use these tools, although ONC disagreed that a significant burden shift was likely. Either way, a provider’s existing duties are separate from the developer certification duties that were removed. In a small practice, the person who turns a tool on is usually the physician owner, the practice manager or a part time IT contractor, and that is who should own the questions below.
2. What exactly were the AI model cards?
The AI model cards were the source attribute requirements in 45 CFR 170.315(b)(11)(iv), adopted in the HTI-1 rule, which required certified EHRs to support 31 named data points for predictive decision support interventions and 13 for evidence based interventions. “Model card” is shorthand; the rule required structured information and access capabilities, not necessarily a standalone document with that title. The current eCFR text, still in force until December 28, groups the 31 predictive attributes into nine categories.
Those nine categories read like a due diligence checklist. They cover the developer’s name, contact information and funding source; the intended use, intended patients, intended users and intended decision making role; cautioned out of scope uses and known limitations; the inclusion and exclusion criteria behind the training data; whether race, ethnicity, language, sex, date of birth, sexual orientation, gender identity, social determinants or health status data are used as inputs; the developer’s fairness approach; external validation; quantitative validity and fairness measures; and the schedule for monitoring, updating and correcting the tool.
Paragraph (b)(11)(v) went further. It required the EHR to give a limited set of identified users access to plain language descriptions of that information for developer supplied interventions, to indicate when specified predictive fields were not available for review, and to let users record and change source attributes. Paragraph (b)(11)(vi) required developers to apply intervention risk management practices to predictive tools, covering validity, reliability, robustness, fairness, intelligibility, safety, security and privacy.
HTI-5 removes all three paragraphs. Commenters told ONC the requirements had been in effect for less than two years. ONC replied that it had found no peer reviewed publication or trade press showing that users had relied on them or that they had improved patient care.
What survives in the decision support criterion
The criterion itself survives in narrower form. Certified products must still let a limited set of identified users select which evidence based and predictive interventions to activate, and must still capture electronic feedback data on evidence based interventions, now redesignated as 170.315(b)(11)(ii)(B), so a practice can export the intervention, the action taken, any feedback, the recorded user, the date and the location. That record does not necessarily prove a clinician meaningfully reviewed an alert. ONC pointed to that feedback capability as a remaining transparency tool, but it applies to evidence based interventions, not predictive ones.
The drug interaction and drug allergy checking reference in (b)(11)(iii) was also removed, although medications and allergies remain data types for evidence based interventions. The revised criterion also narrows the data scope for predictive interventions to the evidence based data categories plus Clinical Notes and Assessment and Plan of Treatment, so it does not preserve every aspect of the prior predictive requirement. Nothing in HTI-5 forces a vendor to stop showing information it already shows. It simply stops requiring it.
| Item | Before December 28, 2026 | After December 28, 2026 | Who carries it now |
|---|---|---|---|
| 31 predictive source attributes | Required for certification, 170.315(b)(11)(iv)(B) | Removed | Practice requests by contract or letter |
| User access to plain language descriptions | Required, 170.315(b)(11)(v) | Removed | Vendor choice; practice keeps its own file |
| Developer intervention risk management | Required, 170.315(b)(11)(vi) | Removed | Practice asks what the vendor still does |
| Selecting which interventions to activate | Required | Still required | Developer |
| Feedback data on evidence based alerts | Required, (b)(11)(ii)(C) | Still required, redesignated (b)(11)(ii)(B) | Developer builds it; covers evidence based alerts only |
| Section 1557 bias duty, 45 CFR 92.210 | Applies to covered entities | Unchanged | Practice, if it is a Section 1557 covered entity |
| HIPAA and business associate agreements | Apply | Unchanged | Practice and vendor |
3. When does the HTI-5 final rule take effect?
The HTI-5 final rule takes effect on December 28, 2026, which is the date the AI model card and risk management requirements leave the decision support criterion. The rule’s DATES section makes three amendatory instructions effective January 1, 2027 instead, and those cover criteria such as family health history, the clinical quality measure filter and two application access criteria, not the AI provisions.
