Most data analytics guides show you toy examples. "Ask Claude to summarize this 10-row table." That's not analytics. That's a party trick.

This guide covers real work: cleaning messy datasets, writing production SQL, building analysis pipelines, finding patterns in thousands of rows, and generating visualizations that actually communicate something.

I run a company where AI handles analytics end-to-end. No dedicated data team. No BI platform subscription. Here's how.

Why Claude for Data Analytics

Three things set Claude apart from other AI tools for analytics work:

Long context. Claude Opus processes up to 1M tokens in a single conversation. That's roughly 750,000 words — enough to paste an entire quarter's transaction log and ask questions about it. ChatGPT-4o caps at 128K. For analytics, context window size directly determines how much data you can reason about at once.

Structured reasoning. Claude doesn't just pattern-match against your data. It thinks through methodology, considers edge cases, and explains its reasoning. When you ask "why did revenue drop in March," Claude will check for seasonality, compare against prior periods, look for confounding variables, and tell you what it can and can't conclude.

Code quality. The analysis scripts and SQL Claude generates are clean, commented, and production-ready. Not notebook-quality throwaway code — actual scripts you can commit to a repo and run in CI.

The Five Core Analytics Workflows

1. Data Cleaning and Preparation

Every analytics project starts with messy data. Claude handles the grunt work.

The prompt pattern:

Here's a CSV export from [source]. I need to:
1. Identify and handle missing values
2. Standardize date formats to ISO 8601
3. Remove duplicates based on [key columns]
4. Flag outliers in [numeric columns]

Output a Python script using pandas that processes this data
and saves the cleaned result. Include logging for every
transformation so I can audit what changed.

Claude won't just write the script. It'll ask about your tolerance for missing data, suggest appropriate imputation strategies based on the column types, and handle edge cases like mixed date formats or encoding issues.

Real example: I pasted 3,000 rows of Stripe transaction data with inconsistent currency formatting, duplicate webhook entries, and timezone-confused timestamps. Claude produced a 45-line pandas script that cleaned everything, deduplicated on charge ID, normalized to UTC, and output a summary of what it changed. Ran perfectly on first execution.

2. Exploratory Data Analysis

This is where Claude genuinely shines. Instead of manually plotting histograms, you describe what you're trying to understand.

The prompt pattern:

I have [describe dataset]. I want to understand:
- Distribution of [key metric]
- Correlation between [variable A] and [variable B]
- Any segments or clusters in the data
- Anomalies or unexpected patterns

Write a Python analysis script that generates visualizations
for each finding. Use matplotlib with a clean style.
Annotate anything surprising.

Claude approaches EDA like a human analyst would — it starts with summary statistics, checks distributions, looks for relationships, and flags things that don't fit the expected pattern. The difference is it does this in seconds.

Pro tip: Upload your data to a Claude Project, then have ongoing conversations about it. Each question builds on previous context. "Now break that down by region." "What happens if we exclude Q4?" "Run the same analysis but segment by customer tier." This iterative exploration is where AI analytics gets genuinely powerful.

3. SQL Query Generation

If you work with databases, this alone justifies using Claude.

The prompt pattern:

Database: [PostgreSQL/MySQL/BigQuery/etc.]
Schema:
- users (id, email, created_at, plan, status)
- orders (id, user_id, amount, currency, created_at, status)
- events (id, user_id, event_type, properties, timestamp)

Question: What's the monthly retention rate by acquisition
cohort for the last 12 months? Include only users who made
at least one purchase in their first 30 days.

Claude generates correct, optimized SQL — not the naive nested-subquery approach most AI tools produce. It uses CTEs for readability, window functions appropriately, and adds comments explaining the logic.

For complex analytical queries — cohort analysis, funnel conversion, LTV calculations, churn prediction — Claude's SQL is often better than what a mid-level analyst would write. It considers index usage, avoids unnecessary full table scans, and structures the output for easy consumption by downstream tools.

Advanced move: Paste your EXPLAIN ANALYZE output and ask Claude to optimize the query. It'll identify missing indexes, suggest materialized views, and rewrite subqueries as joins where appropriate.

4. Statistical Analysis and Hypothesis Testing

This is where most non-statisticians get stuck, and where Claude provides the most value.

