An LLM copilot clinicians use every week
I led the design of LLM-powered analysis for therapy-session data. It cut manual review time by about 35%, and about 70% of eligible clinicians use it every week, with a clinician approving every output.
2 min read
less manual review time
weekly use among eligible clinicians
active users at 10+ customers
The problem
Therapists and BCBAs (Board Certified Behavior Analysts) spent hours reviewing session notes and behavior data to write summaries and update treatment plans. The work was repetitive and error-prone, and it took time away from client care and supervision.
My role
I defined the use cases, designed the review experience, and partnered with Engineering on the integration, evaluation and rollout.
What I did
- Prioritized three use cases: session summaries, progress insights and anomaly flags.
- Designed every AI output as an editable draft that a clinician reviews, edits and approves.
- Partnered with Engineering to connect structured session data to a frontier LLM through a HIPAA-aligned backend.
- Set up clinician feedback loops to tune prompts and measure accuracy.
Safety and evaluation
- No protected health information (PHI) lives on client devices. PHI is encrypted in transit and at rest, and the system de-identifies data before it reaches a prompt, with field-level filters, role-based access and audit logs.
- Senior BCBAs wrote gold-standard session summaries. We compared human-only drafts and clinician-edited AI drafts against them for correctness, completeness and clinical tone.
- Edit distance and the share of drafts accepted with only minor changes guided prompt tuning and guardrails.
The hard call
The copilot assists and the clinician decides. I made clinician review mandatory for every draft and accepted a slower rollout to protect trust, quality and compliance.
Results
- Piloted with about 25 clinicians across three organizations.
- Expanded to about 120 active users across 10+ customers.
- About 70% of eligible clinicians use AI-assisted summaries every week.
- Manual therapy-data review time fell by about 35%, and clinicians kept final control of every output.