AI-Powered Customer Service and Call Center Transformation with Multi-Agent Systems

fastweb_/_vodafone · Telecommunications · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Telecommunications customer service and call-center assistance

Fastweb + Vodafone needed AI systems that could resolve nuanced customer requests across channels and help consultants diagnose issues quickly across customer data, service history, procedures, APIs, and knowledge bases.

Constraints:

• Telecommunications support includes billing, activation, roaming, sales, and technical-support flows
• Customers expect resolution within one interaction
• Internal consultants need compliant, source-backed next steps rather than a customer-facing bot response

Architecture

Fastweb/Vodafone built two LangGraph systems: Super TOBi, a customer-facing agentic chatbot with a Supervisor and specialized Use Case agents, and Super Agent, an internal consultant copilot over Neo4j graph procedures and hybrid RAG. Business-authored templates are converted by ETL and task-specific LLM/ReAct agents into a knowledge graph; consultants' requests route to graph procedures or vector/graph RAG, with LangSmith traces, daily evaluation, and human oversight.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  super-tobi-chatbot["super-tobi-chatbot"]
  supervisor-agent["supervisor-agent"]
  use-case-agents["use-case-agents"]
  chatbot-action-executor["chatbot-action-executor"]
  super-agent-copilot["super-agent-copilot"]
  business-template-etl["business-template-etl"]
  neo4j-knowledge-graph["neo4j-knowledge-graph"]
  consultant-request-router["consultant-request-router"]
  crm-context-injector["crm-context-injector"]
  graph-procedure-executor["graph-procedure-executor"]
  hybrid-rag-chain["hybrid-rag-chain"]
  langsmith-monitoring["langsmith-monitoring"]
  super-tobi-chatbot --> supervisor-agent
  super-tobi-chatbot --> chatbot-action-executor
  supervisor-agent --> use-case-agents
  use-case-agents --> chatbot-action-executor
  super-agent-copilot --> consultant-request-router
  business-template-etl --> neo4j-knowledge-graph
  neo4j-knowledge-graph --> graph-procedure-executor
  neo4j-knowledge-graph --> hybrid-rag-chain
  consultant-request-router --> crm-context-injector
  consultant-request-router --> graph-procedure-executor
  consultant-request-router --> hybrid-rag-chain
  langsmith-monitoring --> super-tobi-chatbot
  langsmith-monitoring --> super-agent-copilot

Components

Data flow

For Super TOBi, a customer query enters the Supervisor, passes guardrails and input shaping, routes to a specialized Use Case or clarification question, then the Use Case plans API calls and may emit structured action tags that the chatbot executes. For Super Agent, business templates flow through ETL into Neo4j; consultant requests route to graph procedures or hybrid RAG, inject CRM context, execute APIs step by step, and return source-backed next steps to the human consultant.

Why this architecture

The team chose LangChain/LangGraph because telecom support maps naturally to graph-based decisions, while Neo4j made procedural relationships first-class for consultant guidance. The split between Super TOBi and Super Agent keeps autonomous customer interactions separate from human-in-the-loop consultant support, and LangSmith observability prevents production agents from becoming opaque black boxes.

Models used: LLMs · LangSmith Evaluators SDK · task-specific LLM agents · ReAct agents

Scale

• Super TOBi serves nearly 9.5 million customers through the Customer Companion App and voice channels.
  “Currently serving nearly 9.5 million customers through the Customer Companion App and voice channels”article, stated
• Super TOBi reports a 90% correctness rate.
  “90% correctness rate, 82% resolution rate, and a Customer Effort Score (CES) of 5.2 out of 7”article, stated
• Super TOBi reports an 82% resolution rate.
  “90% correctness rate, 82% resolution rate, and a Customer Effort Score (CES) of 5.2 out of 7”article, stated
• Super TOBi has a Customer Effort Score of 5.2 out of 7.
  “90% correctness rate, 82% resolution rate, and a Customer Effort Score (CES) of 5.2 out of 7”article, stated
• Super Agent drove One-Call Resolution above 86%.
  “the use of generative AI has driven One-Call Resolution (OCR) rates above 86%.”article, stated
• Daily automated evaluation runs overnight on LangSmith traces.
  “The team has developed sophisticated evaluation processes that run daily, automatically classifying chatbot responses”article, stated

Alternatives considered (roads not taken)

Traditional TOBi chatbot behavior

Rejected because: Traditional TOBi struggled with nuanced requests needing contextual understanding, multiple systems, and end-to-end resolution.