The removal of paragraphs (b)(11)(iv) through (vi) sits in amendatory instruction 11, which carries no delayed date, so it lands on December 28. ONC’s own table lists the decision support interventions criterion as “Revise” with timing “Effective date of final rule.” One more date matters for interoperability planning: the removal of C-CDA and Direct requirements from the transitions of care criterion is pushed to January 1, 2030.
Two dates are worth separating for a practice. December 28 is when certification stops requiring model cards. The date your vendor actually stops publishing or maintaining them is a business decision the vendor makes on its own schedule, which is why we recommend sending the request in section 7 before the holiday slowdown rather than after it. That is a practical target, not a legal deadline imposed on practices.
4. Why ONC removed model cards over most commenters’ objections
ONC removed the model cards because it concluded they were burdensome for developers and unused by clinicians, even though, in ONC’s own words, “a large majority of commenters were not supportive” of the change. ONC reported 298 unique comment submissions on the overall proposal, and separately estimated 616 reviewers for its regulatory impact calculation. Reed Smith likewise reads the preamble as recording that a large majority opposed the model card removal; that does not mean a majority opposed every provision of HTI-5.
The objections recorded in the preamble are the ones a practice owner should take seriously. Commenters said removal would “shift evaluation, governance, and liability burdens to providers and health systems” and weaken the ability to evaluate bias, performance and appropriateness. Others warned that behavioral health and substance use settings would lose information needed to use sensitive data tools lawfully, and that funding source disclosures had helped expose hidden commercial influence on decision support.
ONC’s answer had three parts. Developers had reported that “most clinical users lack the time or technical background” to engage with source attributes during care. ONC had found no evidence the requirements improved care or organizations’ evaluation of AI. And removing the requirement “does not mean that such information about intended use, provenance, training data, limitations, performance, and known risks is unknown or unavailable,” so users should ask their developers for it.
ONC also said it believes the revised criterion makes “any significant increase or shift in the evaluation, governance, or liability burden of providers and healthcare systems unlikely.” Practices do not have to accept that prediction to plan sensibly around it. The McGuireWoods alert on the rule advises providers to obtain transparency information through contracts and internal governance, and Reed Smith calls the model card removal “a consequential change in the final rule.”
ONC left a narrow door open. In discussing the retired clinical decision support criterion for evidence based, non AI tools, it said it will “take under consideration for future rulemaking” whether elements of source attribute transparency should be restored. McGuireWoods notes ONC signaled nothing similar for predictive AI. That is a possibility, not a commitment, and nothing in the rule gives a date.
GOVERNANCE INSIGHT
A certification rule changed. Your accountability did not.
HTI-5 governs what EHR software must do to be certified. It never governed what a practice owes its patients. If a predictive alert performs badly for a patient group you serve, the questions a regulator, a payer or a plaintiff’s lawyer will ask are about your decisions: what you knew about the tool, what you asked, and what you did with the answer. Contemporaneous documentation, gathered while vendors still maintain the certification era fields, is useful evidence of reasonable efforts. It does not replace effective mitigation where a real risk is found.
5. What does HTI-5 mean for medical practices that use EHR AI tools?
For medical practices, the HTI-5 final rule means certification will no longer require the specified transparency capabilities after December 28, so a practice that relies on predictive features should confirm what information its vendor will keep providing, preserve what exists and act on it. The rule places no new obligation on providers. It removes a certification baseline that some practices were relying on, often without knowing it.
Adoption is no longer a niche concern. The AMA’s 2026 Physician Survey on Augmented Intelligence of nearly 1,700 physicians found 81 percent of respondents reporting professional use of AI, up from 38 percent in 2023, with 28 percent generating chart summaries and 17 percent using assistive diagnosis. Those are self reported figures across practice settings, not a measure of small practices specifically, but they show how common clinical AI use has become.
