Customer support · Production

Doctolib builds Alfred, an agentic AI system for healthcare support with human-in-the-loop action execution

The problem

As Doctolib's platform grew, support request volume increased and linearly scaling the support team was unsustainable, while conventional chatbots frustrated users with rigid decision trees and free-text fields that delivered nothing actionable.

Workflow diagram · grounded in source
1
User submits support request
trigger
“hey Alfred, I'd like to give Maria Smith read-only access to my home consultations calendar”
2
Alfred understands and clarifies
ai_action
“Understands what customers need, even when it's not perfectly articulated”
3
Data Retriever gathers customer information
ai_action
“One of our agents is the "Data Retriever" — a specialist focused solely on gathering customer information. While it has deep access to our customer data APIs, it can only use a carefully curated set of endpoints.”
4
RAG engine enriches AI responses
ai_action
“we developed a Retrieval Augmented Generation (RAG) engine that enriches AI responses with our support knowledge base”
5
Alfred crafts action request
ai_action
“an AI agent decides it's time to perform an update on the agenda authorizations it will craft the complete request (url, http method and payload) and pass it to a deterministic node”
6
Deterministic fact-check
validation
“this node will - Ensure the parameters were not hallucinated by the LLM (fact-checking) - Send it to an Action Request Checker responsible for fetching fresh data for all referenced resources, and returning both technical and human-reada…”
7
User confirms proposed action
human_review
“The final step of changing agenda accesses always remains in the users' hands”
8
Action executed
output
“HCP submits the confirmation button”
Reported outcome

Doctolib built Alfred as a proof-of-concept agentic AI system handling calendar access management, with a human-in-the-loop design ensuring the LLM never directly executes sensitive actions, and plans to expand to additional support scenarios.

Reported metrics
Support cases per business day~1,700
Messages daily~17,000
Reported stack
LangGraphLiteral.aiLLMRAGKeycloak
Source
https://medium.com/doctolib/building-an-agentic-ai-system-for-healthcare-support-a-journey-into-practical-ai-implementation-0afd28d716e6
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Doctolib built Alfred as a proof-of-concept agentic AI system handling calendar access management, with a human-in-the-loop design ensuring the LLM never directly executes sensitive actions, and plans to expand to add…

What tools did this team use?

LangGraph, Literal.ai, LLM, RAG, Keycloak.

What results were reported?

Support cases per business day: ~1,700; Messages daily: ~17,000 (source-reported, not independently verified).

How is this customer support AI workflow structured?

User submits support request → Alfred understands and clarifies → Data Retriever gathers customer information → RAG engine enriches AI responses → Alfred crafts action request → Deterministic fact-check → User confirms proposed action → Action executed.

Related customer support cases