Compliance Guide

    Identifying race, age, sex, and disability inputs in Amazon Bedrock Agents under Section 1557 Section 92.210

    45 CFR 92.210(c) requires a covered entity to make reasonable efforts to mitigate the risk of discrimination for each patient care decision support tool identified under paragraph (b). For a Bedrock Agent, that means mapping every path by which race, color, national origin, sex, age, or disability data, or proxies for them, can reach the agent's output, deciding which interactions are in scope, applying mitigation, and keeping records that show both steps were carried out.

    Mapping the regulation to a Bedrock Agent's input surfaces

    1. 1

      User prompts

      Free text can include age, sex, race, language, or disability details. Staff may also paste records into a prompt. Assume anything a person can type is an input.

    2. 2

      Session and prompt session attributes

      Application-supplied context such as demographics, eligibility category, or care setting travels with the session. These values are chosen by your developers, so they are the easiest to inventory and the easiest to forget.

    3. 3

      Knowledge base retrievals

      Retrieved documents and records can contain protected data the user never typed. The corpus, not the query, determines exposure.

    4. 4

      Action group parameters

      The agent decides which values to pass to downstream APIs. Review which parameters can carry demographic fields or proxies.

    5. 5

      Tool and API responses

      Data returned by scheduling, claims, EHR, or scoring systems re-enters the agent's context and can influence the next step.

    What Section 92.210 asks of covered entities

    Section 1557 prohibits discrimination on the basis of race, color, national origin, sex, age, or disability in certain health programs and activities. HHS issued the implementing final rule, published in the Federal Register on May 6, 2024. Section 92.210 addresses nondiscrimination in the use of patient care decision support tools. Paragraph (c), titled mitigation of risk, states that for each tool identified in paragraph (b), a covered entity must make reasonable efforts to mitigate the risk of discrimination resulting from the tool's use in its health programs or activities. As of November 2026, the current eCFR text of that paragraph shows no amendment, delay, or repeal.

    Two points follow from the wording. First, the mitigation duty attaches to tools that have been identified, so identification is the step on which everything else depends. Second, the standard is reasonable efforts. The regulation does not set a numeric threshold or a list of required controls, so each entity has to define its own criteria and document why they are reasonable.

    The 2022 proposed-rule preamble gives useful context. HHS pointed to increasing reliance on clinical algorithms and said their implementation may be discriminatory, and it discussed assessment tools used for resource allocation that screen out, or tend to screen out, individuals with disabilities. That is proposed-rule discussion, not final text, but it shows the kind of harm the provision targets. The final rule also added 'or any combination thereof' to § 92.101(a)(1) to account for people who experience discrimination on more than one protected basis, which matters when you assess inputs that interact, such as age together with disability.

    The rule text does not mention agents, but an agent that recommends or shapes patient care decisions is a reasonable candidate for review under the same logic. Whether a given agent interaction counts as a patient care decision support tool is a scoping decision your organization must make, record, and assign to a named owner.

    Building a repeatable identification practice

    Identification works as an operating routine, not a one-time project. The goal is a record that shows what you examined, what you concluded, and when. The steps below follow the identify-then-mitigate sequence described in practitioner commentary on the rule, and the review cadence is your decision because the regulation does not prescribe one.

    • Define scope: Write down which agent interactions count as patient care decision support, who decided, and on what basis.
    • Inventory agents and tools: List each agent and every tool, knowledge base, and action group it can call.
    • Map input paths: For each agent, record which of the five input surfaces could carry race, color, national origin, sex, age, or disability data or proxies, and how you verified it.
    • Collect vendor input lists: Ask suppliers of embedded decision support tools for their input variables and keep the responses.
    • Record the mitigation decision: For each identified tool, document the action taken, such as removing an input, adjusting a threshold, adding review, or stopping use, with the rationale.
    • Re-review on change: Repeat the mapping when an agent, knowledge base, or action group changes, and on a fixed schedule you set.

    Identification versus mitigation, and where runtime enforcement fits

    Identification tells you where protected inputs can appear. Mitigation is what you do about it. Practitioner commentary lists mitigation options as removing an input, adjusting a threshold, adding review, or stopping the use. These are suggestions, not regulatory requirements, but they map naturally to runtime actions: allow, redact, block, or escalate for human approval.

    That mapping is where a design-time inventory meets operation. A document that says an agent should not receive a patient's disability status does not prevent a retrieval or API response from carrying it. Policy enforcement in the execution path applies your decision to each interaction as it happens. A rule might redact a demographic field from a tool response before the agent sees it, block an action group call whose parameters include a field you have excluded, or route a recommendation to a clinician for approval.

    Trussed provides this kind of enforcement in the execution path, with actions including allow, redact, block, and escalate for human approval. Enforcement applies to individual interactions and tool calls and is enforced at runtime, not reviewed after the fact. It depends on decisions your organization makes: which inputs are in scope, which rules apply, and who approves escalations. It does not replace the scoping and inventory work above, and it does not determine on its own whether a tool is covered by the regulation.

    Evidence that supports a reasonable-efforts position

    An auditor or regulator is likely to ask two questions: how did you identify protected-characteristic inputs, and how do you know mitigation was applied. The regulation does not prescribe evidence formats or retention periods, so you need to design this yourself. A defensible record set typically includes the scope decision and its owner, the agent and tool inventory, the input-path mapping with verification notes, each mitigation decision with rationale and date, review dates, and change records.

    The second question is harder, because design documents show intent and not behavior. Evidence generated by enforcement closes part of that gap. When each allow, redact, block, or escalation decision is recorded as it occurs, you can show that a stated mitigation was applied to specific interactions and tool calls. Trussed produces continuous audit-ready evidence from enforcement, supports self-managed deployment in the customer cloud, and offers configurable retention and data residency, which lets you align records with your own retention policy.

    This evidence supports a documented reasonable-efforts position. It does not guarantee compliance, and it does not cover decisions made outside the interactions and tool calls where policy is enforced. Pair it with the written identification record, assign ownership (the Section 1557 Coordinator has overall responsibility for Section 1557 compliance, but your organization should name who owns 92.210 specifically), and have counsel confirm the current regulatory status before relying on any date or scope assumption.

    Where protected characteristics can enter an agent

    User prompts

    Free text from clinicians, staff, or patients

    Session attributes

    Context passed alongside the conversation

    Knowledge base retrievals

    Retrieved passages and records

    Action group parameters

    Values the agent sends to APIs

    Tool and API responses

    Data returned to the agent mid-run

    Turn your input map into enforced policy

    Discuss how policy enforcement in the execution path and audit-ready evidence can support your 92.210 identification and mitigation records.

    Explore Agent and Tool-Call Enforcement