Analysts spend most of their time on plumbing: finding the right table, writing the query, fixing the join, cleaning the export, reshaping it for the chart, then writing three sentences a VP will read. AI is strong at every step, given the schema and a check on the result.
The tools now live where you work: ChatGPT and Claude run Python on an uploaded file and hand back the chart with the code, Copilot in Excel and Gemini in Sheets write formulas and explain workbooks, Cursor and Claude Code suit dbt models, and Julius is built for talking to a dataset.
The discipline that separates good AI-assisted analysis from embarrassing analysis: make the model write code that computes the answer instead of computing in its head, verify against a number you already know, and never upload data your governance policy says stays inside the warehouse.
Quick wins this week
- Paste your table schema (columns and types, no rows), ask the question in plain English, and request the SQL with a one-line explanation per join. Run it on a date range you can hand-check.
- Upload a messy CSV to ChatGPT or Claude and ask it to profile the file: nulls per column, duplicates, type mismatches, outliers. Read the code it ran before trusting the summary.
- Paste a query someone else wrote and ask for a plain-English explanation plus the assumptions baked into it. You will find a filter nobody remembers adding.
- Give it the chart's numbers and the audience and ask for three sentences: what happened, why it matters, what to do. Cut whichever sentence is speculation.
What AI can do for data analysts, task by task
Writing and debugging SQL
Give the dialect, the schema of the tables involved, a few synthetic rows, and the output you expect. Ask for the query plus the reasoning behind each join and filter. Run it against a period whose totals you know; models invent column names and use functions from the wrong dialect.
Cleaning and transforming data
Ask for a profiling script first (nulls, duplicates, types, outliers), then a cleaning script with every rule as a named step and row counts before and after each. Keep the script in your repo; a cleaned file with no code behind it cannot be reproduced.
Exploratory analysis on an uploaded file
Upload only data your governance policy allows, ideally aggregated or synthetic, and ask for distributions, segments, and hypotheses worth testing. Read the code behind every chart, and treat the 'insights' as hypotheses until checked against the warehouse.
Spreadsheet formulas and models
Use Copilot in Excel or Gemini in Sheets to write and explain formulas, build pivots, and document someone else's workbook. Test any XLOOKUP, LET, or LAMBDA it produces on five rows you can compute by hand, and version the file before letting it restructure anything.
Dashboard measures and semantic layers
Ask for DAX measures, LookML, Tableau calculated fields, or dbt models with tests, and have it explain filter context or grain in plain language. Validate each measure on a handful of rows by hand; an elegant measure at the wrong grain looks fine until finance asks.
Turning results into a narrative and deck
Paste the final numbers, the audience, and the decision they face, and ask for an executive summary, a slide outline, and a chart type per point. Check that every number in the narrative matches your output and delete any claim the data cannot support.
Prompts for data analysts
Replace the bracketed placeholders, paste into any chat assistant, and iterate on the result.
SQL from a schema and a plain-English question
You are a senior analytics engineer writing [SQL DIALECT]. Schema: [PASTE TABLES WITH COLUMN NAMES AND TYPES]. Synthetic sample rows: [PASTE A FEW ROWS]. Question: [THE BUSINESS QUESTION IN PLAIN ENGLISH]. Grain of the result: [ONE ROW PER WHAT]. Write the query with a comment above every join and filter explaining why it is there. List any assumption you made, and which columns I should verify exist before running. Use only columns in the schema.
Tip: Stating the grain up front prevents the most common silent error: double counting.
Profile and clean a CSV, showing the code
You are a data-cleaning assistant. Using the uploaded file [FILE NAME], first profile it: rows, columns, types, nulls per column, duplicates, and obvious outliers. Then propose cleaning rules and wait for my approval. Once approved, apply each rule as a separate named step and report row counts before and after each. Return the complete Python script so I can rerun it. Context: [WHAT THE DATA IS AND HOW IT WILL BE USED]
Tip: The per-step row counts are your audit trail; keep them.
Explain a query and its hidden assumptions
You are reviewing a query written by someone who has left the company. Explain it in plain English, step by step, for an analyst who knows SQL but not this data. Then list every assumption it bakes in: filters, date logic, join types that could drop rows, hard-coded values, and anything that would break if the data changed. Rate each as safe, risky, or unknown. Dialect: [SQL DIALECT] Query: [PASTE THE QUERY]
Tip: Chase every 'unknown' with whoever owns the source table.
Executive narrative from results, no speculation
You are an analyst writing for [AUDIENCE, E.G. THE VP OF SALES]. Results: [PASTE THE KEY NUMBERS OR TABLE]. Decision they face: [THE DECISION]. Write a five-sentence summary: what happened, how it compares with [BASELINE OR PRIOR PERIOD], the most likely explanation supported by the data, what is uncertain, and one recommended next step. Use only the numbers provided, cite each as you use it, and label every inference as an inference.
Tip: If the explanation cites a number you did not paste, it made it up.
