Can Claude do data visualization?

The BI Bottleneck: Why Teams Are Asking AI to Draw Their Charts

Your analyst is buried, your dashboard request is #47 in the queue, and by the time the report lands, the question you needed answered has already changed. That's the BI bottleneck in miniature: building dashboards is slow, data teams are stretched thin, and a static report can't answer the follow-up question you ask two minutes after seeing it.

The frustration is driving real interest in AI-generated visuals. Companies using data visualization reportedly see a 20–30% improvement in decision-making speed, according to McKinsey figures cited by aivancity, and roughly 49% of companies adopting advanced AI report significantly better understanding of their internal data, per Deloitte figures reported through the same outlet. These are second-hand market stats — treat them as directional context, not gospel — but the trend they point to is hard to argue with: teams want answers drawn, not just delivered.

That raises the obvious question: can Claude actually visualize data, or does it just describe charts in prose? The direct answer is yes — and not in one way, but four:

• In-chat chart generation, where Claude draws charts using HTML and XML vector graphics — Anthropic describes it as giving Claude "its own whiteboard" — available since a March 2026 beta launch, per Engadget's coverage.
• Artifacts: interactive, shareable visualization apps built from uploaded files like CSVs, PDFs, and DOCX up to 30MB, exportable as .tsx React files.
• Claude Dashboards (beta): live artifacts that query sources like Snowflake, Databricks, and Microsoft Fabric through MCP connectors and render charts in d3.js.
• Chart comprehension: Claude can also read chart images, and research shows it analyzes data more accurately when it can see the visual.

That last point matters more than it sounds. A 12,000-trial study found Claude 3.5 Sonnet was at least 1.7x more accurate on complex clustering tasks when given a correct scatterplot than when working from raw data alone, according to the researchers' preprint — evidence that visualization is becoming two-directional, not just an output format.

The catch, which practitioners are quick to flag, is that context determines quality. BI author Ryan Dolley warns that contextless AI dashboard generation is "a quick path to good looking, unusable outputs" — and that the same principle applies to any agent skill you bolt onto a workflow. Tools that touch your data deserve the same scrutiny: SkillzBench is building a free directory that verifies, security-scans, and benchmarks agent skills so you can see what actually helps before you connect it.

Claude's Four Visualization Capabilities: Charts, Artifacts, Dashboards, and Chart Reading

Claude doesn't just make charts — it reads them, builds interactive apps from your data, and increasingly decides on its own when a visual would help. Here are the four capabilities that answer the question.

1. In-chat chart generation. Since March 12, 2026, Claude produces charts, graphs, and diagrams directly in conversations using HTML code and XML vector graphics — what Anthropic describes as "like giving Claude access to its own whiteboard," not raster image generation. According to Engadget's coverage, the feature is available to all users including the free tier, and Claude can autonomously decide when visuals would help rather than waiting to be asked — a differentiator versus ChatGPT and Gemini, which generally require explicit prompts or datasets.

2. Artifacts. Upload a data file and Claude builds an interactive, shareable visualization app with no code required. Supported formats include:

• CSV, PDF, HTML, and DOCX files, up to 30MB each
• Exports as .tsx (TypeScript) files for React integration
• Connections to 9,000+ apps via Zapier for live data

Practitioners report Claude works best with CSV files, though it can hallucinate with large datasets or too many filters — once outputs go glitchy, starting a fresh chat is the reliable fix.

3. Claude Dashboards (beta). Released October 8, 2026, Dashboards are live artifacts rendered as a single HTML page with d3.js charts that query sources like Snowflake, Databricks, and Microsoft Fabric through remote MCP connectors. Hands-on testing by the Tabular Editor team confirms size caps of 8MB per dataset and 100MB database storage, and notes that live queries require a remote MCP server — local servers only feed extracts during building. Because MCP connectors are the backbone of live-data integration, agent builders extending Claude's reach can browse and test free skills and MCP connectors on SkillzBench, where every skill is security-scanned and benchmarked before it earns a trust score.

4. Chart comprehension. A 12,000-trial academic study from Harvard and Google researchers found Claude 3.5 Sonnet is at least 1.7x more accurate on complex clustering tasks when given a correct scatterplot than raw data alone. The effect cuts both ways: a misleading linear best-fit plot reduced parabola-detection success by more than half, and models flagged chart-data discrepancies in only 7 of 800 deceptive trials. As the researchers put it, AI systems "like humans, can benefit from visualization" — but may silently fail when visuals and data disagree.

Together, these four mechanisms make Claude a genuinely bidirectional visualization tool — one that both produces and consumes charts, with the caveat that verification still belongs to you.

Where It Breaks: Hallucinations, Context Gaps, and Governance Limits

A chart that looks polished but quietly gets the numbers wrong is worse than no chart at all — and that's exactly the failure mode Claude users need to plan for. This section covers where Claude's visualization capabilities genuinely break down.

Start with hallucination. Zapier's hands-on testing found that Claude "tends to hallucinate when working with large datasets or if you ask it to add too many filters or layers," and once it starts down this glitchy path, it tends to double down. Their recommended fix is blunt: start a new chat rather than trying to talk the model out of its error.

Context gaps cause a subtler problem. BI practitioner Ryan Dolley warns that contextless dashboard generation is "a quick path to good looking, unusable outputs." In his experience, what looks like hallucination is often a missing-context guess — SQL referencing tables the model never saw in a schema isn't fabrication, it's an incorrect assumption. Providing project context (design standards, architecture docs, data source details) is what turns pretty-but-useless outputs into usable ones, which is why his testing shows context is the key to making Claude useful for data apps.

