Testing Microsoft Copilot Studio Patient Agents for Section 1557 Language Access and Disability Accessibility
Scope the test plan by mapping each agent task to the Section 1557 provisions it touches: language access (45 CFR 92.201), effective communication (92.202), accessible information and communication technology (92.204) and patient care decision support tools (92.210(a)). Write test scenarios and pass/fail criteria for each, with escalation to qualified interpreters or staff as an explicit pass condition. Pre-release tests show behavior on scenarios only, so requirements that depend on what the agent did with real patients need runtime enforcement and per-interaction evidence. As of November 2026, Part 92 is current in the eCFR.
Define pass and fail criteria and escalation conditions
- Language accuracy: qualified reviewers rate task outputs against a written rubric. Any error that changes a patient's understanding of cost, coverage or next steps is a fail.
- Timeliness: the patient reaches an answer or a handoff within the limit you set, with no charge or added burden.
- Privacy and independence: the agent does not require a companion to interpret or complete the task for the patient.
- Assistive technology: each in-scope task can be completed with a screen reader and by keyboard alone, on every deployed channel.
- Escalation: when the agent cannot identify or serve a limited English proficiency patient accurately, or cannot provide an effective communication aid, a documented handoff to a qualified interpreter or staff member is the pass condition.
- Procedures alignment: agent routing matches your written language access procedures, including how staff identify limited English proficiency and obtain qualified interpreters and translators.
Scope the test plan by task, channel and provision
Start with an inventory of what the Copilot Studio agent actually does: scheduling, benefits questions, care navigation, or a combination. For each task and each channel through which patients reach it, record which Section 1557 provisions apply. The agent is not a single compliance object. A scheduling flow raises different questions than a care navigation flow that suggests next steps.
Four provisions carry most of the weight. Section 92.201 requires reasonable steps to provide meaningful access for individuals with limited English proficiency, including companions. Section 92.202 requires communications with individuals with disabilities, including companions, to be as effective as communications with people without disabilities. Section 92.204, as described in HHS OCR guidance, requires health programs and activities delivered through information and communication technology to be accessible to individuals with disabilities, subject to fundamental-alteration and undue-burden limits. Section 92.210(a) bars disability discrimination through patient care decision support tools.
Whether a given agent flow counts as a patient care decision support tool is a scoping decision your compliance team should make and document. HHS has also stated, in the Section 504 rulemaking, that disability discrimination can occur through the use of algorithms, automated decision-making and artificial intelligence. A narrow scheduling bot and a care navigation agent should not receive the same scope entry by default.
Build language access scenarios from the regulatory criteria
Section 92.201(b) gives the test design its structure. Language assistance must be free of charge, accurate and timely, and must protect the privacy and independent decision-making ability of the individual. Each of those becomes a scenario family.
For accuracy, run each core task in the languages your organization serves and have qualified reviewers assess the output. Include benefits and cost explanations, where errors carry the most consequence. For timeliness, test whether a patient who writes in a language other than English reaches a usable answer or a handoff without unreasonable delay. For cost, confirm that no flow asks a patient to bring their own interpreter or pay for assistance. For privacy and independence, check that the agent does not push a patient toward relying on a family member or companion to complete the task.
Add companion scenarios, since the obligation extends to companions. Also test the notice of availability under 92.11, which must be provided in English and at least the 15 languages most commonly spoken by individuals with limited English proficiency in the relevant State or States, and in alternate formats for individuals with disabilities who require them. Confirm that a patient can reach it from the agent in each required language.
Treat machine translation with care. HHS OCR guidance says a qualified human translator must review machine translations of critical documents when accuracy is essential. Decide in advance which agent outputs fall into that category and route them accordingly.
Build disability accessibility scenarios around the patient who cannot use the default path
The core question for 92.202 is whether a patient with a disability can complete the same task with communication that is as effective as everyone else receives. Auxiliary aids and services must be provided in accessible formats, in a timely manner, and free of charge, in a way that protects the individual's privacy and independence.
Test the chat experience with assistive technology that patients commonly use, including screen readers and keyboard-only navigation, and across the channels you deploy. Federal training material on this topic stresses that communications and websites must be accessible to people with vision, hearing and speech disabilities, including interoperability with commonly used assistive technology. Test output as well as interface: check that responses are in plain language, that forms and confirmations are readable by assistive technology, and that time-limited steps do not lock out patients who need longer to respond.
Then test the failure path. If the agent is the fastest route to book an appointment, a patient who cannot use it must still have another route that works. Scenarios should confirm that the agent offers that route and does not simply repeat the same inaccessible step.
What pre-deployment testing shows and what production must show
Pre-release tests show how the agent behaved on the scenarios you wrote. They do not show what it did with patients you did not anticipate, in languages or phrasings you did not test, after the next configuration change. Written language access procedures matter here too: OCR considers their effectiveness when evaluating compliance, and effectiveness is a production question.
The split below is a control design, not a list of regulatory requirements. Your organization decides which items it needs to demonstrate.
- Pre-deployment testing: Accuracy of translations on a test set, screen reader and keyboard operability, notice of availability reachable in required languages, escalation path functions in a scripted case.
- Runtime enforcement: Policy applied to each real interaction and tool call: allow, redact, block, or escalate for human approval, enforced at runtime rather than reviewed after the fact.
- Runtime evidence: Per-interaction records of which policy applied, what the agent was allowed to do, and whether an escalation occurred and to whom.
- Ongoing review: Compliance reviews escalation records and changes to scope, languages and procedures, then updates test cases to match.
Where runtime policy enforcement fits
Some requirements can only be shown through production records. Examples include whether a patient who needed an interpreter was actually handed off, whether a response in a covered language was blocked or escalated, and whether sensitive patient data was redacted before reaching a tool. A test report cannot answer these questions.
Trussed is not a testing product and does not replace the test plan above. It enforces policy in the execution path for individual interactions and tool calls, with actions to allow, redact, block or escalate for human approval. Because enforcement generates the record, a regulated enterprise running in healthcare gets audit-ready evidence of what the agent was permitted to do and what it did, mapped to HIPAA among other frameworks. Trussed runs in production with healthcare organizations, and deploys self-managed in the customer cloud with configurable retention and data residency.
The scope is bounded. Enforcement applies per interaction and per tool call. It does not trace an entire agentic workflow end to end, and it does not test translation quality. Your accuracy rubric and accessibility testing remain your own responsibility. Define the policies that correspond to your pass/fail criteria, such as an escalation rule when the agent cannot serve a patient in their language, and then confirm the resulting evidence is retained and retrievable for the period your counsel requires.
From obligation to evidence
Map
Tie each agent task to the Part 92 provisions it touches.
Test
Build language and accessibility scenarios with written pass/fail criteria.
Escalate
Define handoff to qualified interpreters, bilingual staff and auxiliary aids.
Evidence
Record what the agent did with real patients, per interaction.
Plan production evidence alongside your test plan
Talk through which Section 1557 requirements for your patient agents need enforcement and per-interaction evidence in production.
Explore Agent and Tool-Call EnforcementHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.