How is Claude for data analysis?

The Promise and the Problem: Claude as Your Analyst

Can an AI chatbot really replace your analyst? The honest answer is yes — but only under conditions most teams skip. Claude's data-analysis story is one of remarkable capability wrapped around a surprisingly fragile core.

The adoption signal is unmistakable. According to Anthropic's research on roughly 400,000 Claude Code sessions, writing and data analysis roughly doubled from about 10% to 20% of sessions between October 2025 and April 2026. Practitioners back this up with real results: one analyst described analyzing 78 years of London weather data in plain English, with Claude Code writing and executing all the code in the background (Medium Data Science Collective).

Then comes the number that stops everyone cold. Anthropic's own internal deployment found that Claude answered only 21% of analytics questions correctly out of the box — without curated context, it frequently queried the wrong table, applied the wrong filters, or misinterpreted metric definitions, even while writing technically flawless SQL (InfoQ reports).

So which is it: a junior intern or a senior analyst? The evidence says both, depending entirely on what you give it to work with. As one practitioner put it, Claude Code without context performs at roughly "junior intern" level, but with documentation and a structured workflow it produced "correct queries, clear visualisations and a consistent document structure" (Towards AI case study).

What separates the two outcomes comes down to a handful of factors:

• Table documentation and metric definitions — without them, Claude guesses at column meanings
• Example queries and a defined report structure to anchor output format
• Structured workflows that prevent the agent from improvising or batching queries
• Domain knowledge from the human — Anthropic found success depends on understanding the problem, not coding skill

That last point deserves emphasis. Non-coders succeed at nearly the same rate as software engineers, and users make about 70% of planning decisions while the agent handles execution. Claude doesn't replace analytical judgment — it accelerates it, as one practitioner guide puts it.

The gap between 21% and the 95%+ accuracy Anthropic eventually achieved internally is the whole story of this article. Closing it requires skills, context engineering, and ongoing maintenance — which is exactly the problem SkillzBench is being built to help with, by vetting and benchmarking the agent skills that make tools like Claude production-ready. The rest of this piece explains how.

Context Beats Capability: Why Skills Turn 21% Into 95%

The single most striking data point in Claude's data-analysis story isn't the 95% — it's the 21%. When Anthropic deployed Claude for internal business analytics, the model answered just 21% of analytics questions correctly on its own, despite being able to write technically flawless SQL. After encoding analytical workflows, metric definitions, and business context as curated skills, accuracy jumped to roughly 95%, approaching 99% in some domains, according to reporting on Anthropic's internal deployment.

Anthropic's own framing is that this is "a context and verification problem, not a code generation problem." The model rarely fails at syntax; it fails at interpretation. The team identified three recurring failure modes:

• Entity ambiguity — Claude guesses at what a column or metric means when nobody has defined it.
• Staleness — definitions and table structures drift as the business changes, and stale skills silently produce wrong answers.
• Retrieval failure — the agent pulls the wrong table, filter, or join because the correct context isn't surfaced at query time.

Practitioner experience mirrors the pattern. One analyst found that without table documentation, Claude Code generated SQL errors and simply guessed at column meanings; adding documentation, example queries, and a chart style guide was described as "a game changer" that turned output from "junior intern" level into correct queries and consistent report structure.

Two caveats deserve equal attention. First, the gains are perishable: accuracy decayed from ~95% to 65% within a single month when documentation maintenance lapsed. Skills are maintained assets, not one-time setups. Second, more context isn't automatically better — giving the agent thousands of historical queries moved accuracy by less than one point, and letting an LLM auto-generate metric definitions actually made things worse. Humans must own metric definitions, full stop.

There's also a cost side to verification. Adding adversarial review of every query delivered a further +6% accuracy, but at the price of 32% more tokens and 72% higher latency — a trade-off worth weighing for high-volume analytics.

For teams evaluating Claude for data analysis, the takeaway is that the model is only half the system. The other half is the context layer: curated skills, human-owned metric definitions, and validation steps. And since roughly a quarter of published agent skills carry at least one vulnerability, vetting what you install matters as much as writing it — which is why SkillzBench is building free benchmarking, security scanning, and a free MCP connector so teams can equip agents with skills that are tested, not trusted blindly.

From Chat to Agent: Claude's Evolving Analysis Workflow

Data analysis is quietly moving out of the chat window. Instead of pasting a CSV and asking questions one at a time, more people are handing Claude a dataset and letting it work autonomously — writing code, running it, and delivering a finished report.

Anthropic's own Managed Agents cookbook shows the pattern in action: upload a CSV, and the agent autonomously analyzes it using eight tools (bash, read, write, edit, glob, grep, web_fetch, and web_search) plus preinstalled pandas and plotly. In the demo, a 50-row e-commerce CSV was analyzed "in a few minutes," producing a 53,728-byte narrative HTML report with three or more embedded interactive charts. Anthropic's guidance for these setups includes sampling large tables rather than loading everything, and sanity-checking key metrics before building a narrative around them.

Practitioners are landing similar wins outside Anthropic's walls. One analyst analyzed 78 years of London weather data entirely in plain English — describing the data, explaining what he wanted to know, and letting Claude Code write and execute all the code in the background. No SQL, no Python by hand.

The usage data backs up the migration. Writing and data analysis roughly doubled from about 10% to 20% of Claude Code sessions between October 2025 and April 2026, based on roughly 400,000 sessions from ~235,000 people, according to Anthropic's research. Meanwhile, the Economic Index shows a telling counter-trend: Business and Financial Operations usage on Claude.ai declined in relative share, from 6% to 3%, and Management fell from 5% to 3%. The likely explanation is that analysis work is migrating from chat toward agentic tools rather than disappearing.

