Compliance Guide

Washington My Health My Data Act Applicability to LLM Prompts

Whether the My Health My Data Act applies to an LLM prompt depends on two determinations: whether the prompt, retrieved context, output, or tool-call payload contains information that qualifies as consumer health data, and whether your organization acts as a regulated entity or a processor for it. Applicability to prompt data is interpretive, so confirm each conclusion against the statute, Attorney General guidance, and counsel. This guide maps the data flow points to review and the runtime controls that can support the resulting obligations. It is not legal advice.

Start with four definitions, applied to prompt data

The statutory definitions should be applied to prompt data at every point in the data flow, not only to what a user types. Two determinations drive the outcome:

  1. Whether the prompt, retrieved context, output, or tool-call payload contains information that qualifies as consumer health data.
  2. Whether your organization acts as a regulated entity or a processor for that data.

Outputs can infer health-related details as well as repeat them, so document the statutory analysis for inferences with counsel.

Points in an AI data flow where obligations may be triggered

Each of the following points can carry data that warrants review. Walk through them in order when you inventory a system.

Where prompt data can trigger review
Data flow pointWhat to review
User prompt and outputFree text can carry health-related details, and outputs can infer them.
Retrieved contextDocuments and records added to a prompt by retrieval.
Provider transmissionPrompt data sent to a third-party model provider.
Logs and memoryObservability stores and persistent agent memory.
Tool callsPayloads passed to downstream systems by agents.

Separate legal obligations from engineering controls

Decide the legal questions first, then choose controls. Whether data qualifies as consumer health data, and whether you act as a regulated entity or processor, are interpretive conclusions. Runtime controls support the obligations that follow from those conclusions; they do not replace the analysis.

As part of that analysis, review provider terms on retention and model-improvement use.

Runtime controls and their limits

Detection and redaction controls need to be tested rather than assumed. Run them against realistic prompts and record the false-negative rate you observe. Deletion is similarly worth exercising: delete one test identity across logs, memory, and downstream tools, and note where data remains.

Where runtime governance fits

Runtime policy enforcement, tool governance, and audit logging can support your data governance requirements for AI agents. They act at the points listed above: prompts and outputs, retrieved context, provider transmission, logs and memory, and tool calls.

Review checklist

  1. Inventory every place prompts, retrieved context, outputs, and tool-call payloads are sent or stored.
  2. Document the statutory analysis for consumer health data, including inferences, with counsel.
  3. Determine whether you act as a regulated entity or processor, and review provider terms on retention and model-improvement use.
  4. Test detection and redaction against realistic prompts and record the false-negative rate you observe.
  5. Run a deletion exercise for one test identity across logs, memory, and downstream tools.

Map your prompt data flows to runtime controls

Talk with Trussed AI about how runtime policy enforcement, tool governance, and audit logging can support your data governance requirements for AI agents.

Talk to an Expert