	In Browser
	StumbleUpon
	del.icio.us
	Google
	Google Buzz
	reddit
	LinkedIn

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Google Gemini API > Customer Churn Prediction Agent

Customer Churn Prediction Agent

Author: Venkata Sudhakar

Customer churn costs ShopMax India revenue and acquisition budget. An ADK churn prediction agent scores customers on purchase recency, frequency, and spend (RFM), identifies those at high churn risk, and generates personalised retention offers - all in one automated workflow. Running this daily on the active customer base lets the retention team focus their outreach on the highest-risk customers.

The agent uses three tools: one to fetch RFM scores from BigQuery, one to compute the churn risk label, and one to generate a personalised retention offer. The model decides what offer to make based on the customer segment and their purchase history.

from google.adk.agents import LlmAgent
from google.adk.tools import FunctionTool
from google.adk.sessions import InMemorySessionService
from google.adk.runners import Runner
from google.genai.types import Content, Part

# Simulated customer RFM data (in production: BigQuery query)
CUSTOMERS = {
    "C001": {"name": "Priya Sharma",  "city": "Mumbai",    "days_since_last_order": 75,  "orders_last_90d": 0, "total_spend_rs": 145000},
    "C002": {"name": "Amit Verma",    "city": "Delhi",     "days_since_last_order": 12,  "orders_last_90d": 4, "total_spend_rs": 62000},
    "C003": {"name": "Kavya Reddy",   "city": "Hyderabad", "days_since_last_order": 120, "orders_last_90d": 0, "total_spend_rs": 28000},
}

def get_customer_rfm(customer_id: str) -> dict:
    """Get RFM (recency, frequency, monetary) data for a ShopMax India customer."""
    c = CUSTOMERS.get(customer_id)
    if not c:
        return {"error": "Customer not found"}
    return {"customer_id": customer_id, **c}

def compute_churn_risk(days_since_last_order: int, orders_last_90d: int, total_spend_rs: int) -> dict:
    """Compute churn risk score for a ShopMax India customer based on RFM signals."""
    risk_score = 0
    if days_since_last_order > 90:
        risk_score += 3
    elif days_since_last_order > 45:
        risk_score += 2
    elif days_since_last_order > 20:
        risk_score += 1

if orders_last_90d == 0:
        risk_score += 3
    elif orders_last_90d <= 1:
        risk_score += 1

if total_spend_rs > 100000:
        risk_score -= 1  # high-value customers churn less

if risk_score >= 5:
        label = "HIGH"
    elif risk_score >= 3:
        label = "MEDIUM"
    else:
        label = "LOW"
    return {"risk_score": risk_score, "risk_label": label}

def generate_retention_offer(customer_name: str, risk_label: str, total_spend_rs: int, city: str) -> dict:
    """Generate a personalised retention offer for an at-risk ShopMax India customer."""
    if risk_label == "HIGH" and total_spend_rs > 100000:
        offer = f"Rs 500 cashback on your next order + free express delivery in {city}"
        channel = "Personal call from account manager"
    elif risk_label == "HIGH":
        offer = "10% off your next order with code COMEBACK10"
        channel = "SMS + Email"
    elif risk_label == "MEDIUM":
        offer = "Free delivery on your next 2 orders"
        channel = "Push notification"
    else:
        offer = "No intervention needed"
        channel = "None"
    return {"customer": customer_name, "offer": offer, "channel": channel}

agent = LlmAgent(
    name="shopmax_churn_agent",
    model="gemini-2.0-flash",
    tools=[
        FunctionTool(get_customer_rfm),
        FunctionTool(compute_churn_risk),
        FunctionTool(generate_retention_offer),
    ],
    instruction=(
        "You are a ShopMax India customer retention analyst. "
        "For each customer: fetch RFM data, compute churn risk, and generate a retention offer. "
        "Summarise the action plan clearly."
    )
)

session_service = InMemorySessionService()
runner = Runner(agent=agent, session_service=session_service, app_name="shopmax")

for cid in ["C001", "C002", "C003"]:
    session = session_service.create_session(app_name="shopmax", user_id="analyst")
    events = list(runner.run(
        user_id="analyst",
        session_id=session.id,
        new_message=Content(parts=[Part(text=f"Analyse churn risk for customer {cid} and recommend action.")])
    ))
    for event in events:
        if hasattr(event, "content") and event.content:
            for part in event.content.parts:
                if hasattr(part, "text") and part.text:
                    print(f"[{cid}] {part.text.strip()[:200]}")
    print()

It gives the following output,

[C001] Priya Sharma (Mumbai) - HIGH churn risk (score: 5)
Last order: 75 days ago, 0 orders in last 90 days, lifetime spend Rs 1,45,000.
Action: Personal call from account manager with Rs 500 cashback offer
and free express delivery in Mumbai.

[C002] Amit Verma (Delhi) - LOW churn risk (score: 1)
Last order: 12 days ago, 4 orders in last 90 days, lifetime spend Rs 62,000.
Action: No intervention needed - customer is active.

[C003] Kavya Reddy (Hyderabad) - HIGH churn risk (score: 6)
Last order: 120 days ago, 0 orders in last 90 days, lifetime spend Rs 28,000.
Action: SMS + Email with 10% off next order code COMEBACK10.

The agent correctly identified Priya as a high-value churning customer deserving a premium retention approach (personal call + cashback), while Kavya gets a standard discount code. Amit is active and needs no intervention. This three-tier approach maximises retention ROI by reserving expensive interventions for high-value customers.

For ShopMax India production, run this agent as a daily Cloud Run job that reads all customers with RFM data from BigQuery, processes them in batches of 100, and writes the recommended actions to a retention_actions Firestore collection. The CRM system reads this collection to schedule SMS, email, and call centre tasks. Track which offers were redeemed using coupon code analytics in BigQuery to measure the agent's retention effectiveness over time.

Send your comments, suggestions or queries regarding this site to [email protected].