There's a caveat worth flagging before you set expectations. The headline figure — 95% of business analytics queries automated at ~95% accuracy — comes from Anthropic's own heavily governed internal environment and is self-reported, not independently verified. Independent commentary on the results, like BI architect Francesco Mucio's observation that Claude queries a Semantic Layer rather than raw tables, suggests the result depends on infrastructure most teams don't have yet.

What does carry over to ordinary teams is the workflow itself. The agentic pattern works because it bundles the pieces that matter:

• Structured tool access — bash, file operations, and search — so the agent can execute rather than just suggest.
• Skills packages encoding metric definitions, table documentation, and validation steps, which drove accuracy from 21% to 95% in Anthropic's deployment.
• Human-owned context, since LLM auto-generated definitions made accuracy worse, not better.

That skills layer is exactly why accuracy is so sensitive. If you're building agentic analysis workflows, the quality of the skills you attach determines whether your agent behaves like a senior analyst or an unverified guesser. SkillzBench is building free benchmarking and security scanning for agent skills — testing each skill in a sandbox with the same model, with and without it, and publishing the lift in percentage points — plus a free MCP connector to equip agents with vetted skills. For a workflow where one stale or ambiguous definition can quietly corrupt every downstream number, knowing which skills actually help is not optional.

Your Domain Expertise Is the Real Multiplier

Here's the counterintuitive truth about Claude and data analysis: you don't need to code to succeed with it. Anthropic's research across roughly 400,000 Claude Code sessions found that "success is determined by how well a person understands the problem they are trying to solve, not whether they're trained in coding" — non-coders succeed at nearly the same rate as software engineers.

This reframes what matters. Users make about 70% of planning decisions but only around 20% of execution decisions. As Anthropic puts it, "People decide what to build, and the agent decides how to build it." Your domain knowledge — knowing which metric matters, which filter is wrong, which number looks off — is the multiplier, not your Python skills.

That finding also explains why context beats capability. Anthropic's internal deployment saw analytics accuracy jump from 21% to 95%+ purely by encoding business context and analytical workflows as skills — the model didn't change, the context did.

So how do you put domain expertise to work in practice? Four habits consistently show up in practitioner accounts:

• Paste a representative sample of 50–100 rows rather than dumping an entire dataset, per practitioner guidance.
• Provide 2–3 sentences of business context — what the data means, who uses it, what decision it informs.
• Use a todo-list workflow: read task, execute, mark complete, repeat. One analyst found this stopped Claude from improvising and batching queries into one messy notebook.
• Expect 3–5 iterative exchanges per analysis, interrogating Claude's assumptions along the way.

That last habit deserves emphasis. Claude will often surface its own uncertainty when prompted, so ask it directly: "What did you assume about this column?" or "Why did you filter out those rows?" As one case study put it, if you don't tell Claude what a column means, it's going to guess — and a plausible-looking guess is more dangerous than an obvious error.

Without context, practitioners describe Claude as roughly "junior intern" level. With documentation, business framing, and structured iteration, it produces correct queries and clear visualizations. The gap between those two outcomes is entirely filled by what you bring to the conversation.

For teams running recurring analyses, agent skills packages can encode this context once and reuse it everywhere — SkillzBench is building a free directory with security scanning, benchmarking, and a free MCP connector so you can equip your agent with vetted skills rather than reinventing your analytical context every session.

Getting Started: A Practical Playbook (and Where Skills Fit In)

Claude answered only 21% of analytics questions correctly before Anthropic added curated skills and documentation — and over 95% after. The difference wasn't a better model; it was better context. That finding from Anthropic's internal deployment, reported by InfoQ's coverage, is the clearest signal that success with Claude for data analysis comes from setup, not prompting luck.

Start by documenting your tables and metric definitions before you write a single query. A practitioner case study found that without table documentation, Claude generated SQL errors and guessed at column meanings; adding docs, example queries, and a style guide was described as "a game changer". One warning from Anthropic's own experience: LLM auto-generated metric definitions made accuracy worse, so humans must own those definitions.

Next, split query creation from report writing. A structured workflow — one where Claude reads a task, executes it, marks it complete, then repeats — stopped the agent from improvising and batching everything into a single notebook. Treat these as separate prompts with separate verification steps.

Sanity-check key metrics before building narrative around them. Anthropic's own Managed Agents guidance recommends sampling large tables rather than loading everything, and validating headline numbers before writing conclusions. Claude doesn't replace analytical judgment — it accelerates it, so interrogate its assumptions rather than trusting fluent prose.

Finally, budget for ongoing maintenance. This is the step most teams skip:

• Accuracy decayed from ~95% to 65% within one month when documentation upkeep lapsed.
• Adversarial review of every query added +6% accuracy, but cost 32% more tokens — decide where that trade-off is worth it.
• Giving the agent thousands of historical queries moved accuracy by less than one point; curated context beat raw volume.

Skills are clearly the decisive lever — but they require vetting and upkeep. SkillzBench is building free benchmarking, security scanning, and a free MCP connector so you can equip your agent with vetted data-analysis skills rather than guessing which ones help. Since published research found roughly 26% of skills carry at least one vulnerability, verification before installation is not optional. Treat your skills as maintained assets, not one-time downloads, and Claude becomes a genuinely reliable analyst.