Which tools are affected
The removed requirements applied to decision support interventions supplied by the developer of your certified EHR. That usually means risk scores, deterioration and sepsis alerts, care gap flags, readmission predictions, and any predictive model the vendor ships inside its certified module. A no show prediction may be purely administrative in one practice and shape clinical outreach in another, so classify each tool by its actual use. Ambient scribes, patient messaging assistants and third party apps connected through an API were often outside the certified module to begin with, so HTI-5 changes less for them.
That distinction cuts both ways. A practice that assumed every AI feature came with a model card was already wrong about many add on tools. The cleanest approach is to treat every AI feature the same way regardless of which program covered it: one inventory, one request letter, one file.
What does not change
Your HIPAA Privacy and Security Rule obligations, your business associate agreements, state privacy and AI laws, FDA requirements for software that is a medical device, malpractice exposure and payer contract terms all stay exactly where they were. ONC had proposed removing the privacy and security certification criteria in 170.315(d), but none of them appear on its final table of removals, so baseline security functions such as multifactor authentication and audit reports remain certification criteria, as McGuireWoods notes. Certification is not the same as HIPAA compliance, and it does not prove every module implements every listed feature in your configuration.
California practices face more, not less, on January 1, 2027. AB 1979 and SB 503 take effect that day, and SB 503 expects deployers of clinical AI to identify, monitor and mitigate bias risks in the tools they use. For California practices, vendor documentation may be more than a contractual request: new Business and Professions Code 22758(a)(3) requires developers to make available, for clinical decision support systems identified as carrying a risk of biased impacts, a statement of intended uses and foreseeable risks plus documentation on training data and demographic representativeness where available, performance and limitations, bias mitigation, data governance and monitoring recommendations. It must be provided on request or at initial sale, whichever is earlier, and on material updates. Its scope and content differ from HTI-1, so it does not recreate the federal model card exactly. We cover those duties in our guide to California’s new clinical AI laws for small practices. A California practice losing the federal certification baseline four days before those laws start is the clearest example of why vendor documentation should now be requested on purpose.
6. Section 1557 and 45 CFR 92.210: the duty HTI-5 did not touch
The binding federal duty that still reaches clinical AI in a small practice is 45 CFR 92.210, which requires a covered entity to make reasonable efforts to identify uses of patient care decision support tools that employ “input variables or factors that measure” race, color, national origin, sex, age or disability, and to make reasonable efforts to mitigate the discrimination risk from each one. ONC cited this section by number in the HTI-5 preamble when it reminded “users of certified health IT of the importance of complying with applicable requirements to identify potential bias.”
The definitions in 45 CFR 92.4 make the duty broad. A “patient care decision support tool” is “any automated or non-automated tool, mechanism, method, technology, or combination thereof” used to support clinical decision making. That reaches an AI risk score and a paper flowsheet alike. A “covered entity” includes any recipient of federal financial assistance, so whether a particular practice is covered turns on its funding and payer relationships, which counsel should confirm. A Section 1557 covered entity is not necessarily the same thing as a HIPAA covered entity, so assess coverage under each framework separately.
Section 92.210 has three parts. Paragraph (a) prohibits discrimination through the use of these tools. Paragraph (b) imposes “an ongoing duty to make reasonable efforts to identify” tools that use the listed input variables. Paragraph (c) requires reasonable efforts to mitigate the risk for each tool identified, and the compliance date for (b) and (c) was May 1, 2025. Identification does not make a demographic input unlawful. It triggers the reasonable mitigation analysis. And because paragraph (a) is a separate prohibition, a finding that a tool has no demographic inputs is not a complete nondiscrimination conclusion if use reveals discriminatory effects.
Here is the connection a practice should not miss. Under HTI-1, the predictive source attributes included whether race, ethnicity, sex, date of birth and other demographic data were used as input features. That was exactly the kind of fact a covered entity needed for its 92.210 identification step, and date of birth is a factor that measures age even where a vendor says age is not a named field. After December 28, the certification program stops guaranteeing that fact, but the identification duty keeps running.