The most rigorous warning comes from a 12,000-trial academic study of how AI models read charts:

• Models mentioned a chart/data discrepancy in only 7 of 800 deceptive trials — a "silent failure" risk where the model never flags that the visual contradicts the data.
• A misleading linear best-fit plot cut parabola-detection success by more than half, showing that bad visuals actively degrade analysis rather than just failing to help.
• Raw-data-only comprehension "can be brittle," meaning Claude's answers without correct visuals are more fragile than they appear.

Anthropic's own Academy guidance reinforces this: users should manually verify computed outputs, because "an app that looks polished but gets the numbers wrong will mislead the people you share it with."

Enterprise buyers face a different set of limits. The Tabular Editor team's testing of Claude Dashboards flags "clear and present challenges for data governance and compliance": no EU data residency, data routed to Anthropic's US servers, and no zero data retention for the Claude apps where dashboards are built. They conclude Dashboards suit personal BI and small teams but are "a difficult option to consider for enterprise scenarios."

Availability is also murkier than marketing suggests. Engadget reports chart generation reached all users including the free tier, while another report says interactive visualizations rolled out primarily on paid plans — the sources may describe different features or rollout stages, so verify what your plan actually includes before building a workflow around it.

The through-line is trust, not capability. Claude can visualize data impressively, but every source above — academic, vendor, and practitioner — lands on the same advice: verify before you share. That's the same principle behind SkillzBench's approach to agent skills — every skill gets security-scanned, version-pinned, and benchmarked in a sandbox before you rely on it. Skills that automate your visualization workflows deserve the same scrutiny; you can browse and test free ones on SkillzBench as the platform comes online.

Making Claude's Charts Actually Work: A Practical Playbook

A polished chart built on messy data is worse than no chart at all — and Claude will happily build you one. The good news: most of Claude's visualization failures trace back to inputs, not the model, and a short preparation routine fixes most of them.

Start with clean, standardized data. Claude accepts PDF, HTML, DOCX, and CSV files up to 30MB, but hands-on testing by Zapier found it works best with CSV — so standardize your dates, strip duplicates, and convert before uploading. The same testing surfaced a second warning: Claude tends to hallucinate with large datasets or too many filters and layers, and once it starts down a glitchy path, it doubles down. The recommended fix is simply starting a fresh chat rather than arguing with a conversation that has already gone sideways.

Next, give Claude your standards instead of letting it guess. BI practitioner Ryan Dolley, who has built data apps in Claude extensively, calls contextless dashboard generation "a quick path to good looking, unusable outputs." With a PRD, schema, tech stack, and meeting notes attached, Claude follows design conventions, refuses to compute metrics from insufficient data, and even flags mismatches — like a raw CSV where your standard calls for a Snowflake semantic layer. Notably, what looks like hallucination (SQL with wrong table names) is usually just an incorrect guess from missing context, not model failure.

Then verify the numbers yourself. Anthropic's own guidance is blunt: an app that looks polished but gets the numbers wrong will mislead everyone you share it with, and you are the only one who can catch that first. That advice is well-founded — a 12,000-trial research study found models noticed chart/data discrepancies in only 7 of 800 deceptive trials, a silent failure mode that hand-verification is the only defense against.

For developers, a few practical notes:

• Artifacts export as .tsx (TypeScript) files, ready to drop into a React project.
• Dashboards render as a single HTML page with charts built in d3.js.
• Dashboard export destinations include Grafana Cloud, Hex, Sigma, and Amplitude — but not Power BI or Fabric.
• Live queries to Snowflake, Databricks, or Fabric require a remote MCP server; local servers only feed extracts during the build phase.

If you want Claude following your team's conventions automatically, agent skills and MCP connectors are how that happens — SkillzBench is building a free directory of verified, benchmarked skills plus a free MCP connector so you can test what actually improves output before trusting it in production.

Extending the Workflow: MCP Connectors and Verified Agent Skills

Live dashboards are only as good as the pipes behind them. Claude Dashboards, released in beta in October 2026, render charts with d3.js — but every live query runs through a remote MCP server, meaning connectors to Snowflake, Databricks, or Microsoft Fabric are what separate a static snapshot from genuinely live BI (hands-on testing). Local MCP servers can only feed extracts during the build phase; the live part depends entirely on remote infrastructure.

That makes MCP connectors and agent skills the real backbone of AI-driven BI. The same connectors that power dashboards can also automate the surrounding workflow — cleaning CSVs before upload, validating computed outputs, and assembling reporting context. This matters because, as one BI practitioner found, Claude without project context produces "good looking, unusable outputs," while Claude with proper context follows design standards and flags data mismatches on its own.

There is a trust problem, though. Published research found that roughly 26% of published agent skills carry at least one vulnerability, and Anthropic itself advises users to manually verify computed numbers because a polished-looking app that gets figures wrong will mislead everyone who sees it (Anthropic Academy guidance). The risk compounds silently: in a 12,000-trial academic study, models mentioned a chart/data discrepancy in only 7 of 800 deceptive trials — a failure mode the researchers describe as silent.

Before adopting any skill for a BI workflow, look for three safeguards:

• Version pinning — every skill version should be pinned to a specific commit SHA, so what you tested is what runs.
• Security and license scanning — validated, license-checked, and quality-judged before it touches your data.
• Independent benchmarking — tested in a sandbox with the same model, with and without the skill, so you know whether it actually helps.

This is exactly the gap SkillzBench is being built to close. Its free, independent skills directory continuously crawls and dedupes skills from GitHub, then verifies, scans, and benchmarks each one — and its free MCP connector (currently in build, with the waitlist live) will let you connect any agent to only the skills you choose and trust. Everything on the platform is free: the directory, the connector, the benchmarks, and the alerts. If you're wiring Claude into your data stack, that's the place to browse, vet, and test before anything gets near production numbers.