Healthcare AI playbook for private practices and clinics.

Get the compliance layer right

Step zero is not picking a tool. It's confirming your AI usage path is BAA-covered for any workflow touching PHI. Without this, every other decision compounds risk.

• Identify which AI platforms your organization has — or can get — a signed BAA with. Claude Enterprise, ChatGPT Enterprise, Microsoft Copilot on M365 with HIPAA configuration are the common paths.
• Map workflows to data classes: which touch PHI (must be BAA-covered), which touch internal-confidential, which touch fully de-identified or public.
• Get sign-off from your compliance officer in writing on the deployment path. One paragraph is enough; it just needs to exist.

Five workflows with real ROI

1. Patient communications & correspondence

Referral letters, lab result explanations, prior-auth letters, post-visit summaries. Cuts physician documentation time 40-60% with a well-built Project. Highest single ROI workflow we see in practice rollouts.

2. Patient portal triage

Categorize incoming portal messages (clinical-urgent, clinical-routine, billing, refill) and draft initial responses. Reduces MA triage time 60-80%; reduces inbox-related physician burnout.

3. Visit note structuring

Convert dictated notes into structured documentation matching your EHR template. Physicians dictate; AI structures. Frees evening charting time.

4. Prior authorization workflows

Drafting prior-auth narratives from clinical context. Significant time recovery; insurance authorizations are universally hated work.

5. Internal operations

Scheduling letters, employee comms, policy updates, vendor correspondence. Non-PHI work that nonetheless eats admin time.

Workflows where AI is not ready

• Clinical decision support — interesting research, not yet appropriate for production at most practices. Liability exposure outpaces benefit.
• Direct-to-patient chatbots — quality bar is much higher; one bad answer to a clinical question is a catastrophic outcome.
• Automated coding/billing decisions — fraud and abuse exposure if AI codes without human review. Use AI to draft; require human sign-off.
• Imaging interpretation — specialized; use FDA-cleared tools designed for the purpose, not general LLMs.

What practice rollouts look like

What works, what does not

• Works: One physician champion who uses the tools and shows colleagues. Group education sessions led by the champion.
• Works: Letting physicians see the time-savings on their own work before mandating use. Convert with data, not directives.
• Does not work: Top-down adoption mandates. Physicians resist.
• Does not work: Generic AI training. Specific workflow training does work.
• Does not work: AI rollout led by IT without clinical co-leadership.

What to measure

1. Hours of physician evening charting recovered (most important; biggest burnout driver).
2. Inbox response time on patient portal messages.
3. Referral letter turnaround time.
4. Prior authorization submission time.
5. Physician burnout / satisfaction metrics over 6-12 months.

Related frameworks & reading

• Claude for healthcare →
• Case study: 3-doctor practice recovered 12 hrs/week →
• How to use AI in a medical practice →
• Treetop Implementation →