HTI-5 does not amend 92.210. A federal court in Tennessee v. Kennedy vacated the 2024 rule’s provisions to the extent they extended sex discrimination to gender identity, and in a June 1, 2026 notice HHS said it will continue to enforce protections based on race, color, national origin, age, disability and aspects of sex discrimination not affected by the order. Section 92.210 appeared in the eCFR version we retrieved on November 19, 2026. Applicable court orders still need to be assessed separately, so get counsel’s view in writing and keep it in the same file.
7. What to ask your EHR vendor before December 28
We recommend that before December 28 a practice ask its EHR vendor in writing for the source attribute information it currently displays or maintains for every predictive and evidence based intervention the practice has activated, including any fields marked unavailable, and whether the vendor will keep publishing and updating that information after the requirement ends. The removed regulation text is the best possible request template because it lists exactly what vendors were already preparing.
Asking before the effective date matters. Vendors certified to (b)(11) have had to support these fields for developer supplied interventions, although some predictive fields could be marked unavailable. Ask the vendor to explain any gaps. Asking for existing information is a much easier conversation than asking a vendor to recreate it after it has stopped maintaining it.
| Removed source attribute category | Plain question for your vendor | Why it matters to a small practice |
|---|---|---|
| Details and output | Who built this tool, who funded it, and is the output a prediction, a recommendation or something else? | Funding source can reveal commercial influence on a recommendation |
| Purpose | What patients, users and decisions was it designed for? | A tool built for inpatient adults may not fit a pediatric clinic |
| Cautioned out of scope use | Where should we not use it, and what are its known limitations? | Sets the boundary you train staff on |
| Development and input features | Does it use input variables or factors that measure race, ethnicity, sex, age or disability, such as date of birth, language or social determinants? | Directly answers the 45 CFR 92.210(b) identification question |
| Fairness approach | How did you test for and reduce bias? | Supports your 92.210(c) mitigation record |
| External validation | Was it validated outside the training data, by whom, and on what population? | Shows whether it was ever checked on patients like yours |
| Performance measures | What are its validity and fairness results, internal and external? | Gives a baseline to compare against your own feedback data |
| Ongoing maintenance | How often do you monitor it, and can we see results on our own data? | Local performance is what affects your patients |
| Update schedule | How often is it updated, and will you notify us before a change goes live? | A silent model update can change alert behavior overnight |
Three contract terms worth asking for
First, a commitment to keep providing the current source attribute documentation for activated interventions for the life of the contract, in plain language, on request. Second, advance written notice before a predictive intervention is materially changed, retrained or retired. Third, cooperation with your 92.210 identification and mitigation work, including disclosure of input variables and factors, and, for predictive tools, whatever decision logs, override records, subgroup performance data and outcome monitoring the vendor can supply. Many small practices sign standard vendor paper, so treat these as requests at renewal and record the vendor’s answer either way.
The final rule itself refers readers to ONC’s EHR contract guide, and ONC’s HTI-5 final rule page collects the fact sheets and the rule text. If a vendor refuses all three, that answer belongs in your file too. It tells you how much weight to put on the tool’s output and how much human review to build around it.
8. Where the CMS clinical AI roadmap fits
The CMS Clinical Super Intelligence Roadmap, posted November 13, 2026, is a nonbinding eight page federal policy statement, and it creates no compliance duty for a medical practice. Its own “What Will Not Be Done” section says it “does not mandate SI adoption” and “does not create a new agency or duplicate an existing regulator.” “Super Intelligence” is the administration’s current term for what most practices still call AI.
The roadmap still matters for planning because it signals direction. It proposes holding clinical AI to a “better than today” standard for safety and effectiveness, pursuing “national standards” instead of a patchwork of state rules, and partnering with states on AI “sandboxes.” It floats a competency based FDA premarket approach for certain AI devices and notes that certain devices may pursue authorization through established pathways, “including regulatory actions that may trigger express preemption” of differing state requirements under the Food, Drug and Cosmetic Act. The roadmap itself does not preempt state clinical AI laws.
