Section 1557 and healthcare AI agents

    Building a Section 1557 Patient Care Decision Support Tool Inventory for Copilot Studio, Bedrock, and ADK Agents

    Under 45 CFR 92.210, a covered entity must make reasonable efforts, on an ongoing basis, to identify patient care decision support tools that use input variables measuring race, color, national origin, sex, age, or disability, and to mitigate discrimination risk for each tool it identifies. For AI agents, scope by the regulatory definition rather than by platform, record the same fields for every agent on Copilot Studio, Amazon Bedrock, and Google ADK, and use runtime records of what each agent was permitted to do and did to keep the inventory current. The inventory supports the entity's own risk determinations. It does not replace them.

    Inventory fields that work across all three platforms

      What 45 CFR 92.210 requires

      Section 92.210 has three parts. Paragraph (a) prohibits discrimination on the basis of race, color, national origin, sex, age, or disability in a covered entity's health programs or activities through the use of patient care decision support tools. Paragraph (b) creates an ongoing duty to make reasonable efforts to identify uses of those tools that employ input variables or factors that measure any of those characteristics. Paragraph (c) requires reasonable efforts to mitigate the risk of discrimination for each tool identified under (b).

      The definition is broad. A patient care decision support tool is any automated or non-automated tool, mechanism, method, technology, or combination thereof used by a covered entity to support clinical decision-making in its health programs or activities. Under 45 CFR 92.1, compliance with paragraphs (b) and (c) was due within 300 days of July 5, 2024. As of November 2026, the section remains codified in the eCFR.

      Two features shape inventory design. The duty is ongoing, so a point-in-time list will drift. The standard is reasonable efforts, so the record should show how decisions were reached, not only which tools exist.

      Scope agents by the definition, not the platform

      An AI agent is rarely one tool. It is typically a model, a set of instructions, one or more tools or actions, connectors to other systems, and knowledge sources. Because the definition covers combinations, any of those elements may form part of what supports a clinical decision. Scoping by platform, or by whether something is labeled AI, will miss agents and may over-include others.

      Start with one question per agent: with its tools and data, is it used to support clinical decision-making in a health program or activity? Record the answer, who made the determination, and the reasoning. Record out-of-scope decisions as well, since a reviewer will want to see why an agent was excluded. Whether a given agent qualifies is the covered entity's determination.

      For each in-scope agent, then record whether its inputs include variables that measure the protected characteristics in paragraph (b). Those inputs may come from the user's message, a knowledge source, or a tool output. Some entities also note variables that could stand in for those characteristics. That is a conservative practice for the entity to decide, not a field the regulation spells out.

      Keeping one inventory current across independent teams

      Different teams build agents on different platforms, and the tools, connectors, and knowledge sources attached to an agent can change after launch. An inventory built from a one-time survey will not reflect those changes, which is a problem when the duty to identify is ongoing.

      Define review triggers as a design choice: a new connector, a change of model, a new or modified knowledge source, a material change to system instructions, or a new clinical use. Each trigger should route the agent back to the scoping question. Restricting who can change or activate an agent, and recording that authority in the inventory, makes the triggers easier to enforce.

      The preamble to the final rule ties mitigation to the covered entity becoming aware of potential discrimination. Complaints, monitoring results, and clinical review findings should therefore feed back into the inventory as well, so that a mitigation determination reflects what the entity has learned since the last review.

      Where runtime enforcement produces evidence

      An inventory states what an agent is supposed to be. Runtime records show what it did. Trussed enforces policy in the execution path on individual interactions and tool calls, with actions that include allow, redact, block, and escalate for human approval. Enforcement happens at runtime, not in review after the fact, and each decision generates audit-ready evidence that is mapped to HIPAA, including the 164.312(b) audit control requirement. Trussed runs in production with healthcare and pharma organizations.

      This evidence supports the inventory in three ways. First, a covered entity can compare the tools and data an agent actually invoked with what the inventory lists, and treat a mismatch as a review trigger. Second, a policy that blocks calls to unlisted tools, or redacts specified fields before they reach a model or tool, turns an inventory attribute into an enforced constraint with a record behind it. Third, escalation for human approval records a human review point for the interactions where the entity decided one is needed.

      The scope is deliberately narrow. Enforcement and evidence apply to individual interactions and tool calls. They do not constitute governance of an entire agentic workflow, and they do not decide whether an agent is a patient care decision support tool.

      The limits of an inventory

      The duties in paragraphs (b) and (c) belong to the covered entity. A well-kept inventory and strong runtime evidence make identification and mitigation demonstrable, but neither performs the risk identification or makes the mitigation determination. Those judgments, including whether an input variable measures a protected characteristic and what mitigation is reasonable, stay with the entity's clinical, compliance, and legal reviewers.

      The standard is reasonable efforts, so the most useful artifact is a record of reasoning: why an agent was scoped in or out, what mitigation was chosen and why, who approved it, and when it will be reviewed again. Runtime evidence adds the observed behavior that lets a reviewer check those determinations against reality.

      A practical next step is to take one agent per platform, run it through the scoping question and the six fields above, and note where the entity cannot currently answer from existing records. Those gaps show where ownership, change control, or runtime evidence needs attention first.

      From agent definition to inventory evidence

      Scope

      Is the agent, with its tools and data, used to support clinical decision-making?

      Record

      Model, instructions, tools, connectors, knowledge sources, inputs, owner.

      Mitigate

      Document the determination, approver, and review date for each identified tool.

      Evidence

      Compare what each agent was permitted to do with what it did on individual interactions and tool calls.

      See what runtime evidence looks like for your agents

      Review how policy enforcement on interactions and tool calls produces audit-ready evidence that supports your Section 1557 inventory and your own mitigation reviews.

      Request a Demo