Customer support · Production

Fastweb + Vodafone deploys Super TOBi and Super Agent with LangChain and LangGraph for AI-powered customer service at enterprise scale

The problem

Traditional TOBi struggled with nuanced customer requests requiring contextual understanding, multiple system access, and end-to-end resolution, while call center agents had to manually consult multiple systems and knowledge bases for each interaction.

First attempt

Traditional TOBi, the existing chatbot, could not handle nuanced requests requiring contextual understanding, access to multiple systems, or end-to-end resolution.

Workflow diagram · grounded in source
1
Supervisor filters and routes query
routing
“The Supervisor acts as the central entry point for all user queries. Its first responsibility is to apply guardrails, filtering, and shaping inputs to ensure they are valid and safe. Beyond this, it manages special scenarios such as the …”
2
Use Case agents execute API calls
ai_action
“They operate with access to a well-defined subset of customer APIs and follow the LLM Compiler pattern. This approach allows them to reason about which APIs should be invoked, to coordinate multiple steps where required, and to generate …”
3
Action tags execute transactions
output
“an action tag might initiate an offer activation, disable an ongoing service, or trigger a payment method update. When such action tags are returned, the ChatBot automatically executes them in the conversation interface.”
4
Business template ETL to knowledge graph
integration
“an automated ETL pipeline—powered by LangGraph and task-specific LLM Agents (including ReAct Agents)—parses the document into JSON, extracts verification APIs, performs consistency checks, and refines step definitions. The content is dec…”
5
Intent routing for consultant requests
routing
“Incoming requests from consultants, whether troubleshooting support requests on user issues or general knowledge base questions, are first processed by a LangGraph Supervisor who determines whether the request matches a graph-based proce…”
6
Graph-based procedure executor
ai_action
“For structured troubleshooting and fault-isolation scenarios, the Supervisor activates a procedural sub-graph executor. Using LangChain and LangGraph, the system retrieves the first Step node of the procedure from Neo4j along with its as…”
7
Graph RAG for open-ended questions
ai_action
“Generic or unstructured questions about the company knowledge base are routed to a hybrid Retrieval-Augmented Generation (RAG) chain that combines a vector store with the Neo4j knowledge graph. The vector store finds a broad set of poten…”
8
Daily LangSmith automated evaluation
feedback_loop
“The team has developed sophisticated evaluation processes that run daily, automatically classifying chatbot responses and providing structured feedback for continuous improvement”
Reported outcome

Super TOBi serves nearly 9.5 million customers achieving a 90% correctness rate, 82% resolution rate, and a Customer Effort Score of 5.2 out of 7.
Super Agent drives One-Call Resolution rates above 86%.

Reported metrics
Super TOBi correctness rate90%
Super TOBi resolution rate82%
Customer Effort Score (CES)5.2 out of 7
One-Call Resolution (OCR) rateabove 86%
Show all 5 reported metrics
Super TOBi correctness rate90%
Super TOBi resolution rate82%
Customer Effort Score (CES)5.2 out of 7
One-Call Resolution (OCR) rateabove 86%
customers served by Super TOBinearly 9.5 million
Reported stack
LangChainLangGraphNeo4jLangSmithReAct Agents
Source
https://blog.langchain.com/customers-vodafone-italy/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Super TOBi serves nearly 9.5 million customers achieving a 90% correctness rate, 82% resolution rate, and a Customer Effort Score of 5.2 out of 7.

What tools did this team use?

LangChain, LangGraph, Neo4j, LangSmith, ReAct Agents.

What results were reported?

Super TOBi correctness rate: 90%; Super TOBi resolution rate: 82%; Customer Effort Score (CES): 5.2 out of 7; One-Call Resolution (OCR) rate: above 86% (source-reported, not independently verified).

What failed first in this deployment?

Traditional TOBi, the existing chatbot, could not handle nuanced requests requiring contextual understanding, access to multiple systems, or end-to-end resolution.

How is this customer support AI workflow structured?

Supervisor filters and routes query → Use Case agents execute API calls → Action tags execute transactions → Business template ETL to knowledge graph → Intent routing for consultant requests → Graph-based procedure executor → Graph RAG for open-ended questions → Daily LangSmith automated evaluation.

Related customer support cases