It also names a liability framework as a goal, saying clearer regulatory pathways “will naturally lead to clearer ownership of liability.” The roadmap itself creates no new practice level mandate, implementation deadline, liability safe harbor or across the board preemption rule. Some of its initiatives build on existing FDA guidance, authorization pathways and CMS programs; others remain proposals. Read together with HTI-5, the federal posture in late 2026 is fewer certification mandates on developers, more emphasis on post market monitoring and outcomes, and an unresolved question about who is responsible when an AI tool goes wrong.
For a small practice, the honest reading is that the federal floor for AI transparency in EHRs dropped on paper this quarter, while much of the roadmap’s new agenda is still direction rather than rule text. State law, Section 1557 and your own documentation are what you can actually manage today. If you run behavioral health services, our guide to state AI therapy restrictions shows how far state law has already moved without waiting for Washington.
Getting your AI governance in order
Dynamic Comply helps small and medium sized businesses build a real AI governance foundation using the NIST AI RMF and ISO/IEC 42001.
9. What HTI-5 left unresolved
HTI-5 left the proposed information blocking changes unresolved, left the future of source attribute transparency open, and said nothing about how practices should document AI tools outside certified modules. The final rule contains none of the information blocking proposals, including one that would have expressly extended “access” and “use” to autonomous AI systems, and law firm summaries report that HHS says those comments remain under consideration.
That matters because providers are the subject of nearly 80 percent of information blocking complaints, according to the ONC data cited by McGuireWoods. Complaint share is not the same as substantiated violations or sanctions, but the existing rules and disincentives continue to apply unchanged, so a practice should not read HTI-5 as loosening anything on that front. Deferring the autonomous AI clarification also does not establish that automated requests fall outside existing information blocking law, a distinction Reed Smith flags. Covington’s summary confirms the rule is narrower than the December 2025 proposal and that later rulemaking may revisit what was left out.
The other open question is product behavior. Removal from certification is not removal from the product. Some vendors will keep model card screens because large health system customers demand them. Others will let them go stale. Until your vendor answers the section 7 request, you will not know which kind you have.
Finally, the rule does not define what “reasonable efforts” look like under 92.210 for a practice that can no longer get source attributes from its EHR. A dated request and a documented review can support evidence of reasonable efforts. Sufficiency depends on the tool, the information available, the clinical consequences, the risks identified and the mitigation actually implemented. This is our operational recommendation, not an agency approved checklist or safe harbor.
Talk to Ross about your AI governance needs
Every business has different AI governance requirements. Let us talk about yours.
10. A practical 40 day plan for a small practice
We recommend that a small practice use the roughly 40 days from mid November to December 28, 2026 to inventory its AI tools, preserve the vendor documentation that exists today, confirm future access, and record a bias review for each tool used in clinical decisions. This is a practical plan, not a federal mandate or deadline. A basic inventory can be done internally by one person with authority to own it; harder clinical, statistical, legal or security questions may need specialist help.
Days 1 to 7: inventory. List every AI or predictive feature turned on in your EHR, plus every connected app and scribe. For each, record the vendor, the module, who activated it and whether it supports a clinical or an administrative decision. Include features a vendor switched on by default during an upgrade, not only the ones someone in the practice chose.
Days 8 to 14: request. Send each vendor a written request built on the nine categories in section 7. Ask for the information currently displayed or maintained, including fields marked unavailable, and for a statement of whether it will be maintained after December 28. Keep the dated request even if the reply is slow.
