The Data Science Workflow: From Question to Value

From Vague to Clear

Data science rarely starts as a clean math problem. It often begins with a vague question like "Are our users happy?" or "How can we make more money?" The first job is to turn this into a clear, answerable problem.

Three Levels of a Question

1. Business question: big picture, e.g., "How can we reduce customer churn?" 2. Analytical question: how data can help, e.g., "Can we predict who will churn?" 3. Data question: what we actually do, e.g., "Can we train a weekly churn model?"

Work With Stakeholders

Data scientists work with business stakeholders to define success, constraints, and metrics. Ask: What counts as a win? What limits do we have (budget, time, privacy)? Which metric will we try to move?

Common Failure Modes

Projects fail when questions are too vague ("Use AI to improve the app"), when there is no clear success metric, or when stakeholders do not agree on the goal from the start.

Quick Checklist

For a good problem definition, answer: Who is this for? What decision will this help? What metric will we improve, and roughly by how much? By when do we need a first result?

Why Stakeholders Matter

A data science project only creates value if it helps someone make a better decision or power a useful product. So you must understand who cares about the result and how they will use it.

Typical Stakeholders

Stakeholders include business leaders, product managers, engineers, legal/privacy teams, and end users. Each group has different goals: revenue, UX, reliability, compliance, or usability.

Key Early Questions

Ask: Who will use the result? How often? What decision will change because of it? These answers shape everything else: accuracy needs, speed, and allowed data.

Example: Churn Prediction

In a churn project, marketing wants a list to target, finance wants better revenue forecasts, and engineering must integrate the model. Understanding this alignment keeps the project useful.

Data Discovery

After defining the problem, you ask: What data do we have, and what data do we need? This is the data discovery or data collection stage of the workflow.

Where Data Comes From

You explore internal databases, logs, and transactions, request access from data engineers, collect new data (like surveys), or use external datasets and APIs, respecting licenses and privacy rules.

Check Relevance and Legality

For each source, ask: Is it relevant? Is it available for the users and time we care about? Is it legal and ethical to use, for example under GDPR and similar privacy laws?

Example Sources for Churn

For churn, you might use sign-up and cancel dates, subscription plans, app usage logs, support tickets, and payment failures. Together they help explain why users leave.

Common Data Problems

You may find that key data does not exist yet, is hard to access, poorly documented, or biased. Here you often work closely with data engineers and IT/security teams.

Exploring a Dataset

Here is a small Python example using pandas to explore a churn-like dataset. Focus less on syntax and more on the questions we ask about the data.

Key Questions in Code

We ask: What columns do we have? How many rows? Any missing values? What are basic summaries? How does sessions per week differ between churned and active users?

Why This Matters

These quick checks are part of data discovery and initial exploration. They help you see structure, quality, and rough patterns before deeper modeling.

Messy to Usable Data

Real-world data is messy. The longest stage is often cleaning, exploring, and turning raw data into useful features that models and humans can work with.

Data Cleaning

Cleaning means fixing invalid values, handling missing data, and standardizing formats like dates and categories so they are consistent and usable.

Exploratory Data Analysis

EDA uses plots and summaries to understand distributions and relationships, and to check for issues like data leakage or strange outliers.

Feature Engineering

Feature engineering turns raw logs into meaningful variables, like sessionsperweek, failedpaymentslast3months, or hasrecentcomplaint for a churn model.

Collaboration and Impact

Data scientists work with domain experts and data engineers here. Clean data and good features often matter more than using the most complex model.

When Modeling Happens

Modeling usually comes after you define the problem and prepare the data. You often start with simple baseline models before trying anything complex.

Evaluating Models

You pick metrics that match the business goal, use train/validation/test splits, and check performance across user groups to avoid overfitting and unfair bias.

Iteration, Not a Line

If results are poor, you often loop back: improve features, fix data, or even adjust the problem definition. Data science is a cycle, not a strict one-way pipeline.

Stakeholders in This Stage

Data scientists lead modeling; product managers judge if metric gains matter; legal and ethics teams may review fairness and compliance under modern AI regulations.

From Model to Real Use

A model or analysis only creates value when it is used. Deployment means putting it into production systems or dashboards so people can act on it.

Monitoring in Practice

After deployment, teams monitor technical health (errors, speed), model health (performance over time), and business impact (are key metrics improving?).

Feedback Loops

User and stakeholder feedback helps reveal if the model stays useful or causes side effects. Teams retrain or redesign when needed, creating a continuous loop.

Example: Churn in Production

A daily job scores users for churn risk. Marketing uses a dashboard to target offers. Teams then track whether churn rates actually go down after campaigns.

Lifecycle and Governance

By 2026, many sectors must document and monitor important models for compliance and ethics. The focus is on governing models across their full lifecycle.