Bot speaks directly to customers for every support case

Rejected because: Super Agent is deliberately internal; it augments consultants with next steps rather than replacing them in every interaction.

Flat documents for troubleshooting knowledge

Rejected because: Neo4j stores steps, conditions, actions, APIs, documents, and relationships as first-class graph artifacts.

Black-box production agents without deep observability

Rejected because: LangSmith traces, latency metrics, evaluation scores, and daily evaluation make the workflows operationally inspectable.

Patterns

multi-agent-orchestration (customer-service-routing)

Super TOBi uses a Supervisor and Use Case agents implemented as LangGraph graphs for customer interactions.

tool-use-orchestration (procedure-execution)

Super Agent retrieves graph procedure nodes, executes required APIs, checks conditions, and iterates until a solution is found.

rag-retrieval (knowledge-base-answering)

Open-ended questions use hybrid RAG that combines a vector store with the Neo4j knowledge graph and source citations.

structured-output-contract (procedure-authoring)

Business specialists write procedures in structured templates, which LLM agents parse into JSON and knowledge-graph steps/conditions/actions.

observability-tracing (production-agent-monitoring)

LangSmith streams traces, latency metrics, and evaluation scores into dashboards for continuous optimization.

eval-driven-development (daily-agent-quality)

The team runs daily automated evaluations over LangSmith traces and uses human oversight to maintain correctness targets.

human-in-the-loop-gate (consultant-assist)

Super Agent does not speak to customers; it gives human consultants diagnostics, compliant guidance, and source-backed explanations.

Tradeoffs

How to support human call-center consultants.

Chose an internal AI copilot that recommends next steps over letting a bot speak directly to customers for every support case.

Rationale: Consultants retain the customer relationship while AI speeds diagnostics, procedure lookup, and compliant guidance.

How to represent troubleshooting knowledge.

Chose Neo4j graph procedures with first-class relationships over flat documents or unstructured knowledge-base passages alone.

Rationale: Graph representation supports conditions, actions, API links, validation rules, traversals, and source-grounded procedure execution.

Gotchas (what bit them)

Traditional TOBi struggled with nuanced requests.

Root cause: Telecom support often requires contextual understanding, multi-system access, and end-to-end resolution.

Resolution: Fastweb + Vodafone transformed TOBi into an agentic system with graph-based planning and use-case agents.

Unstructured consultant questions need grounding and policy compliance.

Root cause: Pure vector retrieval may find broad passages but not the right procedural context or company-policy constraints.

Resolution: The hybrid chain anchors vector-store retrieval in the Neo4j knowledge graph and adds source citations.

Production agent systems become black boxes without deep observability.

Root cause: Graph workflows reason, route, and act across tools and subgraphs, making failures hard to diagnose from outputs alone.

Resolution: LangSmith provides end-to-end trace visibility and daily evaluation.

Evaluation

Fastweb + Vodafone evaluates with correctness, resolution, customer-effort, one-call-resolution metrics, LangSmith traces, latency metrics, evaluation scores, overnight automated evaluation, and human business-stakeholder review.

Metrics: correctness rate · resolution rate · Customer Effort Score · One-Call Resolution · latency · evaluation scores · daily performance metrics

Outcomes

• Super TOBi serves nearly 9.5 million customers.
• Super TOBi achieved a 90% correctness rate and 82% resolution rate.
• Super TOBi achieved a Customer Effort Score of 5.2 out of 7.
• Super Agent drove One-Call Resolution rates above 86%.
• Updated knowledge graphs can be promoted to production within hours without downtime.

The systems support Customer Companion App, voice channels, customer-facing chatbot flows, and consultant-assist workflows for a major Italian telecommunications customer base.