AI Assistant for Therapists (EHR Copilot)

An effective EHR copilot integrates at three key surfaces: the clinical workspace, the client record, and the document composer. In platforms like TherapyNotes or TheraNest, this means connecting via secure APIs to read and write to objects like Client, Appointment, ProgressNote, and TreatmentPlan. The assistant should be invoked contextually—for example, from within a SOAP note editor to generate a "Subjective" section based on the last session's note, or from a client dashboard to summarize recent PHQ-9 trends.

Implementation typically involves a middleware layer that sits between the EHR and LLM APIs (like OpenAI or Anthropic). This layer handles PHI de-identification for external calls, maintains a vector index of practice-specific documents (consent forms, treatment protocols), and enforces a clinician-in-the-loop pattern where all AI-generated content is presented as a draft for review and edit. Key workflows include:

• Data Retrieval: "Find all clients with a GAD-7 score increase >5 points in the last month."
• Documentation Drafting: "Draft a progress note for today's CBT session focusing on cognitive restructuring."
• Workflow Navigation: "Show me the steps to submit a prior authorization for this client's service."

Rollout requires a phased, role-based approach. Start with a pilot group of clinicians, enabling the copilot for low-risk tasks like data queries and note templating. Govern usage with audit logs that track all prompts and generated content, linked to the clinician's EHR user ID. Crucially, the system must never act autonomously; all writes back to the EHR (saving a note, updating a treatment plan) require explicit clinician approval. This architecture reduces documentation time from hours to minutes while keeping the clinician firmly in control of the clinical record.

The copilot is triggered by clinician actions within the EHR interface—such as opening a client record, starting a progress note, or reviewing a treatment plan. A secure middleware layer, acting as a HIPAA-compliant orchestrator, captures these events via the EHR's API (e.g., TherapyNotes' REST API) or a secure webhook. This layer packages only the necessary, de-identified context—like client ID, note type, and previous session summaries—into a request to a Business Associate Agreement (BAA)-covered LLM provider such as Azure OpenAI or Anthropic. PHI is never sent directly to a third-party model; all calls are logged with full audit trails for compliance under 42 CFR Part 2.

The AI's response—a draft SOAP note, a suggested intervention, or retrieved data points—is returned to the middleware and presented to the therapist within a dedicated UI panel or inline suggestion in the EHR. The clinician retains full control: they can accept, edit, or reject the AI's output. Any accepted content is written back to the appropriate EHR object (e.g., a ProgressNote or TreatmentPlan record) via the same secure API, maintaining the system of record. For complex retrievals, we implement a RAG (Retrieval-Augmented Generation) system using a vector database like Pinecone, indexing practice-specific documents (consent forms, clinic policies) and anonymized, aggregated client history to ground suggestions in relevant context.

Rollout follows a phased governance model: starting with read-only data retrieval (e.g., "summarize last three sessions") to build trust, then progressing to draft generation with mandatory review (e.g., note drafting), and finally to automated workflow triggers (e.g., flagging a missed PHQ-9). Each step includes clinician training, clear opt-in/opt-out controls, and continuous monitoring of AI usage logs and feedback. The architecture is designed to fail gracefully; if the AI service is unavailable, the EHR workflow continues uninterrupted, preserving the clinician's primary tool.

Every interaction must be designed for PHI protection. This starts with a Business Associate Agreement (BAA) with your LLM provider (e.g., Azure OpenAI, Google Vertex AI) and extends to your integration architecture: data is encrypted in transit and at rest, prompts and responses are never used for model training, and all AI activity is logged to a dedicated audit trail within your EHR or a separate SIEM. For platforms like Valant or SimplePractice, this means API calls must be tokenized and scoped to the minimum necessary data, and any AI-generated content written back to the EHR must be tagged with its AI origin for full traceability.

The rollout should be phased, starting with low-risk, high-reward assistive functions. Phase 1 often focuses on documentation support: the AI suggests a SOAP note objective section based on the session transcript, which the therapist reviews, edits, and finalizes. Phase 2 introduces data retrieval: a therapist can ask, "What was the client's PHQ-9 score trend over the last three sessions?" and the RAG system fetches and summarizes only that client's data. Phase 3 expands to proactive support, like flagging a missed assessment or suggesting a treatment plan update based on progress note themes—always requiring clinician review and sign-off.

Governance is continuous. Establish a clear protocol for handling AI errors or hallucinations, including a quick feedback mechanism within the EHR interface. Regularly review audit logs for unusual access patterns and re-evaluate prompt designs to minimize clinical risk. The goal is not to replace clinical judgment but to create a clinician-in-the-loop system where the AI acts as a tireless, knowledgeable assistant, and the therapist remains the ultimate decision-maker. This approach builds trust, ensures compliance with HIPAA and 42 CFR Part 2, and allows the practice to scale AI's benefits safely across documentation, care coordination, and patient communication workflows.