Turn a vague request into an analysis spec
You are a senior analyst scoping a request. The stakeholder asked: [PASTE THE REQUEST VERBATIM]. Ask me up to eight clarifying questions covering the metric definition, grain, date range, filters, segments, comparison baseline, and what decision the answer will inform. After I answer, write a one-page spec with precise metric definitions, the tables likely needed, known edge cases, and what 'done' looks like. Flag anything I answered vaguely.
Tip: Send the spec to the stakeholder before writing any SQL.
Want a prompt for something else? Use the Prompt Builder or browse career and job search prompts, data and spreadsheets prompts.
Skills to build
Schema-first prompting
Why: The model does not know your tables; without the schema it invents columns that look right and run wrong.
How: Keep a text file with your core tables, columns, types, and grain (no rows) and paste it at the start of every session, or load it into a Project or custom GPT.
Verification habits
Why: Every generated query or script is a hypothesis until it reproduces a number you already trust.
How: Check totals against a known report, compare row counts before and after each step, and reconcile one segment by hand before sharing anything.
Reading generated code well enough to catch the wrong join
Why: A left join that should be inner, or a filter on the wrong date column, produces a clean answer that is simply wrong.
How: Read every join condition and WHERE clause, and ask the model to explain them; if its explanation surprises you, the code is probably wrong.
Data governance literacy
Why: Uploading the wrong table to the wrong tool is a reportable incident in most companies, and the tool will not warn you.
How: Learn your company's data classification levels, which AI tools are approved for each, and how to build synthetic or aggregated samples that carry the structure without the people.
Statistical judgment the model lacks
Why: Models happily report a correlation as a cause, ignore seasonality, and declare significance on a sample of twelve.
How: Ask it to list the assumptions behind any test it suggests, check the sample size and the baseline yourself, and read the chart before the sentence about it.
Tools worth knowing
ChatGPT
The general-purpose AI assistant most of your coworkers already use.
Claude
A careful writing and analysis assistant that shines on long documents.
Microsoft Copilot
AI inside Word, Excel, Outlook, and Teams, with your company's data protections.
Gemini
Google's assistant, strongest when your work already lives in Google Workspace.
Julius AI
Chat with your spreadsheets and data files and get charts, stats, and answers back
Claude Code
Anthropic's agentic coding tool that works in your terminal and IDE, on your real codebase
Cautions for data analysts
Customer records, health data, payment data, and financial results carry legal duties under GDPR, CCPA, HIPAA, PCI, and SOX-style controls, and consumer AI tools may retain what you upload. Use only tools your data team has approved, prefer aggregated or synthetic samples, and where the warehouse has its own assistant (Snowflake Cortex, Gemini in BigQuery, Databricks Assistant), prefer it: access controls and lineage come along.
Language models miscount, misadd, and confidently produce percentages that do not sum to 100. Never let one compute in its head: ask for code that computes the answer, run it, and compare against a figure you already know.
Without the schema in front of it, a model invents column names, uses Postgres functions in BigQuery, and assumes a date column that does not exist. Paste the schema, run everything, and treat a query that errors as the good outcome; the dangerous one runs and returns something plausible.
The narrative it writes will outrun the data: correlation becomes cause, a small sample becomes a trend. You own every claim in the deck, so label inferences and keep the chat's code in a repo with the data version and date; a result you cannot reproduce next quarter is not a result.
Your 30-day plan
- Week 1: Learn your data classification and which AI tools are approved for each level. Build a schema file (no rows) and load it into a Project or custom GPT.
- Week 1: For one week, have the assistant write every query with explanations, and reconcile each against a known number before use.
- Week 2: Move your cleaning scripts into a repo with the model's code and your edits; rerun one old analysis to confirm it reproduces.
- Week 3: Try Copilot in Excel or Gemini in Sheets on one recurring report, and use the narrative prompt for a real readout.
- Week 4: Write your personal verification checklist, share it with the team, and tally the hours saved against the errors caught.
Frequently asked questions
Will AI replace data analysts?
Can I upload company data to ChatGPT?
Is ChatGPT good at SQL?
Do I still need to learn SQL and Python if AI writes them?
Terms used on this page
Related roles
- AI for Financial AnalystsAI writes the first pass of your variance commentary, audits your model for hard-codes and sign errors, and turns a 10-K into a table with page references. You decide what the numbers mean and what to tell the CFO.
- AI for Software DevelopersCoding assistants write, refactor, test, and explain code, and agentic tools run the whole loop. Treat every generated line as a pull request from a fast, overconfident junior: useful, never merged unread.
- AI for MarketersDraft campaigns faster, mine customer feedback for messaging, and turn reporting into a ten-minute job, without drifting off-brand or into a CAN-SPAM problem.
- AI for Product ManagersAI turns forty interview transcripts into themes, a rough idea into a PRD draft, and a napkin sketch into a clickable prototype, and it will also confidently invent a customer need that does not exist. Here is how product managers get the first without the second.