Tencent Cloud AI Retention Blueprint
Menu
Use caseWhy TI-ONECloud stackModel logicTry demo
Checking model service

TI-ONE customer churn solution

Predict customer churn.
Prioritize who to retain.

Customer data goes in. Tencent Cloud TI-ONE returns a real-time churn risk score. Retention teams use that score to prioritize customers for human review and follow-up.

Built on TI-ONERisk score APIHuman-reviewed action

LIVE TI-ONE SERVICE

Managed inference endpoint

customer-churn-service

Checking
SingaporePublic gatewayAuthenticated
Platform
Tencent Cloud TI-ONE
Model state
Checking model
Prediction API
/v1/models/m:predict
Output
Churn probability

The live demo below sends each prediction request to this managed TI-ONE service through a secure proxy.

01

Business challenge

Retention teams cannot review every account.

02

AI signal

Prioritize customers with a churn probability.

03

Business action

Route high-risk records to a human review queue.

Why TI-ONE is the center

Move from notebook
to managed model service.

TI-ONE covers the AI lifecycle around the model: development, training, model management, deployment, and online service operations.

DEVELOP

Interactive model development

Use TI-ONE development machines with Jupyter Notebook or VS Code for data preparation, debugging, and model training.

DEPLOY

Managed online service

Deploy a model with a built-in or custom runtime instead of maintaining a standalone inference VM and process manager.

OPERATE

Service-level controls

Configure authentication, throttling, health checks, logging, replicas, and manual or automatic scaling from the service layer.

INTEGRATE

Cloud-native connections

Connect storage, container images, permissions, monitoring, and logs through the broader Tencent Cloud service portfolio.

Verified capabilities: based on current Tencent Cloud TI-ONE documentation. No unverified performance uplift, cost reduction, or ROI claim is made in this prototype.

One solution, connected services

Sell the workflow—not an isolated model.

Each cloud service has a clear architectural role. TI-ONE remains the inference and model-operations core.

Target cloud architectureCurrent prototype uses a local secure proxy in place of COS + EdgeOne.

Customer request path

WEB EXPERIENCE

COS website

Hosts HTML, CSS, and JavaScript

SECURE API ENTRY

EdgeOne Function

Protects the token and proxies requests

MANAGED INFERENCE

TI-ONE Online Service

Runs the authenticated prediction API

Core
RUNTIMETCRImage
MODELCFSPackage
OPERATIONSCAM + CLSAccess + logs
BUSINESS OUTPUT

Churn risk score

Feeds review or CRM workflows

How it connects: COS serves the page, EdgeOne Function securely forwards the request, and TI-ONE runs inference. TCR supplies the runtime image; CFS supplies the model package; CAM and CLS support access control and operations.

How the current model separates risk

What low and high risk look like
in this synthetic dataset.

These profiles summarize holdout records grouped by model score. The model evaluates all fields and their interactions; it does not apply the values below as fixed business rules.

LOW< 30%Normal monitoring
WATCH30–50%Watch for changes
ELEVATED50–70%Consider review
HIGH≥ 70%Prioritize review

Observed median profile

Lower-risk group

Day minutes
2.75
Evening minutes
3.95
Night calls
300
Customer service calls
5
Voicemail messages
200

Holdout records scored below 30% had a 2.36% observed churn rate.

MODELED SIGNAL→

Observed median profile

Higher-risk group

Day minutes
7.66
Evening minutes
6.01
Night calls
200
Customer service calls
6
Voicemail messages
0

Holdout records scored at least 70% had a 96.77% observed churn rate.

Global model sensitivity

Fields the model relied on most

Permutation importance on the holdout set ranks Night Calls, Evening Minutes, Day Minutes, and Customer Service Calls as the strongest global signals.

  1. 01Night calls
  2. 02Evening minutes
  3. 03Day minutes
  4. 04Customer service calls
Interpret carefully: these are descriptive patterns in a synthetic dataset, not causal telecom rules. A high score means the feature combination resembles high-risk training records. Field-level reasons for one individual prediction require an explanation layer such as SHAP.

How to use this prototype

From profile
to review signal.

The interface demonstrates model scoring, not an automated retention decision.

  1. 01
    Choose a starting point

    Load a validated low-risk or high-risk sample, or enter a profile manually.

  2. 02
    Describe the customer

    Provide account, plan, usage, and customer-support information.

  3. 03
    Run the model

    Select Analyze churn risk. The web app sends one JSON record to the inference API.

  4. 04
    Interpret the signal

    Compare the probability with the 50% prototype threshold. A higher score means stronger modeled risk.

  5. 05
    Apply business judgment

    Validate the signal against current context and policy before outreach or an offer.

Important: an 82% score is a model output above the prototype threshold. It does not mean the customer is guaranteed to leave.

Live model input

Customer data

Choose a profile for a one-click prediction, or edit any field below.

All profiles come from the model’s holdout dataset. Select a profile to fill the form, review the inputs, then choose Analyze churn risk to run the model.

01 Account
02 Plans
03 Usage
PeriodMinutesCalls
Day
Evening
Night
International
04 Service history