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.
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 6 CRISP-DM phases
1. Business Understanding
Before touching data, understand the problem. What is the business goal? How will success be measured?
2. Data Understanding
Collect and explore available data. What columns exist? Is the data clean? What patterns jump out?
3. Data Preparation
Clean and transform data into a format ready for modeling. Handle missing values, encode categories, scale numbers.
4. Modeling
Select and train algorithms. Compare multiple models (logistic regression, XGBoost, neural nets) to find the best fit.
5. Evaluation
Does the model actually solve the business problem? Measure accuracy, precision, recall, and business impact.
6. Deployment
Ship the model so the business can use it. Could be an API, a dashboard, batch predictions, or a recommendation system.
CRISP-DM is iterative, not linear
The key insight: projects rarely move straight from Deployment back to the beginning. Real work loops backward.
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:
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
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.
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
Quiz: 5 questions to test your knowledge
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
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.
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 →