Back to playbooks

AI for Care (with Guardrails)

AI can help clinicians draft and organize. It must not become an unaccountable decision machine.

AI should do more of

• paperwork drafting (with clinician review)
• summarizing patient history (with citations/links to source)
• benefits/billing navigation for patients
• prior auth packet assembly (not decision-making)
• routing and triage support (with human accountability)

AI should NOT do

• opaque denial decisions
• “optimize revenue” coding without transparency
• replace informed consent conversations
• high-stakes triage without clear human ownership and appeal paths

Always required

• auditability (why did it recommend that?)
• an appeal path (for patients and clinicians)
• responsibility assigned to a human role

Data boundaries

• know whether the tool is allowed to handle PHI at all
• know where processing happens and whether the environment is vendor-approved
• know whether inputs or outputs are retained, logged, or used for model training
• do not move PHI into a random tool because the workflow is annoying

Sign-off

Any production use should have sign-off from:

• clinical lead
• privacy / compliance
• IT / security

Non-negotiable documentation rule

AI-generated text is still documentation. The licensed human who signs it owns the content and the decision.

Plain-language safety rule

If you can’t explain it, audit it, and appeal it, it doesn’t belong between a patient and their care.

Back to playbooks