📞
Real-world scenario
Wanjiku at TelcoMax Nairobi: "Can we stop 23% of customers leaving?"

Wanjiku is a data scientist at TelcoMax, a telecom operator in Nairobi serving 5,000 enterprise customers. This morning, the CEO walks into her office with a problem: every quarter, 23% of customers churn—switching to Safaricom, Airtel, or going out of business. Lost revenue is climbing.

"Can data science help?" the CEO asks.

Wanjiku could jump straight to Python and build a model. But first, she needs a map. The CRISP-DM framework is that map. It guides her through 6 phases—from understanding the business problem to shipping a churn-prediction system that the sales team can use to save customers before they leave.

5,000
Customers
23%
Quarterly churn
13
Data features

What is CRISP-DM?

CRISP-DM stands for Cross-Industry Standard Process for Data Mining. It's a structured methodology, not a tool, that breaks a vague business problem into 6 actionable phases. Think of it as a blueprint for any data science project, whether you're predicting churn, detecting fraud, or recommending products.

Unlike waterfall software projects, CRISP-DM is iterative. You move forward through phases, but you loop back when new information demands it. Build a model, learn something unexpected about the data, go back to Data Preparation and try again. This flexibility is what makes it work in the real world.

The core idea: Good data science is 80% problem framing and data understanding, 10% modeling, 10% deployment. Most beginners flip this ratio and fail.

The 6 CRISP-DM phases

1. Business Understanding

Before touching data, understand the problem. What is the business goal? How will success be measured?

TelcoMax: Reduce churn from 23% to 15%. Success metric: identify 80% of at-risk customers before they leave.

2. Data Understanding

Collect and explore available data. What columns exist? Is the data clean? What patterns jump out?

TelcoMax: Load 5,000 customer records. 13 columns: tenure, monthly_charges, contract_type, internet_service. Check for missing values (found 150 nulls).

3. Data Preparation

Clean and transform data into a format ready for modeling. Handle missing values, encode categories, scale numbers.

TelcoMax: Fill missing payment info with median charges. Convert contract_type (text) to numbers. Scale numeric features so that tenure and charges are on the same 0–1 scale.

4. Modeling

Select and train algorithms. Compare multiple models (logistic regression, XGBoost, neural nets) to find the best fit.

TelcoMax: Train logistic regression (simple, fast) and XGBoost (more complex, captures non-linear patterns). Which has better precision on at-risk customers?

5. Evaluation

Does the model actually solve the business problem? Measure accuracy, precision, recall, and business impact.

TelcoMax: Model has 92% accuracy overall, but only 68% precision on churners. We catch 680 true churners but falsely flag 320 loyal customers. Is that good enough for the retention team?

6. Deployment

Ship the model so the business can use it. Could be an API, a dashboard, batch predictions, or a recommendation system.

TelcoMax: Deploy as a REST API. Sales team uploads a customer list each week. API returns churn probability for each. Team targets the top 200 with retention offers.

CRISP-DM is iterative, not linear

The key insight: projects rarely move straight from Deployment back to the beginning. Real work loops backward.

The CRISP-DM cycle showing feedback loops between phases
Business Data Prep Modeling Evaluation Deployment Data Understand Data Science Project

Example feedback loop: After evaluating your model, you discover it misses 40% of churners. You loop back to Data Preparation—maybe tenure is being scaled incorrectly. Fix it, retrain, evaluate again.

The data pipeline within each project

Inside the CRISP-DM phases, individual projects follow a technical pipeline. Here's what a churn-prediction project looks like:

From raw data to deployed model
Raw CSV Explore Clean Features Split Train Evaluate Deploy

Each box is a step that takes hours or days of work. Each arrow is a point where you might discover something and loop backward.

Types of analytics in CRISP-DM

During the Business Understanding and Evaluation phases, you'll work with these 4 types of analysis:

Descriptive Analytics

"What happened?" Report past facts without prediction. Example: TelcoMax had 23% churn last quarter.

Diagnostic Analytics

"Why did it happen?" Drill into root causes. Example: Month-to-month customers churn at 33%, yearly at 11%.

Predictive Analytics

"What will happen?" Forecast future outcomes. Example: This customer has 68% churn probability this month.

Prescriptive Analytics

"What should we do?" Recommend actions. Example: Offer this customer a 2-year contract discount to reduce churn risk.

Where CRISP-DM shows up across industries

📞
Telecom: Customer churn
Predict which subscribers will switch networks. Deploy to sales team weekly.
🏦
Banking: Fraud detection
Identify suspicious transactions in real-time. Models must be deployed as low-latency APIs.
🛡️
Insurance: Claims prediction
Estimate claim likelihood and cost. Deployment feeds risk assessment dashboards.
🛒
E-commerce: Recommendations
Predict which products a user will buy. Deployed as APIs that power website recommendations in real-time.

How AI systems follow CRISP-DM

When ChatGPT answers your question, it followed this exact lifecycle. Engineers defined the business goal (helpful, safe responses). They collected and understood text data (billions of conversations). They prepared data (filtered toxic content, tokenized). They trained models (transformer neural networks). They evaluated (does it reduce harmful outputs?). Finally, they deployed as an API. Every AI system, whether a recommender, a detector, or a language model, follows the same framework.

