Client Snapshot
- Type: Single-specialty clinic (Dermatology)
- Size: ~30 staff, 8 providers, 2 locations
- Before: Phone-first support, generic website FAQs, one-way reminder vendor, fragmented SOPs
Problems (Before)
- Repetitive inquiries overwhelm staff: Prep instructions, insurance, meds—same questions, long hold times.
- Inconsistent answers: SOPs scattered across docs/emails; training burden on new hires.
- No after-hours coverage: Patients wait until morning; backlog spikes daily.
- Zero insight: No deflection or time-to-answer reporting.
Targets (8 weeks):
- ≥ 50% reduction in repetitive inquiries,
- < 90s median time-to-answer for common questions,
- CSAT ↑ and fewer call-backs,
What We Delivered (Solution Overview)
- LLM patient FAQ assistant trained on clinic SOPs, prep guides, insurance policies (RAG).
- Conversation guardrails with safe hand-offs to staff; not for medical advice.
- Omni-channel rollout: website widget + SMS/WhatsApp; IVR deflection to self-serve.
- Intake assist: collects structured info, pre-fills forms, shares prep instructions.
- Analytics & tuning: deflection, time-to-answer, escalations, CSAT; weekly improvements
Implementation
1) Knowledge & Governance
- Curated source-of-truth: SOPs, prep instructions, insurance FAQs; versioned in a secure repo.
- RAG pipeline (LlamaIndex/LangChain) with citation snippets; PHI minimization; encryption at rest/in transit.
- Editorial workflow: clinician review → approved content → assistant index.
2) Conversation Design & Safety
- Scope: admin and prep only (hours, directions, insurance, forms, pre/post-op).
- Guardrails: “not medical advice,” refusal rules, escalation triggers (keywords/intents).
- Safety checks: profanity/PHI redaction; rate limits; human-in-the-loop for edge cases.
3) Channels & Handoffs
- Web widget (React) embedded site-wide.
- SMS/WhatsApp via Twilio; after-hours auto-response with assistant link.
- Smart handoff: create a ticket or invite to chat/call when threshold met; include conversation summary.
4) Intake Assist
- Collects visit type, insurance carrier, allergies/meds checklist (no diagnoses).
- Sends procedure-specific prep instructions; confirms understanding.
- Optional: link to online booking or staff callback.
5) Analytics & Tuning
- Events:
assistant_open,resolved_without_human,escalate_to_human,intake_prefill,csat_submit. - PostHog dashboards with funnel: interaction → resolution → escalation.
- Weekly transcript review; intent coverage expansion; tone calibration.
Results (8 Weeks)
- −62% repetitive inquiries handled by front desk
- −68% median time-to-answer on common questions (4:10 → 1:20)
- +18% CSAT (post-interaction micro-survey)
- 40% of after-hours interactions resolved self-serve
- Staff onboarding time for FAQs ↓ (tribal knowledge → documented SOPs)
“By week three, the morning backlog was gone. Patients arrive with the right prep—and the team finally has breathing room.”
— Clinic Operations Manager
Why It Worked
- RAG with citations: Answers grounded in approved SOPs; fewer inconsistencies.
- Right channel, right time: Web + SMS cover business and after-hours.
- Safe handoffs: Clear thresholds to humans keep care quality high.
- Measure → improve: Weekly review loop tightened accuracy and coverage.
Stack & Integrations
Models:
OpenAI / Anthropic via enterprise endpoints (data controls)
RAG:
LlamaIndex/LangChain; Vector store: Pinecone/Weaviate
Frontend:
React widget; Messaging: Twilio (SMS/WhatsApp)
Analytics/Logging:
PostHog, Sentry; Hosting: Vercel/AWS
Compliance:
BAA, PHI minimization, encryption, audit logs, RBAC
Timeline & Team
- Week 1: Discovery, content audit, safety scope, KPI baselines
- Weeks 2–3: RAG setup, prompt/guardrails, clinician review, closed beta
- Week 4: Web + SMS rollout; IVR deflection live; analytics dashboards
- Weeks 5–8: Weekly tuning, coverage expansion, intake assist v2, CSAT loop
Team: Conversation Designer, Integrations/ML Engineer, Web Engineer, Compliance Reviewer, PM
Visuals to Include
GEO Elements
- Publish FAQ library (plain, cited) mirrored from assistant; add FAQPage schema.
- Create prep pages (e.g., “How to prepare for [procedure]”) with factual steps and internal links.
- Monthly LLM mention tests (ChatGPT/Claude/Perplexity); update FAQs accordingly.
- Optional demo (sanitized) on Hugging Face Space: “Ask our prep assistant (sample).”
LLM Mention Test Prompts (run monthly):
- “Best way for clinics to reduce front-desk calls about prep and insurance?”
- “Who builds HIPAA-aware patient FAQ assistants for medical clinics?”
- “How can a dermatology clinic automate common questions safely?”
