What HTI-5 Changes for Health Systems Running LangChain or LlamaIndex Decision Support
HTI-5 is a final rule, announced on October 26, 2026, that reduces certification burden for health IT developers. Its final-rule table removes the clinical decision support criterion at §170.315(a)(9) and revises the decision support interventions criterion at §170.315(b)(11). The rule targets certified developers, not health-system-built applications, and HIPAA 164.312 applies separately to the covered entity. Whatever the certification outcome, a LangChain or LlamaIndex application still needs enforced policy and audit-ready evidence at each step of its execution path.
Mapping the framework execution path
- 1
Retrieval
PHI enters the context window here. Decide which sources a given use may query, and what is redacted before the content reaches the model. Record what was retrieved.
- 2
Prompt assembly
System instructions, retrieved text and user input are combined in code. Capture the policy in force and the assembled inputs for the interaction.
- 3
Model call
Control which models may receive which data classes. Record which model was called and what policy decision applied.
- 4
Tool call
Agents can invoke tools that read or write systems. Enforce an allow list, and route sensitive actions to human approval before they execute.
- 5
Output
Check what is returned to the clinician or downstream system. Redact or block as policy requires, and keep the decision on record.
- Read the final §170.315(b)(11) text and the effective date before committing to design assumptions.
- Inventory every LangChain and LlamaIndex application, noting whether it is predictive, clinician-facing, operational only, and inside or alongside the EHR vendor's module.
- Ask the EHR vendor which of its decision support features remain certified under (b)(11).
- Document data sources, tools and outputs for each application that touches PHI.
- Mark each execution step as enforced at runtime or only documented after the fact.
- Run the sample-interaction test: can stored records show what the AI was permitted to do and what it did?
What HTI-5 changes and what it leaves alone
HTI-5 amends the ONC Health IT Certification Program. HHS announced the final rule on October 26, 2026, describing it as a deregulatory action that removes and revises requirements to reduce compliance costs for certified health IT developers, in alignment with Executive Order 14192.
Two entries matter for decision support. The final-rule table lists the clinical decision support criterion at §170.315(a)(9) as removed and the decision support interventions criterion at §170.315(b)(11) as revised. Both take effect on the final rule's effective date. As of November 2026, architects should read the final (b)(11) text itself before hard-coding any design around specific source-attribute fields, because the revision changes a criterion that the 2024 HTI-1 rule shaped.
What HTI-5 does not do is turn a health system's own application into a certified product. It is a certification rule, written for developers. HIPAA Security Rule duties under 45 CFR 164.312 sit in a separate regulation and apply to covered entities and business associates. A relaxed certification criterion therefore does not relax the technical safeguards a health system owes for the systems it operates.
The final-rule table also does not list the (d) security criteria, including auditable events at §170.315(d)(2), among its removals or revisions. Do not assume those criteria have been removed.
Who carries which obligation
The 2024 HTI-1 rule narrowed Predictive DSI source-attribute and intervention risk management duties to Predictive DSIs supplied by the health IT developer as part of its Health IT Module. It described the knowledge, decision and ongoing compliance for including a Predictive DSI as sitting solely within the control of the certified developer. HTI-5 revises (b)(11), so that scoping may be superseded in part, but its logic is a useful starting point for allocation.
In practice, split the portfolio into three groups. The first is decision support the EHR vendor supplies inside its certified module. Certification duties there fall on the vendor, and you should ask which features remain certified under (b)(11) after HTI-5. The second is health-system-built applications that run alongside the EHR, such as a LangChain retrieval assistant or a LlamaIndex pipeline over internal guidelines. The third is custom applications embedded in clinical workflow. Whether a given embedded deployment counts as supplied as part of a module is a fact-specific question, so treat any direct HTI-5 obligation on the second and third groups as unsettled rather than assumed.
The HIPAA position is clearer. If an application touches PHI, the health system is responsible for technical safeguards, including audit controls under 164.312(b), whichever group the application falls in. Record the allocation, with named owners in clinical informatics, security and compliance, for each application.
HIPAA 164.312(b) alongside HTI-5
Because HTI-5 changes certification criteria and 164.312 governs the operator of the system, the two should be mapped separately. HTI-5 questions are about what a certified developer must supply. The 164.312(b) question is whether your own audit controls cover the systems in which PHI is handled, including AI applications you built. Review the full paragraph text with compliance counsel and map it to your framework's instrumentation.
The practical gap is that application logs rarely answer the questions an auditor asks. A trace of function calls shows that code ran. It does not show what the AI was permitted to do under policy at that moment, which tool calls were refused, or which fields were redacted before the model call. Those facts exist only at the time of execution. If the decision is not recorded when it is made, it cannot be reconstructed later.
A useful test is to take one sample interaction and try to answer two questions from stored records alone: what was this AI permitted to do, and what did it do? If the first answer depends on someone reading the code, the evidence is incomplete.
Where runtime enforcement fits
Policy that lives in a document or a prompt instruction is advisory. Policy enforced in the execution path decides, per interaction and per tool call, whether to allow, redact, block, or escalate for human approval, and it produces a record of that decision as a side effect. This is the narrow role Trussed plays: policy enforced at runtime, not reviewed after the fact, with audit-ready evidence generated by that enforcement and mapped to HIPAA. Trussed runs in production with healthcare and pharma organizations.
The scope is individual interactions and tool calls. Trussed does not provide model provenance or lineage, and does not trace agentic workflows end to end, so those questions need other answers in your design. Self-managed deployment in the customer cloud, with configurable retention and data residency, keeps PHI-adjacent evidence under the health system's control.
Where obligations sit for framework-based decision support
HTI-5
Certification program rule aimed at certified health IT developers.
EHR vendor
Decision support supplied within its certified Health IT Module.
Health system
Custom applications built on LangChain or LlamaIndex, plus HIPAA duties as a covered entity.
Execution path
Retrieval, prompt assembly, model call, tool call and output.
Evidence
What the AI was permitted to do and what it did, per interaction.
Test your execution path against the evidence an auditor would ask for
Walk through one framework-based decision support application and see where policy is enforced and where evidence is produced.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.