The Client
A cosmetic surgery practice whose revenue lives or dies on the front of the funnel: website and chat inquiries that need fast, accurate, empathetic responses. Every consultation is a high-value decision for the patient — and every slow reply, unanswered question, or missed follow-up is a booked procedure that walks to a competitor.
The Problem
The patient journey in a surgical practice is really eight different jobs: qualifying new leads, reading how a nervous patient actually feels, educating them about procedures, preparing them for consultation, scheduling, checking in after surgery, tracking what the funnel is doing, and keeping the CRM honest. Clinics usually split those jobs across a coordinator's inbox and memory. Things fall through — not because staff don't care, but because no human can watch every conversation at once.
A single general-purpose chatbot doesn't fix this. One prompt trying to be a salesperson, a nurse, a scheduler, and an analyst at once does each job badly — and in an elective-medical setting, "badly" carries real regulatory and reputational risk.
The Solution
We built CosmeticOS as an AI workforce, not a chatbot: eight specialized agents, each owning one job, coordinated by a workflow engine and constrained by a governance layer that decides what an agent may say and when it must hand off to a human.
01 Lead Conversion — qualifies and scores incoming patient chats
02 Sentiment — reads hesitation, anxiety, and buying signals in real time
03 Education — answers procedure questions from approved clinic material
04 Consultation Prep — briefs patient and surgeon before the visit
05 Scheduling — books and manages consultations
06 Post-Op — follow-up care check-ins after procedures
07 Analytics — funnel and conversation metrics
08 CRM — keeps every patient record current automatically
- A patient chat is scored live by the agent workforce — qualification, sentiment, and next-best-action, updated turn by turn
- A governance layer bounds each agent: approved knowledge only, defined escalation triggers, human handoff for clinical questions
- Three role-specific live dashboards — Coordinator, Practice Manager, and Surgeon — each seeing the same patients through their own lens
- Clinic memory persists context across the whole journey, so the post-op agent knows what the lead-conversion agent learned weeks earlier
- Graceful degradation: if the LLM is unreachable, a rule-based fallback keeps core flows working instead of going dark
- Full test suite across the agent, core, and governance packages
The Stack
A TypeScript monorepo: the agent workforce and workflow engine in one package, clinic domain logic (qualification, governance, memory) in a core package, and a React front end with the patient chat and the three dashboards. An Express server ships it all as a single container, with Claude as the reasoning model via OpenRouter.
The Outcome
The practice gets a front office that never sleeps: every inquiry answered immediately and scored honestly, coordinators working from a live prioritized queue instead of a cold inbox, surgeons walking into consultations pre-briefed, and no patient forgotten after surgery. The system is live — you can open the demo below and watch the agents score a conversation in real time.