Data Science Pipeline ML Engineering Model Evaluation Production Deployment
🎯 Mini Challenge: Map a project to CRISP-DM phases

Scenario: You work at SafeGuard Insurance in Bangalore. Your manager says: "We lose money on false claims. Can you predict which claims are fraudulent?"

Task: Classify each task below into a CRISP-DM phase:

(a) Meet with claims investigators to define what "fraudulent" means and set a success target
(b) Pull 10,000 claim records from the warehouse (dates, amounts, injury types, policyholder history)
(c) Handle missing injury descriptions; encode claim type (auto vs. health); scale claim amounts
(d) Train logistic regression and XGBoost; compare precision and recall on held-out claims
(e) Build a Flask API. Underwriters upload daily claims; API flags top 50 for manual review

Answers: (a) Business Understanding, (b) Data Understanding, (c) Data Preparation, (d) Modeling & Evaluation, (e) Deployment

Interview questions: test your understanding

Cross-Industry Standard Process for Data Mining. It's a framework for organizing data science projects, industry-agnostic, and iterative.
1. Business Understanding, 2. Data Understanding, 3. Data Preparation, 4. Modeling, 5. Evaluation, 6. Deployment.
Because you discover new information at every phase. A bad model may reveal that your features are poorly scaled—loop back to Data Preparation. A failed deployment might require rethinking the business goal. Real projects are messy, not straight lines.
Descriptive answers "What happened?"—past facts (23% churn last quarter). Predictive answers "What will happen?"—future estimates (this customer has 68% churn risk). Both matter: descriptive informs the baseline; predictive is what your model produces.
Business Understanding or Evaluation. High accuracy doesn't guarantee business value. Maybe the model catches all non-churners (high accuracy) but misses 90% of actual churners (useless for retention). Or the business goal was never clearly defined, so accuracy is the wrong metric. You should have aligned on success metrics in phase 1 and validated in phase 5.

Quiz: 5 questions to test your knowledge

Question 1 of 5
How many phases are in the CRISP-DM lifecycle?
Question 2 of 5
Which CRISP-DM phase comes first?
Question 3 of 5
"What will happen next?" is which type of analytics?
Question 4 of 5
Is CRISP-DM linear or iterative?
Question 5 of 5
Cleaning missing values and scaling features belongs to which CRISP-DM phase?
0
/5
Excellent! You understand the CRISP-DM framework.

Cheatsheet: CRISP-DM phases at a glance

Phase Key Question TelcoMax Example
Business Understanding What problem are we solving? Reduce churn from 23% to 15%. Success: identify 80% of at-risk customers.
Data Understanding What data exists? Is it clean? 5,000 customers, 13 features (tenure, charges, contract). 150 missing values.
Data Preparation Is data model-ready? Fill missing values (median). Encode contract_type. Scale numeric features to [0, 1].
Modeling Which algorithm fits best? Train logistic regression and XGBoost. Compare precision, recall, AUC on test set.
Evaluation Does it solve the business goal? XGBoost: 92% accuracy, 68% precision on churners. Catch ~680 true churners, flag 320 false positives.
Deployment How do users consume it? REST API. Sales uploads weekly customer list. API returns churn probability. Team targets top 200.

Common mistakes: what goes wrong

Mistake 1: Skipping Business Understanding
Jumping straight to modeling without defining success metrics. You build a 95% accurate model, then the business won't use it because you optimized the wrong thing. Spend a week on phase 1—align on the goal, agree on metrics.
Mistake 2: Confusing accuracy with business value
A churn model with 95% accuracy but 30% precision on churners misses 70% of at-risk customers. Useless. The business goal was to catch 80% of churners, not to achieve overall accuracy. Use the right metric.
Mistake 3: Treating CRISP-DM as waterfall
Assuming you move forward in order without looping back. Real projects loop constantly. Poor evaluation means going back to Data Prep. Unexpected patterns in exploration mean rethinking the business framing. Embrace iteration.
Mistake 4: Ignoring deployment from the start
Building a model in a Jupyter notebook that never leaves your laptop. The business can't use it. Think about deployment at phase 1: Will it be an API? A batch job? A dashboard? Design with deployment in mind.

Key takeaways

  • CRISP-DM is a 6-phase framework for organizing data science projects from problem to deployment. It's industry-agnostic and proven across telecom, banking, insurance, e-commerce.
  • Phases are iterative, not linear. You loop backward when new information demands it. A failed model evaluation sends you back to Data Preparation to fix your features.
  • Business Understanding is foundational. Before touching data, align on the business goal and success metric. Most failures happen because teams optimize the wrong objective.
  • Data Preparation is where value accrues. 80% of project time goes here—handling missing values, encoding categories, scaling features. Garbage in, garbage out.
  • Deployment is the goal. A model in a notebook doesn't help the business. Design for deployment from phase 1. Think API, dashboard, or batch job.
  • Success metrics matter more than model accuracy. A 95% accurate churn model that misses 70% of churners is useless. Define the right metric in phase 1, validate in phase 5.
Africa & India's Best AI Academy
🤖 AI-first 🎓 Placement Assistance

Ready to build real data science projects?

Join the Best AI Program: The AI-Era Analyst — 5 months, fully online, with placement assistance. Week 1 is free.

Explore The AI-Era Analyst →