Predicting Bank Customer Churn with Agentic AI: What 10,000 Customer Records Taught Us About Retention

Customer churn is one of the most expensive problems in retail banking. Acquiring a new customer costs far more than keeping an existing one, which makes spotting at-risk customers early a top strategic priority. We analyzed 10,000 bank customer records using the Terno Agentic AI platform — connected directly to a MySQL database hosted on Aiven Cloud Data Services — to find out exactly which customers are leaving, and why.

The headline finding: roughly one in five customers (20.37%) has already exited the bank, and churn rates swing dramatically depending on geography, age, and account activity. That means retention budgets spent evenly across the whole customer base are likely wasted — the money should be concentrated on a few specific, identifiable segments.

Here's everything we found, and what to do about it.

Key Takeaways Up Front

• The dataset shows an imbalanced 80/20 split between retained and churned customers, which has real consequences for how prediction models need to be built.

• Age and member activity status are the two strongest predictors of churn.

• Germany churns at roughly double the rate of France and Spain.

• Machine learning classifiers hit up to 86.65% accuracy and 0.8612 ROC-AUC at predicting who will leave.

• All charts below were generated interactively inside Terno's agentic AI environment, querying live data with no manual exports.

Data & Methodology

The Dataset

We worked with the Bankdata customer churn dataset, sitting in a MySQL relational database and accessed through Terno AI's live database connector. The data infrastructure runs on Aiven Cloud Data Services, giving us managed, secure, scalable MySQL hosting. Terno AI connected directly to this cloud database, running real-time SQL queries with no manual exports and no pipeline lag.

The dataset covers:

• Sample size: 10,000 validated customer records from the churn_data table

• Demographics: gender, age, surname, geography (France, Germany, Spain)

• Financial attributes: credit score, account balance, estimated salary, number of products held

• Behavioral indicators: tenure, active member status, credit card ownership

• Target variable: binary churn outcome (Exited: 1 = churned, 0 = retained)

Technology stack: The churn_data table lives in a MySQL instance provisioned through Aiven Cloud Data Services — a fully managed open-source data platform with automated backups, high availability, and enterprise-grade security. Terno Agentic AI connected to this database via its native database integration layer, running every analytical query directly against the live MySQL instance (Dataset ID: 677). No pipeline delays, no stale data — every result reflects the current operational database.

Our Five-Phase Analytical Framework

1. Initial Exploration — row/column counts, data type verification, schema review, sample inspection
2. Data Quality Assessment — missing value checks, duplicate detection, unusual value identification across all fourteen columns
3. Univariate & Segmentation Analysis — churn distribution, age profiling, balance comparison, geographic breakdown, behavioral factor evaluation
4. Correlation Analysis — a numerical feature correlation heatmap highlighting variables most related to churn
5. Predictive Modeling — development and comparison of Logistic Regression, Random Forest, and XGBoost classifiers

Consistent train/test splits and encoding strategies were used throughout to keep model comparisons fair.

Data Preprocessing

Before modeling, we:

• Removed non-predictive identifiers (RowNumber, CustomerId, Surname)

• Converted types — Balance and EstimatedSalary parsed from object strings to numeric format

• Checked for missing data — zero missing values across all 14 columns; no imputation needed

• Encoded categories — one-hot encoding of Geography and Gender with first-level dropping to avoid multicollinearity

• Scaled features — StandardScaler normalization applied to numeric columns for Logistic Regression training

• Split the dataset — a stratified 80/20 partition yielding 8,000 training observations and 2,000 validation cases

Customer Demographics & Profile

Overall Churn Distribution

The class distribution is meaningfully imbalanced: most customers stay, but a substantial minority have left.

Operational concern: With 20.37% of customers churned, the bank faces a material retention challenge. That imbalance also has direct modeling implications — standard classifiers trained without adjustment will inherently bias toward predicting retention, letting at-risk customers slip through undetected.

Status	Count	Percentage
Not Churned	7,963	79.63%
Churned	2,037	20.37%
Total	10,000	100%

Gender: Not a Differentiator

Gender-based churn analysis shows male and female customers exhibit similar overall churn tendencies. Neither gender stands out as a dominant at-risk segment — retention strategies don't need to be gender-differentiated, and should instead focus on behavioral and financial attributes.

