What Is Data Science? A Decision Discipline, Explained

What is data science? Put simply, it is the work of using data, statistics and code to reduce uncertainty about a decision someone has to make. In practice that means framing a question, finding and cleaning the data, analysing it, checking whether the result holds up, and explaining it to the person who will act.

Most definitions stop at the Venn diagram of statistics, programming and domain knowledge, which is technically true and tells you nothing about the job. The decision-first framing explains what is otherwise confusing: why so much of the work is unglamorous data cleaning, why communication is a graded interview skill, and why the fanciest model frequently loses to a simple one.

This overview is for students, career changers and managers who want to know what the work involves before investing in it. It covers the official duty list, a project from question to memo with a worked churn example, how data science differs from its neighbours, the skills in priority order, and where to start.

What is data science, according to official job data?

Official job data describes a role built mostly around data, judgement and communication, and it is more concrete than most textbook definitions. The US Bureau of Labor Statistics lists five core duties for data scientists:

• deciding which data are available and useful for a project, before any analysis starts
• collecting, categorising and analysing data
• creating, validating, testing and updating algorithms and models
• presenting findings with data visualisation software
• making business recommendations to stakeholders

Only one of the five is about models. The rest are about data, judgement and communication, and that balance also matches what the job feels like from the inside.

What does a data science project look like?

A realistic project runs roughly like this.

1. Someone asks a badly formed question. "Why is churn up?" or "Should we open in Pune?" The first real task is turning that into something answerable, which usually means going back and negotiating what would count as an answer.
2. You find out what data exists. There is almost always less than you hoped, in a worse state than you were told, with a definition that changed in 2023 and nobody documented.
3. You clean it. On most projects this takes more time than any other step, because real data is messy. It is also where the important discoveries happen: the reason the numbers look strange is frequently the answer to the question.
4. You analyse. Sometimes that means a model, but not always. More often it means a well-constructed comparison, a cohort breakdown or a properly designed experiment. The sophistication should match the question, not your résumé.
5. You check whether you believe it. Is the effect real or noise? What else could explain it? What would change your mind? This step, above all, separates data science from making charts.
6. You communicate. A memo, a deck, a conversation. If the decision-maker cannot understand it, act on it or disagree with it for good reasons, the analysis might as well not have happened.

That balance is especially uncomfortable for people who came for the modelling. Getting and cleaning data usually takes longer than the analysis, and communication takes longer than most syllabi admit.

A worked example: the churn question

Take "Why is churn up?" at a subscription business. First, you agree what churn means — cancelled within the month, or simply not renewed? Next, you break cancellations down by signup month and find that the spike comes almost entirely from customers who joined during a discount campaign. Then you check whether customers from earlier campaigns churned faster too.

Suppose they did. In that case the answer is not "the product got worse"; instead, it is "discount cohorts churn faster, and we just ran a big one". No machine learning was involved, yet the finding changes next quarter's marketing plan. That, in miniature, is the answer to what is data science: a better decision, backed by evidence someone can check.

How is data science different from analytics, statistics and machine learning?

The neighbouring fields overlap heavily, so the clearest difference is the question each one starts from.

The labels blur in real job ads. For example, O*NET, the US Department of Labor's occupation database, files business intelligence analysts as a specialism under the data scientist occupation. So read the responsibilities rather than the title. The most common confusion gets its own guide: data analyst vs data scientist.

Why do organisations pay for data science?

There are three reasons, in rough order of how much they are worth.

Decisions under uncertainty. Should we change the price? Is the new onboarding flow better? These carry real money, yet whoever argues most confidently often decides them. A good data scientist replaces confidence with evidence.

Scale. A human can review a hundred loan applications, whereas a model can review a hundred thousand, consistently, with a measurable error rate. Consistency is often the bigger win.

Discovery. Sometimes the data contains something nobody suspected — a segment behaving differently, or a process quietly failing. This is the least predictable value, but occasionally the largest.

Official figures reflect that demand. As of September 2026, the latest BLS data put the median data scientist salary at $120,230 (May 2025) and project employment to grow 35% from 2025 to 2035, much faster than the average for all occupations. The largest employers include publishing and broadcasting, insurance carriers and computer systems design firms, so the work is not confined to tech companies.

Which data science skills matter most?

Here they are, honestly ordered.

SQL matters more than anything else on this list, and weak SQL is the most common reason capable candidates fail early interview rounds.

Statistics — distributions, inference and experiment design — comes next, because the ability to say whether a result is real is the core of the discipline.

Programming in Python, mostly for pandas and modelling, needs to be competent rather than brilliant. Both SQL and Python are mainstream well beyond data work: in the 2025 Stack Overflow Developer Survey, 58.6% of respondents had done extensive work in SQL over the past year and 57.9% in Python.

Communication is graded explicitly in every interview process worth joining. Analyses that nobody acts on have no value, whatever their technical quality.

Domain knowledge is underrated and hard to teach. Knowing which numbers are unreliable and what a plausible effect size looks like in your industry takes years — which is why career switchers with deep industry backgrounds often outperform fresh graduates.

Has AI made data science obsolete?

No. The tooling has genuinely improved: AutoML fits models, notebooks can draft code, and a competent assistant can produce a passable exploratory analysis.

What has not changed is that someone must decide which question is worth asking, whether the data can answer it, whether the answer is trustworthy, and what the organisation should do about it. Those are judgement tasks, and they were always the hardest parts anyway.

The realistic effect is a raised floor. For example, work that consisted of producing routine reports is shrinking. Meanwhile, work that consists of framing problems and defending conclusions is in more demand, because there is now far more output that needs someone to evaluate it.

Where should you start?

If this sounds like work you would enjoy, the sequence that works is:

1. SQL first.
2. Statistics alongside it, rather than after.
3. Python after that.
4. Modelling last, once the foundations hold.

Our data scientist roadmap then lays that out as a stage-gated path with exit criteria. If you are not sure this is the right field, AI vs machine learning vs deep learning explains how the adjacent fields fit together.

Who should think twice? If you dislike ambiguity, or you would rather build systems than argue for decisions, then software or machine learning engineering may suit you better than data science.

Related reading

Tools and techniques for data science covers the practical stack, such as SQL, pandas and scikit-learn. The role of programming in data science addresses how much code you actually need.