A multi-agent AI system for proactive customer churn detection and Next Best Action (NBA) recommendations.

AI Proficient Capstone Project — Q2 2026
Author: Anantakumaar Raviprasath

The Customer Retention Intelligence Agent is a multi-agent AI solution designed to proactively identify customers who may be at risk of churn and recommend an appropriate Next Best Action (NBA) before the customer leaves.

The system combines:

• LangGraph for multi-agent orchestration
• ChromaDB + Retrieval-Augmented Generation (RAG) for retention playbook retrieval
• Azure OpenAI GPT-4o for final response generation
• Deterministic business rules for risk-segment classification
• CLV-tier-based personalization for differentiated retention actions

The architecture is intentionally designed so that critical customer-risk classification is deterministic and auditable, while generative AI is used for the final narrative and recommendation output.

Customer churn can be difficult to detect because disengagement often happens gradually. Traditional CRM workflows may identify a customer only after the customer has already decided to leave.

Common challenges addressed by this project include:

• Silent churn: declining app usage and email engagement without an explicit complaint
• Reactive detection: identifying risk only after significant disengagement
• Inconsistent treatment: different retention actions across agents, regions, or shifts
• Limited scalability: manually assessing large customer populations

The proposed system acts as a proactive intelligence layer that can identify at-risk customers, retrieve the relevant retention playbook, and generate a structured NBA recommendation.

The system uses a Centralized Supervisor Pattern in LangGraph.

A Supervisor Agent coordinates three specialized agents in a mandatory sequence:

1. Customer Profiler

   • Retrieves the customer's behavioral profile.
   • Uses signals including recency, app sessions, email engagement, NPS, support tickets, and customer lifetime value (CLV).

2. Churn Diagnostician

   • Applies deterministic, priority-ordered business rules.
   • Classifies the customer into one of four predefined risk segments.

3. Retention Strategist

   • Uses RAG with ChromaDB to retrieve the most relevant retention playbook.
   • Adapts the recommended treatment based on the customer's risk segment and CLV tier.

4. Response Generator

   • Uses Azure OpenAI GPT-4o to generate the final structured retention response.
   • The LLM does not determine the risk segment or the underlying retention action.

                         ┌─────────────────────────┐
                         │    Supervisor Agent     │
                         │       LangGraph         │
                         │  Orchestrates Workflow  │
                         └────────────┬────────────┘
                                      │
             ┌────────────────────────┼────────────────────────┐
             ▼                        ▼                        ▼
┌─────────────────────┐   ┌─────────────────────┐   ┌──────────────────────┐
│ Customer Profiler   │   │ Churn Diagnostician  │   │ Retention Strategist │
│                     │   │                     │   │                      │
│ fetch_customer_     │   │ classify_risk_      │   │ recommend_retention_ │
│ record              │   │ segment             │   │ action               │
└──────────┬──────────┘   └──────────┬──────────┘   └──────────┬───────────┘
           │                         │                         │
           │                         │                    ChromaDB RAG
           │                         │                         │
           └─────────────────────────┼─────────────────────────┘
                                     ▼
                         ┌─────────────────────────┐
                         │   Response Generator   │
                         │    Azure OpenAI GPT-4o │
                         └────────────┬────────────┘
                                      ▼
                         ┌─────────────────────────┐
                         │    Next Best Action     │
                         │  Segment + Decision +   │
                         │  Rationale + Treatment │
                         └─────────────────────────┘

Risk classification is intentionally handled through deterministic rules rather than an LLM. This provides consistency, transparency, and auditability.

Retention decisions require repeatability and governance. With deterministic rules, the same customer signals produce the same risk classification, making the decision easier to validate and audit.

Retention playbooks are stored separately from the application logic.

The RAG workflow is:

Risk Segment
     │
     ▼
Embedding Query
(text-embedding-ada-002)
     │
     ▼
ChromaDB
Semantic Search
     │
     ▼
Top Matching Playbook
     │
     ▼
CLV-Tier Adaptation
     │
     ▼
NBA Report

Using RAG instead of hardcoded playbook lookups allows:

• Retention playbooks to be updated without changing application code
• Semantic retrieval instead of brittle string matching
• Business teams to maintain the knowledge base independently
• New or rephrased segment descriptions to be handled more flexibly

The LLM is used for narrative generation; it does not decide the customer's risk segment or underlying retention action.