The Age Factor

Age is one of the most discriminating variables in the dataset. Churned customers average 44.84 years old, versus 37.41 years for those who stay — a gap of over seven years. Churned customers cluster in the 40–60 age band, while the retained majority peaks at 30–40.

Retention priority: Customers aged 40 and above warrant enhanced engagement programs. This cohort likely has accumulated assets elsewhere, has greater financial literacy for comparing offerings, and may be less price-sensitive but more quality-sensitive — demanding relationship-grade service that standard retail banking products often fail to deliver.

Churn Status	Average Age (Years)
Not Churned	37.41
Churned	44.84

Financial Patterns & Segmentation

Account Balance: A Counterintuitive Signal

Balance analysis challenges conventional assumptions about high-value customer loyalty. Churned customers show a wider balance range with a higher median — elevated account value does not confer loyalty.

• Churned customers: average balance $91,108.54 | median $109,349 | Q3 $131,435 | max $250,898

• Retained customers: average balance $72,745.30

Strategic implication: High-balance customers likely have superior financial sophistication and awareness of competing offers. They may perceive insufficient return on deposits, inadequate personalized service, or better yield opportunities elsewhere. Retention incentives — preferential interest rates, dedicated relationship managers, premium product access — should be concentrated on this high-value segment before exit signals appear.

Churn Status	Average Balance (USD)
Not Churned	72,745.30
Churned	91,108.54

Geographic Segmentation: The Germany Anomaly

Geographic breakdown uncovered a stark regional disparity. Germany's churn rate of 32.44% is roughly double that of France (16.15%) and Spain (16.67%).

Geography	Total Customers	Churned	Churn Rate
Germany	2,509	814	32.44%
Spain	2,477	413	16.67%
France	5,014	810	16.15%

This regional concentration suggests either product-market fit issues specific to Germany, competitive pressure from local banking alternatives, or service delivery gaps affecting German customers disproportionately.

Regional intervention required: Germany should be treated as a priority market for retention programs. Root cause investigation — through customer exit surveys, competitive benchmarking, and product utilization analysis — should precede any broad marketing spend in this region.

Behavioral Factors: Activity Status and Credit Cards

Two behavioral dimensions were tested for their independent relationship with churn — and the results diverge sharply.

Activity Status	Total Customers	Churned	Churn Rate
Inactive Members	4,849	1,302	26.85%
Active Members	5,151	735	14.27%

Card Status	Total Customers	Churned	Churn Rate
No Credit Card	2,945	613	20.81%
Has Credit Card	7,055	1,424	20.18%

Two clear conclusions emerge:

1. Member activity is critical. Inactive members churn at 26.85% versus 14.27% for active members — a 12.6 percentage point gap. Engagement and product utilization are powerful retention levers.
2. Credit card ownership is negligible. A 0.63-point difference in churn rates between cardholders and non-cardholders shows that credit card products alone don't meaningfully bind customers to the institution. Bundled product depth, not card issuance, likely determines loyalty.

Strategic Drivers: Root Cause Analysis

The "Engagement-First" Reality

Conventional banking wisdom emphasizes product breadth and rate competitiveness as the primary retention mechanisms. Our correlation and feature importance analysis challenges that assumption, surfacing behavioral engagement and demographic age as the dominant churn predictors instead.

Key correlation findings:

1. Age (r ≈ 0.285, positive): The strongest numerical predictor of churn. Older customers churn at higher rates regardless of balance size or product holdings — underscoring the importance of life-stage-aligned banking products.
2. IsActiveMember (r ≈ −0.156, negative): Member engagement negatively correlates with churn — customers who actively use accounts, transact regularly, or interact with bank services are substantially less likely to exit. This relationship is causal and directly actionable.
3. Balance (r ≈ 0.119, moderate positive): Higher balances weakly but consistently associate with increased exit probability, reinforcing the high-value customer vulnerability finding from segmentation.
4. NumOfProducts (r ≈ −0.048, non-linear pattern): Customers with two products show the lowest churn rates, while those with three or more products show elevated exit behavior — suggesting over-selling or product mismatch drives dissatisfaction rather than loyalty.
5. CreditScore, EstimatedSalary, Tenure: Weaker or negligible standalone correlations with churn, though they contribute meaningfully within ensemble models.

