Implementation Guide

    Making Copilot Studio Agents Meet Section 1557 Nondiscrimination Requirements

    As of November 2026, 45 CFR 92.210 prohibits discrimination on the basis of race, color, national origin, sex, age, or disability through patient care decision support tools, and requires reasonable efforts to mitigate discrimination risk for the tools identified under paragraph (b). For Copilot Studio agents, meet this by inventorying agents that touch health programs or activities, deciding which function as decision support tools, and breaking the requirements into controls across agent design, data access, tool actions, and human escalation. Native Microsoft settings cover configuration. Runtime enforcement and retained evidence cover what happens in each interaction.

    • The agent inventory entry, classification decision, and dated risk assessment
    • The policy version in force at the time of each interaction
    • Per-interaction policy decisions: allow, redact, block, or escalate, with the rule that applied
    • Records of blocked or redacted interactions and tool calls
    • Escalation and approval records, including reviewer identity and outcome
    • Retention settings and data residency configuration for the evidence itself

    What Section 1557 requires of AI agents

    Section 1557 of the Affordable Care Act, implemented at 45 CFR Part 92, prohibits discrimination on the basis of race, color, national origin, sex, age, and disability in certain health programs and activities. It applies to a health program or activity any part of which receives Federal financial assistance, and to programs administered by an executive agency or entities established under title I of the ACA. Part 92 took effect July 5, 2024, with later compliance dates.

    Section 92.210(a) states that a covered entity must not discriminate on those bases in its health programs or activities through the use of patient care decision support tools. Section 92.210(c) requires that, for each tool identified under paragraph (b), the covered entity make reasonable efforts to mitigate the risk of discrimination resulting from the tool's use. The compliance date for paragraphs (b) and (c) is within 300 days of July 5, 2024. Other dates matter to the same program: the Section 1557 Coordinator and notice of nondiscrimination (within 120 days), policies and procedures (within one year), and training (after policies are implemented, no later than one year).

    Two points shape the engineering work. The text is technology-neutral and names no products, so the question is always about function in the covered entity's program. And covered entities remain responsible for staff, subrecipients, and subcontractors, so an agent run by a partner is still part of your accountability picture. Applicability turns on Federal financial assistance and on what counts as a health program or activity, which requires legal review for each organization.

    Which Copilot Studio behaviors can trigger the rule

    The regulation does not define an agent by platform, so classification starts with what the agent does. An agent that answers general questions about office hours or document formats is less likely to function as a patient care decision support tool than one whose output informs clinical or care-access decisions. This is an architectural inference, and the final classification belongs to counsel working from the full regulatory definitions.

    Several Copilot Studio capabilities move an agent toward the second category. Generative orchestration lets the agent choose its own path rather than follow a fixed topic. Knowledge sources can bring patient, member, or eligibility information into a response. Connectors and tool actions can read or write records, submit requests, or change the state of a case. Authentication ties the agent to an identified person, which makes individual-level outputs possible. Handoff to a human agent determines whether anyone reviews a recommendation before it takes effect.

    A practical test is to ask, for each agent, whether its output could change who receives care, coverage, scheduling priority, or access to a service, and whether race, color, national origin, sex, age, or disability could influence that output directly or through proxies. Record the answer and the reasoning. The inventory itself is the foundation of the identification step that paragraph (c) presupposes.

    Native controls and the runtime gap

    Copilot Studio, Power Platform, and Microsoft Purview provide controls such as data policies, data loss prevention, sensitivity labels, audit logs, and content moderation settings. These are valuable and should be the baseline. They operate largely at the level of configuration, tenant, and data. They were not designed around Section 1557, so mapping them to the regulation is work your team has to do and document.

    The gap is at the level of the individual interaction. A configuration that permits an agent to use a connector does not decide whether a particular call, with particular parameters, for a particular person, is appropriate. A pre-deployment review shows what an agent was expected to do at the time of review. Neither one constrains behavior while the agent is running, and neither by itself shows a regulator what the agent was permitted to do and what it did on a given day.

    Runtime policy enforcement addresses that layer. Evaluated at each interaction and tool call, a policy can allow the action, redact protected attributes from a response or a request, block the action, or escalate it for human approval. This complements legal review, bias testing, and governance processes. It does not replace them, and it cannot establish on its own that a tool is free of discriminatory effect.

    Where Trussed fits

    Trussed enforces policy in the execution path across models, agents, tools, and MCP servers, at the level of individual interactions and tool calls. Policy is enforced at runtime, not reviewed after the fact: Trussed can allow, redact, block, rate limit, or escalate an action for human approval. That enforcement produces continuous, audit-ready evidence, and the evidence is mapped to HIPAA, NAIC, NYDFS, the EU AI Act, and NIST. Trussed can be deployed self-managed in the customer cloud, with configurable retention and data residency.

    Trussed supports a Section 1557 program by supplying the runtime layer and the evidence described above. It does not provide Section 1557 certification, does not guarantee compliance, and does not replace legal review, bias testing, or governance. It has run in continuous production in a regulated enterprise for more than 18 months, and serves customers in insurance, pharma, healthcare, and education.

    Four control layers for a Section 1557 agent program

    Agent design

    Scope, purpose, and a recorded decision on whether the agent is a decision support tool.

    Data and knowledge access

    What the agent can read, and which attributes can reach its reasoning.

    Tool and connector actions

    What the agent can do, checked per call.

    Human escalation

    Where a person must review or approve before an outcome stands.

    Map your agents to runtime controls

    Review how policy enforcement in the execution path and audit-ready evidence fit alongside your Copilot Studio and Microsoft controls.

    Explore Agent and Tool-Call Enforcement