How to learn data analysis in 2026, when AI writes the SQL

“Why learn SQL if AI will write the query?” It’s a fair question. The answer changes what you should learn, not whether.

An AI assistant can turn “show me revenue by category” into a correct query. Without the relevant context, it may miss that your finance team excludes refunds, that last week’s data is incomplete, or that marketing means something else by “active customer”. The part of analysis that remains human is the question, the context and the responsibility for the answer. That is what a good learning path should train.

Short version: learn to read tables and write basic SQL (a few weeks of practice), then spend most of your time on defining metrics and checking numbers on realistic business questions. Use AI from the start, but as a colleague whose work you check.

What still matters when AI writes the query

1. The question. What does the business actually want to know, and which decision depends on it? “How many customers do we have?” has several right answers depending on who asks — we built a three-question game around exactly this.
2. The definition. Which rows, which status, which period. An unstated definition can make a correct calculation answer the wrong question.
3. The check. Is the number plausible? Does it agree with a total you trust? Did a join multiply rows? See the 7-question checklist.
4. The decision. What do you recommend, how sure are you, and what would change your mind?

SQL is still worth learning — not to type faster than an AI, but because it is the language in which you check. Practise reading and changing queries until you can explain which rows they count, including whether cancelled orders belong in the result.

For a worked example of the skills behind that judgement, see Data literacy for AI: a correct growth calculation can still lead to an unsupported budget recommendation.

A learning path from zero

Stage 1: tables and your first queries (1–2 weeks)

What a table, a row and a column are; SELECT, WHERE, ORDER BY, LIMIT; then GROUP BY with COUNT, SUM and AVG. Practise on data that looks like a real business, not on toy tables of pets. Aim for 20–30 minutes a day.

Stage 2: joins, CTEs and window functions (1–2 weeks)

JOIN and LEFT JOIN, and above all the traps: joins that multiply rows and joins that silently drop them. Then WITH for step-by-step queries and window functions for rankings and running totals. Our guide to JOIN duplicates covers a trap that can inflate totals.

Stage 3: metrics and business questions (ongoing)

Conversion, retention, average order value, GMV, unit economics, A/B tests. For each metric learn the definition and its failure modes: the wrong denominator, the incomplete period, the average that hides the spread.

Stage 4: work-like practice

Practise the whole loop: an unclear request from a colleague → a definition → a query → a check → a short written recommendation. This exercises the connection between technical work and a decision. Look for opportunities to practise that connection alongside syntax.

Where to practise

• Free SQL tutorials such as SQLBolt, SQLZoo or the Mode SQL tutorial are good for syntax.
• Public datasets (Kaggle and similar) are good for exploration, but they rarely come with a stakeholder, a deadline or a check of your answer.
• Atterna was built for the parts in between: the Atterna Start course takes you from zero to joins and window functions in ten short days on one marketplace dataset (days 1–2 without an account, all ten free with an account), and the Meridian simulator (part of Atterna Analyst) gives you analyst work with colleagues, deterministic checks on two data extracts and, in every episode, a decision backed by your numbers (episode 1 ends in a memo). Not sure where to begin? The level check has six questions.

How to use AI while you learn

• Write the query yourself first, then ask the AI for its version and compare.
• Ask the AI to explain a query line by line in plain words.
• Ask it to find the mistake in your query — then verify its claim.
• Never paste a number into a report without the seven checks.

FAQ

Is it still worth learning SQL in 2026?

Yes, but as a language of reading and checking rather than of typing. You need to recognise when a query counts the wrong rows, and writing, reading and testing queries helps build that skill.

How long does it take to learn data analysis from scratch?

Basic SQL takes about two to four weeks of regular practice. Being useful on real business questions takes a few months of practice on realistic tasks, because the hard part is defining metrics and checking results, not syntax.

Can I become a data analyst without a degree?

Many analysts don’t have a degree in the field. Many employers look for evidence that you can take an unclear question to a checked answer. A trace of real work — how you defined the metric, what you checked, what you recommended — is stronger evidence than a certificate alone.

What is the best free way to start?

Start with a task that shows why analysis matters: the free three-question Wrong Number game needs no SQL and no account. Then take the level check or day 1 of Atterna Start.