Investment reorientation: Rate competition and product proliferation may deliver less retention improvement per dollar than engagement programs, life-stage advisory services, and proactive outreach for inactive account holders. This finding carries profound budget implications for retail banking strategy.

Predictive Modeling: The Machine Learning Approach

Beyond understanding current churn drivers, we built computational models to forecast individual customer exit probability — enabling preemptive retention intervention before customers disengage for good.

The Classification Problem

We framed churn forecasting as binary classification:

• Input variables (X): credit score, geography, gender, age, tenure, balance, number of products, credit card status, activity status, estimated salary

• Output variable (y): binary churn state (Exited = 1 or 0)

• Learning goal: derive a function f: X → y that generalizes to previously unseen customers

Three Models, Three Approaches

1. Logistic Regression — a linear probability framework offering coefficient interpretability and a baseline benchmark
2. Random Forest — a tree ensemble using bootstrap aggregation to minimize variance while capturing non-linear feature interactions
3. XGBoost — sequential gradient-boosted trees that iteratively correct prior model errors for superior pattern recognition

Baseline Model: Logistic Regression

• Overall accuracy: 81.15%

• Precision (churned): 0.5563

• Recall (churned): 0.2010

• F1 score (churned): 0.2953

• ROC-AUC: ≈ 0.70

Classification matrix:

Actual / Predicted	Not Churned	Churned
Not Churned	1,544	63
Churned	314	79

Though it delivers solid baseline accuracy with transparent coefficients, the logistic model's low recall of 20.10% for churned customers is operationally problematic — it misclassifies the majority of actual churners as retained. This gap reflects both class imbalance and non-linear relationships that linear models can't fully capture.

Advanced Model: Random Forest

Random forest builds numerous decision trees independently, then aggregates their predictions through majority voting. The feature importance chart below closely mirrors our earlier correlation findings.

• Overall accuracy: 86.65%

• Precision (churned): 0.7582

• Recall (churned): 0.4707

• F1 score (churned): 0.5808

• ROC-AUC: ≈ 0.8500

The jump in recall — from 20.10% to 47.07% — and F1 score — from 0.2953 to 0.5808 — shows Random Forest's clear edge at identifying at-risk customers, which is the core operational goal of any churn prediction system.

State-of-the-Art Model: XGBoost

XGBoost represents current best practice in gradient boosting, offering computational efficiency alongside superior predictive performance.

• Overall accuracy: 86.55%

• ROC-AUC: ≈ 0.8612 — the strongest discrimination score across all models tested

That 0.8612 ROC-AUC shows a strong ability to distinguish churners from non-churners across all classification thresholds, making XGBoost particularly suitable for threshold-tuning applications where recall can be prioritized over precision depending on intervention cost economics.

Model Comparison & Selection

Algorithm	Accuracy	Precision	Recall	F1 Score	ROC-AUC
Logistic Regression	81.15%	0.5563	0.2010	0.2953	≈ 0.70
Random Forest	86.65%	0.7582	0.4707	0.5808	≈ 0.85
XGBoost	86.55%	—	—	—	≈ 0.8612

Production recommendation: XGBoost demonstrates the strongest ROC-AUC (0.8612), indicating superior class discrimination capability. For operational deployment, XGBoost combined with probability threshold tuning — rather than a default 0.5 cutoff — will optimize the precision/recall trade-off based on the bank's retention intervention cost structure.

Feature Importance Summary

The Random Forest feature importance chart (Figure 8) corroborates all prior analytical findings, producing a consistent ranked ordering:

1. Age (≈ 0.23): Highest importance by a substantial margin — the primary churn driver across all analytical methods
2. EstimatedSalary (≈ 0.15): Second-ranking position in the ML importance hierarchy
3. CreditScore (≈ 0.15): Near-equal to EstimatedSalary — reflects financial health and alternative credit access as churn proxies
4. Balance (≈ 0.14): Matches the segmentation hierarchy, reinforcing the high-value customer vulnerability finding
5. NumOfProducts (≈ 0.13): Significant predictor, particularly through the non-linear churn relationship at higher product counts
6. Tenure (≈ 0.08): Moderate contribution capturing relationship longevity effects
7. IsActiveMember (≈ 0.04): Despite a strong segmentation signal, partially absorbed by correlated features in the ensemble context
8. Geography_Germany, Gender_Male, HasCrCard (< 0.03): Lowest tier, though Geography_Germany's presence aligns with the documented regional disparity

This convergence between traditional statistical correlation and machine learning importance metrics strengthens confidence in the strategic recommendations below, since it's consistent across independent analytical methods.