The prompt pattern:

I'm testing whether [change/intervention] affected [metric].

Control group: [describe]
Test group: [describe]
Data: [paste or describe]

Run the appropriate statistical test, check assumptions,
and tell me:
1. Is the effect statistically significant?
2. What's the effect size?
3. What's the practical significance?
4. What could confound this result?

Claude doesn't just run a t-test and hand you a p-value. It checks whether a t-test is even appropriate — normality, equal variance, sample size. It suggests alternatives when assumptions are violated (Mann-Whitney U, bootstrap confidence intervals). And critically, it distinguishes between statistical significance and practical significance — something many analysts miss.

5. Dashboard and Visualization Generation

Claude generates publication-quality visualization code across matplotlib, seaborn, plotly, and D3.js.

The prompt pattern:

Create a dashboard showing:
1. [Metric] over time (line chart, monthly)
2. [Metric] by [category] (horizontal bar, sorted)
3. [Metric A] vs [Metric B] scatter with trend line
4. KPI summary cards for [list key numbers]

Use plotly for interactivity. Color palette: [specify].
The audience is [executive/technical/etc.].

The key insight: tell Claude who the audience is. Executive dashboards get fewer charts with bigger numbers and trend arrows. Technical dashboards get more detail, statistical annotations, and drill-down capability.

Custom Instructions for Analytics Work

Set these in your Claude Project instructions to get consistently better analytics output:

When I share data for analysis:
1. Start with data quality assessment before any analysis
2. State assumptions explicitly
3. Use reproducible code (set random seeds, pin library versions)
4. Distinguish correlation from causation
5. Include confidence intervals, not just point estimates
6. Flag when sample size is too small for reliable conclusions
7. Output code as complete, runnable scripts — not fragments
8. Use descriptive variable names, not x, y, df1, df2

These instructions prevent the most common AI analytics failures: jumping to conclusions on dirty data, confusing correlation with causation, and generating code that only works inside the conversation.

Real Workflow: Monthly Business Review

Here's how I actually use Claude for analytics every month:

  1. Export data. Stripe revenue, Google Analytics traffic, Search Console queries, email subscriber counts. CSV exports, 30 seconds.

  2. Upload to Claude Project. One project called "Business Analytics" with the custom instructions above. Upload the fresh exports.

  3. Run the analysis. One prompt: "Generate the monthly business review. Compare to last month and same month last year. Flag anything that needs attention." Claude produces a structured report with charts, commentary, and recommended actions.

  4. Drill into anomalies. "Why did organic traffic drop 15% week over week?" Claude cross-references Search Console data with content publishing dates and algorithm update timelines.

  5. Generate the report. "Format this as a one-page executive summary with the three most important findings." Done. What used to take a data analyst half a day takes 10 minutes.

What Claude Can't Do (Yet)

Honesty matters more than hype:

  • Real-time streaming data. Claude works on snapshots, not live feeds. For real-time dashboards, use Claude to build the system, then run it independently.
  • Datasets larger than context window. For truly massive data (millions of rows), use Claude to write the analysis code, then execute it on your infrastructure with pandas, Spark, or DuckDB.
  • Proprietary tool integrations. Claude can't directly query your Snowflake instance or pull from your Tableau server. It works with data you bring to it.
  • Regulatory compliance verification. Claude can help structure GDPR-compliant data pipelines, but don't rely on it as your compliance authority.

Getting Started Today

You don't need a data engineering background. Start here:

  1. Pick one recurring report you currently build manually.
  2. Export the underlying data as CSV.
  3. Paste it into Claude with the prompt: "Analyze this data and produce the same insights I'd get from [describe your current report]. Generate the analysis as a Python script I can rerun next month."
  4. Iterate. Refine the prompts, add custom instructions, build up a Project with your schema documentation and business context.

Within a week, you'll have an analytics workflow that runs in minutes instead of hours. Within a month, you'll be asking questions about your data that you never had time to ask before.

That's the real unlock — not just faster analytics, but more analytics. Questions you'd never bother investigating because the manual effort wasn't worth it. Claude makes the marginal cost of curiosity nearly zero.

Sophia Cave is the founder of Like One, where AI runs the entire operation. She writes about building real AI systems — not demos.