Add a PHI-detector step in Zapier before calling Claude: if PHI is found, redact/store in your HIPAA queue and route to human review; if clean, send redacted text to Claude for JSON-only intent classification. Use a few-shot prompt and strict schema: {intent, confidence(0–1), phi_flag, recommended_route}. Escalate for confidence <0.8. Claude often follows safety/system prompts more conservatively than ChatGPT, but still validate with logs and tests.

See Zapier vs Make for integration options: Compare Zapier and Make