Recommendations

Drawing from both correlation analysis and machine learning feature rankings, we recommend focused intervention in high-leverage customer segments rather than diffuse retention spending across the entire customer base.

1. Age-Segmented Engagement Programs

Justification: As both the strongest correlation (r ≈ 0.285) and the highest ML importance score (≈ 0.23), age-based segmentation is the single most impactful targeting criterion for churn prevention.

Implementation steps:

• Cohort definition: flag customers aged 40+ as elevated-risk and auto-enroll them in proactive relationship management workflows

• Product realignment:

  • Develop life-stage-appropriate products (wealth management, retirement planning, estate services) for mid-career and pre-retirement priorities
  • Assign dedicated relationship managers for customers aged 45+ with balances exceeding $50,000
  • Launch advisory touchpoints (quarterly reviews, financial health checks) that reinforce relationship value

• Success metric: target a 15% reduction in churn rate among the 40+ cohort within 12 months of launch

Projected impact: Even a 5-percentage-point reduction in the 40+ segment's churn rate could prevent hundreds of high-balance exits, preserving significant deposit volume and associated revenue.

2. Member Activation Campaign

Justification: Inactive members churn at nearly double the rate of active members (26.85% vs 14.27%). Converting inactive accounts into engaged ones is the highest-leverage behavioral intervention available.

Implementation steps:

• Inactivity detection: define operational triggers for accounts with no transactions in 60-, 90-, or 180-day windows

• Re-engagement sequences:

  • Personalized outreach (email, push, SMS) highlighting underutilized product features
  • Limited-period incentives (bonus interest rates, cashback, fee waivers) tied to activity milestones
  • Route persistently inactive high-balance accounts to direct relationship manager outreach

• Freemium activation architecture: offer complimentary premium service trials (priority customer service, premium card upgrades) to lapsed accounts, converting feature exposure into habitual engagement

Projected impact: Re-activating 20% of currently inactive at-risk customers could reduce overall churn by 2–3 percentage points, translating directly into retained deposit balances and cross-sell revenue.

3. Germany-Specific Retention Program

Justification: Germany's 32.44% churn rate — double that of France and Spain — demands targeted regional investigation and intervention independent of the global retention strategy.

Implementation steps:

• Exit survey deployment: mandate structured exit interviews for German customer closures to identify primary departure reasons

• Competitive intelligence: map the German banking landscape for recently launched competing products, rate improvements, or digital-native entrants attracting the churning cohort

• Localized product development: consider Germany-specific innovations (higher-yield savings instruments, digitally-native account formats, local partnership integrations)

• Churn prediction threshold adjustment: lower the ML model's intervention threshold for German customers from 0.7 to 0.5, triggering proactive outreach earlier in the exit journey

Projected impact: Reducing Germany's churn rate to match Spain's (16.67%) would retain approximately 395 additional customers annually — a material improvement given the elevated average balances observed among churned customers.

4. Predictive Intervention System

Justification: With XGBoost achieving 0.8612 ROC-AUC, the model can identify at-risk customers weeks before actual exit, creating intervention windows that reactive retention strategies can't access.

Implementation steps:

• Real-time scoring pipeline: integrate the model into core banking CRM systems for weekly churn probability refresh, querying live data from the Aiven-hosted MySQL instance via Terno's database connector

• Tiered intervention triggers:

  • Probability 0.5–0.7: automated digital outreach (personalized email, in-app notification) with product recommendations
  • Probability 0.7–0.9: relationship manager assignment with a defined contact obligation within 5 business days
  • Probability above 0.9: senior advisor escalation with discretionary retention offer authority (rate enhancements, product upgrades, fee reversals)

• Model maintenance: quarterly retraining with fresh customer data from the live Aiven MySQL database, keeping accuracy current as behavioral patterns evolve

• Feedback loop: capture intervention outcomes (retained, churned despite intervention, declined contact) to continuously refine model training and intervention strategy

Projected impact: Proactive intervention on high-probability churners could prevent 15–25% of forecasted exits from materializing, preserving deposit volumes and relationship revenue that reactive approaches would otherwise lose entirely.

