How to Investigate a Patient Complaint About an AI-Drafted Message
Treat the complaint as an evidence problem, not a model-quality debate. Record the complaint and its dates, preserve every record tied to the message, reconstruct what the AI was permitted to do, what it accessed and what it produced, check whether a human approved it, then determine whether PHI was disclosed. Document the findings and remediation so the record stands on its own. Evidence generated at runtime, tied to the policy decision for each interaction, makes this faster and more defensible.
Reconstruct the interaction and verify the controls
- 1
Establish the timeline
Order every record by timestamp and confirm clock consistency across systems.
- 2
Define the permitted scope
Document what the AI was authorized to access and do for this use case at the time.
- 3
Trace the PHI
List the PHI the AI received, what appeared in the draft and what reached the recipient.
- 4
Verify the policy decision
Confirm the policy version applied and whether it allowed, redacted, blocked or escalated.
- 5
Analyze the review checkpoint
Determine whether a human saw the draft, what they changed and whether review was required.
- 6
Confirm delivery
Match the sent message and recipient to the messaging or EHR record.
Preserve and collect the evidence
- Input and context: the prompt, retrieved content and any patient record context supplied to the AI
- PHI accessed: which fields or documents the AI could see, and which were redacted before drafting
- Policy applied: the policy and version in force at that moment, and the decision it returned
- Human approval events: reviewer identity, edits made, and approval time
- Output delivered: the final message as sent, the recipient, the channel and the delivery time
- Timestamps and identities: the user, service and system identities attached to each step
Scope the complaint at intake
Start by classifying what the patient is actually alleging. Common categories are inaccurate content, inappropriate tone, and disclosure of PHI to the wrong person. Each category implies different evidence and a different level of regulatory exposure, so let the allegation determine the investigative tasks rather than applying one template to every complaint.
Record the date of the complaint and the date the individual says they learned of the issue. As of November 2026, HHS OCR may act on a complaint filed within 180 days of the alleged violation, or within 180 days of when the individual should have known of it. Those dates also shape how quickly you need to preserve evidence. Note that OCR reviews each complaint to decide whether it has authority to investigate and may close it based on the facts presented. Your internal record may be the only one that is ever examined, so build it to stand alone.
Identify the roles involved. If a vendor or other third party participated in drafting or delivering the message, record which party is the covered entity and which is a business associate, because OCR's complaint jurisdiction extends to both. As a general investigative practice, interview both the person raising the concern and the staff member who sent or approved the message, and consider whether the order of those interviews could identify the complainant to the people being questioned.
Make the determination and document it
The human-review checkpoint often decides how the finding is framed. If review was required and the approval event is missing, that is a control failure. If review occurred, the record should show what the reviewer saw. A reviewer who approved a draft containing PHI intended for another person is a different finding from an automated send with no reviewer at all. Keep both questions separate: whether the control existed and whether it operated.
Then assess whether a HIPAA-relevant event occurred, such as PHI reaching the wrong recipient. Document how you reached that conclusion, who was affected and how you determined the number of affected individuals. The count matters because OCR investigates all breaches affecting 500 or more individuals, while smaller breaches may be investigated depending on resources and enforcement priorities. Run the formal breach assessment and any notification analysis with your privacy officer and counsel, using the full text of the Breach Notification Rule.
Complaints about inaccurate content or tone may not involve PHI disclosure, but they still need a recorded outcome. Write down the allegation, the evidence reviewed, the policy and control findings, the determination, the remediation and the owners. Remediation typically covers the policy rule that failed or was absent, the review requirement, and any retraining. Test the fix against the same scenario before closing the record.
What the execution path should produce
Most investigations are slow because evidence is assembled after the fact from systems that were never designed to explain a decision. The alternative is to enforce policy in the execution path and have that enforcement emit the evidence. For each interaction, the enforcement layer should record the policy version, the decision (allow, redact, block or escalate for human approval), what was redacted, the approver and approval time, the output released and the identities involved. These are design requirements derived from the reconstruction steps above. They are not prescribed by regulation.
Trussed enforces policy in the execution path for individual interactions and tool calls, and produces audit-ready evidence mapped to HIPAA, including 164.312(b) audit controls. Because enforcement happens at runtime, not reviewed after the fact, the decision and its evidence come from the same event. Whether a given evidence set satisfies 164.312(b) for AI-generated communications is a question for your privacy officer and counsel against the full regulatory text.
Preservation also depends on deployment choices. Trussed supports self-managed deployment in the customer cloud, with configurable retention and data residency, so organizations can keep interaction evidence under their own access controls and set retention to match their legal hold practices. Trussed runs in production with healthcare and pharma organizations, and it generates more than 30 GB of audit evidence per day across its production traffic.
When evaluating any tooling, ask whether it can show, for one message, the policy version applied, the decision taken, the PHI redacted, the approval event with identity and timestamp, and whether records can be placed on legal hold and exported to the privacy officer.
Investigation lifecycle
Intake and triage
Record the complaint, dates, parties and allegation type.
Evidence preservation
Hold prompts, outputs, policy records and delivery logs.
Reconstruction
Rebuild inputs, PHI accessed, output and delivery.
Policy and control verification
Confirm which policy applied and what it decided.
Human-review analysis
Establish who reviewed, edited or approved.
Determination and remediation
Decide on PHI disclosure, document, and fix.
Make the next investigation faster to defend
See how enforcement in the execution path produces per-interaction evidence your privacy and compliance teams can use.
Request a DemoHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.