For a high-risk customer, the system:

1. Retrieves the customer profile.
2. Evaluates the predefined risk rules.
3. Identifies the highest-priority matching risk segment.
4. Retrieves the corresponding retention playbook through ChromaDB.
5. Applies CLV-tier-specific treatment.
6. Generates a structured NBA report.

The final response contains:

• Customer risk segment
• Next Best Action
• Reasoning/rationale
• Recommended retention treatment
• Engagement policy

The solution was evaluated using two complementary approaches.

Agent outputs were compared directly against expert reference answers.

GPT-4o evaluated five dimensions for each validation customer:

Key finding: Treatment Completeness was the only dimension below 5.0. The evaluation found no wrong risk segments, incorrect engagement policies, or structural errors across the validation set.

The proposed solution can be operationalized as a nightly automated pre-triage layer within an existing CRM workflow.

• Create a high-priority retention case
• Route the customer to a retention specialist
• Initiate personalized outreach within 24 hours
• Avoid promotional content until the service issue is resolved

• Launch targeted win-back engagement
• Provide loyalty incentives and personalized recommendations
• Monitor engagement over the win-back period
• Close the win-back window after prolonged inactivity

• Continue standard engagement
• Avoid unnecessary retention spending
• Consider referral or advocacy programs for highly satisfied customers
• Continue passive monitoring and re-classify if risk signals change

A recommended GitHub repository structure is:

customer-retention-intelligence-agent/
│
├── README.md
│
├── code/
│   └── customer_retention_intelligence_agent.ipynb
│
├── data/
│   ├── validation_customer_records.json
│   ├── test_customer_records.json
│   ├── risk_segments.json
│   └── retention_playbooks.json
│
├── results/
│   └── submission.csv
│
├── documentation/
│   ├── business-brief.pdf
│   └── capstone-presentation.pdf
│
├── requirements.txt
├── .gitignore
└── .env.example

Important: API credentials and environment files containing secrets should never be committed to GitHub.

The original notebook installs the following packages:

pip install langchain==0.3.25
pip install langchain-core
pip install langchain-community
pip install langchain-openai
pip install langchain-text-splitters
pip install langchain-chroma==0.2.4
pip install python-dotenv
pip install langgraph==0.3.31
pip install pandas
pip install chromadb

The notebook loads credentials from:

retention_agent.env

The environment configuration includes values for:

MODEL_ENDPOINT
CHAT_MODEL_NAME
AZURE_OPENAI_API_KEY
api_version
EMBEDDING_MODEL_NAME
MODEL_ENDPOINT_EMBEDDING
api_version_embedding

Do not upload the actual .env / environment file or API key to GitHub.

Use a local environment file or another secure secret-management mechanism when running the notebook.

The notebook expects the following input files:

validation_customer_records.json
test_customer_records.json
risk_segments.json
retention_playbooks.json

The customer data is used to create a unified customer pool for the Customer Profiler. Risk segment definitions are loaded as priority-ordered deterministic rules, while retention playbooks are transformed into documents and indexed in ChromaDB.

The system generates structured NBA responses for customers and exports the final test results to:

submission.csv

The submission contains:

customer_id
agent_response

The current solution has several limitations identified during the project:

• Risk thresholds are based on predefined business rules and may require periodic review as customer behavior changes.
• Additional customer attributes could improve personalization but are not currently available in the dataset.
• Customers who do not clearly match an existing risk segment may require manual review.
• For enterprise-scale deployment, the local vector database should be replaced with a managed production-grade vector search platform.
• The current implementation is a capstone/prototype and would require additional production engineering, governance, security, monitoring, and integration work before enterprise deployment.

The project demonstrates how business rules, retrieval-based knowledge, and generative AI can work together rather than relying on an LLM to make every decision.

The architecture prioritizes:

Auditability → Consistency → Explainability → Personalization → Scalability

The central principle is:

AI in CRM does not have to be a black box. Every decision should be traceable to the signals that triggered it, the playbook it came from, and the action owner responsible for the next step.

The repository can include:

• 📘 Business Brief
• 📊 Capstone Presentation
• 💻 Jupyter Notebook
• 📄 Sample/approved project data
• 📈 Evaluation results
• 🧾 Submission output

Anantakumaar Raviprasath

AI Proficient Training Program — Q2 2026
Dentsu Global Services × Analytics Vidhya

This project was developed as an AI Proficient Capstone Project for learning and demonstration purposes. It is not presented as a production-ready customer-retention system.

Any production implementation should use appropriately governed, anonymized or authorized data and secure handling of credentials and business-sensitive information.