Conclusion

This analysis shows how modern agentic AI capabilities — connected directly to live cloud-hosted databases — turn operational banking data into strategic intelligence for customer retention. By linking Terno AI to a MySQL database hosted on Aiven Cloud Data Services, the whole workflow required no manual data exports, no pipeline engineering, and no batch processing delays: insights came straight from the source of truth.

Combining traditional statistical techniques, interactive visual analytics, and advanced machine learning surfaced five key findings:

1. Churn driver identification: Age and member activity status are the dominant predictors, challenging the assumption that financial product breadth or rate competitiveness are the primary retention levers.
2. Geographic disparity: Germany's churn rate of 32.44% is nearly double that of France and Spain, demanding a market-specific strategic response rather than a uniform global approach.
3. High-value customer vulnerability: Customers with higher account balances churn at greater rates — a counterintuitive risk that conventional loyalty assumptions overlook.
4. Forecasting capability: We achieved 0.8612 ROC-AUC prediction accuracy, enabling proactive rather than reactive retention strategies.
5. Evidence-based prioritization: Quantitative analysis supports specific investment priorities — age-segmented engagement, activation campaigns, and regional programs — that maximize churn prevention relative to capital deployed.

Retail banking today sits at the intersection of relationship depth and digital convenience. This work gives institutions a quantitative framework for directing retention resources toward genuinely high-risk segments, anticipating customer exits through predictive intervention, and converting operational data into a sustainable competitive advantage in deposit retention.

Note: The dataset shows a 79.63% / 20.37% class imbalance. Teams extending this work are advised to apply SMOTE oversampling or scale_pos_weight class-weight adjustment in future optimization cycles to further improve recall on the minority churned class.

References & Methodology

This investigation employed contemporary agentic AI workflows ensuring analytical transparency and result reproducibility. Complete methodological details and source materials appear below:

1. Terno.ai — Agentic AI Platform for Data Science Applications. Available at: https://terno.ai
2. Complete Analysis Documentation: Full conversation history and computational steps recorded on Terno AI. https://manoj3.app.terno.ai/chat/share/9467979e-9444-42b1-aea9-b1a041d1697e?ui_version=v2
3. Data Infrastructure: The churn_data table is stored in a MySQL relational database provisioned through Aiven Cloud Data Services (https://aiven.io) — a fully managed open-source data platform providing automated backups, high availability, end-to-end encryption, and multi-cloud deployment. Terno Agentic AI connected to this Aiven-managed MySQL instance (Dataset ID: 677) using its native database integration layer, enabling live SQL query execution without manual data extraction or pipeline configuration. The dataset was uploaded directly to MySQL and made available to Terno through this cloud-hosted connection.
4. Computational Environment: Python 3.9+ employing scikit-learn (Logistic Regression, Random Forest, evaluation metrics), xgboost (gradient boosting), pandas (data manipulation), and matplotlib/plotly (interactive visualization generation). All charts and visualizations were generated natively within Terno's agentic AI environment and exported as HTML and PNG outputs.
5. Dataset Source: Bank Customer Churn Dataset, churn_data table (10,000 validated customer records), uploaded to MySQL and hosted via Aiven Cloud Data Services. Dataset originally sourced from publicly available churn modeling benchmarks commonly referenced in financial analytics literature.
6. Validation Approach: Standard machine learning evaluation employing Accuracy, Precision, Recall, F1-Score, ROC-AUC metrics, and Confusion Matrix analysis. Stratified 80/20 train-test split guarded against overfitting while ensuring robust performance estimation across both majority and minority classes.
7. Class Imbalance Handling: Imbalanced class distribution (79.63% retained, 20.37% churned) noted throughout. Modeling teams are advised to apply SMOTE oversampling or scale_pos_weight class-weight adjustment in subsequent optimization cycles to further improve recall on the minority churned class.

See the complete analysis and conversation history on Terno AI: https://manoj3.app.terno.ai/chat/share/9467979e-9444-42b1-aea9-b1a041d1697e?ui_version=v2

*Read the full whitepaper*