How GenAI changes a data analyst's day, task by task
In this article (6 sections)
Related programme: the Data Analytics with Generative AI course (3–4 months; ₹90,000 (inclusive of GST) · EMI from ₹4,000/month).
Generative AI changes a data analyst's day unevenly. It takes over some drafting and formatting work, speeds up a larger group of tasks where you still check every output, and leaves the judgement-heavy work — defining metrics, choosing what to investigate, and signing off numbers — firmly with you. The useful question is not "will AI do my job" but "which of my recurring tasks moves into which bucket".
Start from your own week, not a generic job description
Before reading the table below, write down what you actually did last week in 30-minute blocks. Most analysts find their week is a mix of five or six recurring activities: pulling data, cleaning it, building or updating a report, answering ad-hoc questions, explaining numbers in meetings, and documenting. The table uses that mix. Your own list will differ; the reasoning method is what transfers.
The three buckets used here:
- Automate — the model produces an output you can use after a light check, because errors are cheap to spot.
- Assist — the model gives you a draft or options, but you must verify substance, because an error would be hard to see and costly.
- Still human — the task depends on context, accountability or relationships the model does not have.
The task table
| Task | Bucket | Why |
|---|---|---|
| Writing a first-draft SQL query from a clear spec | Assist | Syntax is usually fine; the risk is a wrong join or grain that returns plausible numbers. You must read it. |
| Converting a query between dialects (for example, T-SQL date functions to PostgreSQL) | Automate | Mechanical translation with an easy test: run both and compare row counts and totals. |
| Explaining an inherited query you did not write | Assist | Good for a first reading; the model can misstate intent when column names are ambiguous. |
| Cleaning inconsistent city or product names | Assist | It proposes mappings quickly, but "Noida Ext" versus "Greater Noida West" is a business decision, not a spelling fix. |
| Writing regular expressions for a known pattern | Automate | Easy to test against a handful of sample strings before use. |
| Choosing which metric answers the stakeholder's question | Still human | Requires knowing why the question was asked and what decision it feeds. |
| Defining a metric (what counts as an "active customer") | Still human | This is an agreement between people; a model can list options but cannot own the definition. |
| Building a dashboard layout | Assist | Useful for suggesting chart types; the priority order depends on who reads it and what they act on. |
| Writing DAX or Excel formulas | Assist | Filter context and absolute references are exactly where plausible-looking errors hide. |
| Summarising a 40-message email thread before a meeting | Automate | Low stakes; you were on the thread and will notice a wrong summary. |
| Writing the narrative for a monthly business review | Assist | A draft saves time, but models tend to invent causes for movements. You decide what is claimed. |
| Explaining a number to a sceptical manager | Still human | Trust comes from you knowing the data lineage and admitting limits. |
| Spotting that a number "looks wrong" | Still human | This is pattern memory about your business; a model has no baseline for your data. |
| Documenting a report's logic and sources | Automate | A draft from the query and your notes is reliable enough to edit. |
| Deciding whether to share customer-level data with a tool | Still human | A privacy and policy decision with legal consequences. |
Where the time actually moves
Notice that the automate bucket is mostly work that was already small: translation, regex, summaries, documentation drafts. The assist bucket is where most of your hours sit, and there the gain is real but conditional — it depends on whether checking the output is faster than producing it yourself. A SQL draft you can verify in two minutes is a gain; a narrative paragraph that sends you back to the data to confirm each claim may not be.
A practical rule: if you cannot describe how you would detect an error in the model's output, the task is not in the automate bucket, however good the output looks.
The new tasks that appear
GenAI also adds work to an analyst's day that did not exist before:
- 1Writing context for the model — table descriptions, metric definitions and example questions. This is documentation work that pays off twice, because humans read it too.
- 2Verifying generated outputs — reconciling an AI-drafted total against a trusted report, checking that a stated trend exists in the data.
- 3Deciding what data may leave your environment — checking your company's approved-tool list before pasting anything.
- 4Answering "can AI just do this?" from managers, with a reasoned yes, no or partly.
That fourth task is a career opportunity. The analyst who can explain which tasks moved and why becomes the person others ask, which is covered in becoming the AI person on your team.
A worked example: one ad-hoc request
A regional manager asks: "Why did repeat orders fall in the east region last month?"
- You ask a model to draft a query for repeat-order rate by region and month. It writes one using
COUNT(DISTINCT customer_id)over all orders. Assist — you notice it counts customers, not repeat customers, and fix the logic. - You ask it to list possible causes. It offers pricing changes, stock-outs, a competitor promotion and seasonality. Assist — useful as a checklist, but none of these is evidence.
- You check stock-out data yourself and find two SKUs unavailable for eleven days. Still human — you knew where that data lived and that it mattered.
- You ask the model to tighten your three-paragraph explanation. Automate — the facts are yours; it only edits wording.
The model saved you drafting time in three places. The finding came from you.
What to practise next
If you want to test yourself, take one real report you own and classify each step using the three buckets, writing the reason in one line. Where you wrote "assist", note how you would verify the output. That list is also a strong interview answer when someone asks how you use AI at work.
For the tasks that should stay away from a model entirely, read analyst tasks you should not hand to an LLM. The Data Analyst programme teaches SQL, Power BI and GenAI together so that verifying a generated query is a habit rather than an afterthought.
Frequently asked questions
Which data analyst tasks can generative AI do on its own?
Very few end to end. Drafting boilerplate SQL, reformatting a table, writing a first pass of documentation and summarising a long thread are the closest. Even these need a quick check, because the model does not know your table grain or your business definitions unless you give them.
Does GenAI make SQL skills less important for analysts?
It makes reading and checking SQL more important. If a model writes a query, you are the person who has to notice a wrong join or a missing filter, and you cannot do that without understanding the query yourself.
Pankit Kumar has 10 years in Data Science & AI, building and shipping production systems in regulated pharma and clinical environments. He is a freelance trainer at Boston Institute of Analytics, AnalytixLabs and Scaler, and has taught this material to thousands of working professionals.
This article is part of our Data Analytics with Generative AI programme — 3–4 months. The full analyst stack — Excel, SQL, Power BI and Python pipelines — then a generative-AI layer you can prove is right.
Explore Data Analytics with Generative AI