Answer phone and internet customers' questions with agents
Fastweb + Vodafone built a customer-service agent for nearly 9.5 million customers; the company reports 90% correct answers and 82% of queries resolved.
Done with LangGraph
What they did
Fastweb + Vodafone, an Italian phone and internet company, rebuilt its TOBi chatbot as Super TOBi using LangGraph, a tool for building agents. A lead agent applies safety rules, passes chats to humans when needed and sends each question to a specialist agent. Those agents look up customer accounts and can take actions like switching on an offer or updating payment details. A second system, Super Agent, guides call-center staff through troubleshooting steps stored as a linked map of procedures.
How it went
The company reports Super TOBi serves nearly 9.5 million customers with 90% correctness and an 82% resolution rate. Customers rated the effort needed at 5.2 out of 7. It says Super Agent pushed problems solved in a single call above 86%.
Worth knowing
Super Agent never talks to customers directly; it tells the human staff member the next step. Business experts write procedures in fill-in templates, and these go live within hours.
Try it yourself with LangGraph
Help me design a LangGraph customer-support agent for [your business]. Start with a Supervisor node that filters unsafe or invalid input, handles greetings and handovers to a human, and routes each question to one of these use cases: [list of use cases, e.g. billing, account changes]. For each use case, define which [your APIs or data sources] it may call and what it should answer. Any action that changes an account, such as a payment or a cancellation, must be shown to the customer or a staff member for confirmation before it runs. Show me the graph design and a test plan with sample conversations, and wait for my approval before writing code.
Swap the highlighted [parts] for your own details.
Source: LangChain blog · LangChain Team · Dec 16, 2025
Written with AI assistance from the original post and checked against our editorial guidelines. Results are as reported by the source, not independently verified.
