Implementation Guide | Healthcare

    Section 1557 limited English proficiency AI agent handoff: routing patients from Bedrock Agents to qualified interpreters

    Implement the handoff as an enforced runtime control, not a prompt instruction. When a limited English proficiency signal appears, stop substantive agent responses, route the patient to a qualified interpreter service, pass only the minimum PHI needed, and record the trigger, policy decision, payload and outcome as evidence that supports HIPAA audit controls under 164.312(b). This is not legal advice.

    Handoff sequence, triggers and fallback

    1. 1

      1. Signal detected

      A trigger from the inventory is recorded with its source and confidence.

    2. 2

      2. Policy decision

      Runtime policy evaluates the signal and moves the session to handoff state.

    3. 3

      3. Responses blocked

      Substantive agent output stops; only the fixed handoff message is allowed.

    4. 4

      4. Payload built

      A minimized context payload is assembled and redacted before leaving the agent.

    5. 5

      5. Interpreter connected

      The session routes to a qualified interpreter service or queue.

    6. 6

      6. Timeout fallback

      If no interpreter connects, the patient stays queued and a human owner is alerted.

    Implementation checklist

    • Define which agent outputs are critical to rights, benefits or meaningful access, and whether the agent may respond in a non-English language at all.
    • Write the LEP trigger inventory with an action and an ambiguity rule for each signal.
    • Enforce a block on substantive responses and PHI-retrieving tool calls once a trigger fires.
    • Verify the interpreter service meets the Part 92 qualified interpreter definition, and settle the HIPAA arrangements before routing PHI.
    • Specify payload fields, redaction rules and fallback behavior with no patient-supplied interpreter.
    • Test trigger, block, redaction, connection and timeout paths end to end, and retain the evidence.

    What Section 1557 requires and what it leaves open

    As of November 2026, the language access provisions of 45 CFR Part 92, including § 92.201 on meaningful access for individuals with limited English proficiency (LEP), remain in the current text. Several provisions bear directly on an agent deployment.

    Part 92 defines a qualified interpreter as someone who, through a remote interpreting service or an on-site appearance, has demonstrated proficiency in speaking and understanding both spoken English and at least one other spoken language. Relay interpreters must demonstrate proficiency in two non-English spoken languages. The rule prohibits requiring an LEP individual to provide their own interpreter or to pay for one, and it prohibits relying on staff who are not qualified interpreters, qualified translators or qualified bilingual or multilingual staff. It includes a temporary-measure exception for emergencies involving an imminent threat to safety or welfare when no qualified interpreter is immediately available. Under § 92.201, machine translation used where the text is critical to rights, benefits or meaningful access, where accuracy is essential, or where the language is complex, non-literal or technical must be reviewed by a qualified human translator. Covered entities must also give notice that language assistance services are available free of charge.

    The rule text addresses machine translation of text. How it applies to live agent dialogue is a question for counsel. The architectural objective is clearer: an LEP patient should not be left relying on the agent's own unreviewed machine-generated language where a qualified interpreter or qualified human review is required.

    Treat the handoff as a runtime control point

    A prompt telling the agent to offer an interpreter is a request to the model, not a control. The model can ignore it, misread the signal or keep answering. A reviewable design places a policy decision at the boundaries where the agent produces a response or calls a tool. In a Bedrock Agents deployment, that means the response path to the patient and the tool-call path, such as Lambda-backed actions that read or write patient data.

    Once an LEP trigger fires, the policy decision should change the session state. Further substantive clinical or administrative responses are blocked. Tool calls that retrieve PHI for the conversation are blocked or limited. The agent may emit only a fixed, pre-approved handoff message that states interpreter services are available at no cost. That message supports the notice of availability, though how the notice duty applies to an agent interface is a matter for counsel.

    The routing target should be an organization-vetted interpreter service or queue that can show its interpreters meet the Part 92 definition. Remote video or audio interpreting is a recognized delivery mode. Family members, friends and untrained staff are not valid targets, and the flow must never offer bring-your-own-interpreter as a fallback.

    Trussed enforces policy in the execution path for individual agent interactions and tool calls, with allow, redact, block, rate limit and escalate for human approval available as decisions. That is the layer where the blocked response and the redacted payload described here would be enforced at runtime, not reviewed after the fact.

    Minimize PHI in the handoff payload

    Build the payload by minimization. Pass what the interpreter session needs to connect the patient with the right human service, such as the language, the reason for contact at a general level, and a session reference. Withhold or redact the rest, including free-text transcript content, identifiers not needed for routing, and clinical detail the interpreter will receive directly from the patient or the clinician in the live session.

    The Security Rule applies to electronic PHI that is created, received, used or maintained by a covered entity or its business associate, so the transfer path matters as well as the content. Its technical safeguards include access controls, audit controls, integrity, person or entity authentication and transmission security. Each of these maps to a design question: who can read the payload, how the interpreter service is authenticated, and how the payload is protected in transit.

    Agree the contractual and HIPAA arrangements with the interpreter vendor before any PHI is routed. Earlier rulemaking indicates that a covered entity may share information with a qualified interpreter who is a workforce member or another type of agent, such as a contractor, but the current application should be confirmed with counsel. Redaction should be a policy decision applied to the payload at runtime, so the same rule applies to every handoff and not only to those a developer remembered to handle.

    Evidence each handoff should produce

    A regulator or auditor will want to know what the agent was permitted to do, when the handoff triggered, and what the agent did afterward. Design the record around those questions. For each handoff, retain the trigger signal and its source, the policy that was in force, the decision taken (allow, redact, block or escalate), the payload as sent after redaction, the escalation target, the interpreter connection time and the outcome, including any fallback. Also retain responses the agent blocked after the trigger, since those show the control held.

    These fields are a design recommendation aimed at the audit controls standard at 164.312(b), not a list the regulation prescribes. Pull the exact standard text from the current regulation when you finalize your mapping, and set retention periods with counsel and your compliance team.

    Evidence is strongest when it is produced by the enforcement decision itself, not reconstructed from application logs. Trussed generates continuous audit-ready evidence from enforcement and maps it to HIPAA, and it runs in production with healthcare organizations. It can be deployed self-managed in the customer cloud, with configurable retention and data residency. Whatever tooling you use, test the full handoff and keep the test records as evidence too.

    LEP handoff control path

    Detect

    Language preference, interpreter request or non-English input raises an LEP signal.

    Enforce

    Policy stops further substantive agent responses and limits tool calls.

    Route

    A minimized payload goes to a qualified interpreter service or queue.

    Evidence

    Trigger, decision, payload and outcome are retained for audit.

    Make the handoff an enforced control

    See how policy enforced in the execution path can block agent responses, redact handoff payloads and produce audit-ready evidence for each interaction and tool call.

    Request a Demo
    Next step

    How does your AI governance program compare?

    See where your program has gaps in less than 2 minutes.