Days 15 to 30: review. For each clinical tool, answer the 92.210(b) question: does it use input variables or factors that measure race, color, national origin, sex, age or disability? Record the finding, any uncertainty, the tool’s clinical role and any discrimination concerns seen in use; a “no” is not a complete compliance conclusion. Where a risk is identified, write down what you will do to reduce it, such as limiting the tool to advisory use or requiring clinician review before action. For predictive tools, ask what decision logs, override records, subgroup performance data and outcome monitoring are actually available. The surviving certification feedback requirement covers evidence based interventions and does not guarantee a predictive model monitoring dataset.
Days 31 to 40: file and assign. Put a short record per tool in a shared folder: purpose, inputs and factors, vendor response, review findings, mitigation and owner. Add contract requests to the renewal calendar. Set a quarterly reminder to recheck vendor updates. If you want a structure for the whole program, ISO/IEC 42001 and the NIST AI RMF both scale down to a practice of this size, and our article on what ISO 42001 does and does not prove explains where the standard helps.
Practices that run a patient facing chatbot should add it to the same inventory. Our earlier piece on patient intake chatbot compliance covers the HIPAA and state disclosure duties that apply to those tools regardless of HTI-5.
Frequently Asked Questions
What does the HTI-5 final rule change for AI in EHRs?
The HTI-5 final rule removes the requirement that certified EHRs support and expose source attributes, often called AI model cards, for decision support interventions, and removes developers’ intervention risk management duties for predictive tools. Predictive tools had 31 required source attributes under HTI-1. Certified EHRs must still let users choose which interventions to activate and must still capture feedback on evidence based alerts, but the rule does not require vendors to delete information they already provide.
When does the HTI-5 final rule take effect?
The rule was published in the Federal Register on October 28, 2026 at 91 FR 67916 and takes effect December 28, 2026, including the AI model card removal. A few other certification changes take effect January 1, 2027, and changes to the transitions of care criterion are delayed to January 1, 2030.
Does HTI-5 mean my practice no longer has to check EHR AI tools for bias?
No. HTI-5 changes developer certification requirements, not provider duties. If your practice is a covered entity under Section 1557, 45 CFR 92.210 still requires reasonable efforts to identify uses of decision support tools that employ input variables or factors that measure race, color, national origin, sex, age or disability, and reasonable efforts to mitigate the discrimination risk. ONC reminded providers of that section in the final rule, and HTI-5 does not amend it.
What should a small medical practice ask its EHR vendor after HTI-5?
Ask in writing for the source attribute information the vendor currently displays or maintains for each activated predictive and evidence based intervention, including fields marked unavailable, whether it uses inputs or factors that measure protected characteristics, how it was validated and monitored, and whether the vendor will keep that information current after December 28, 2026. Ask for advance notice before a model is changed or retired, and keep the request and response on file.
Does the CMS clinical AI roadmap create new rules for medical practices?
No. The Clinical Super Intelligence Roadmap posted by CMS on November 13, 2026 is a nonbinding eight page policy statement that says it does not mandate AI adoption or create a new agency. It signals future national standards, state sandboxes and FDA actions for certain devices that could trigger existing statutory preemption, but the roadmap itself creates no practice level mandate, deadline or across the board preemption.
Did HTI-5 change the information blocking rules?
No. The final rule does not include any of the information blocking changes ONC proposed in December 2025, including the proposal addressing autonomous AI access, and law firm summaries report those comments remain under consideration. The existing information blocking rules and exceptions continue to apply to providers unchanged, and the deferral does not mean automated requests fall outside them.
About the author
Ross J. is the founder of Dynamic Comply, an AI governance, compliance, and cybersecurity consulting firm based in Leesburg, Virginia. He brings more than 15 years of federal cybersecurity experience across the Department of State, the Department of Defense, and the Department of Homeland Security, and holds the CGRC certification along with credentials as a GSDC AI Compliance Lead Implementer and Auditor and Certified Ethical Hacker.
Related articles
This article is provided for general informational purposes and reflects the state of the law as of November 2026. It is not legal advice. Regulations in this area are changing quickly. Confirm current requirements and consult qualified counsel before making decisions for your organization.