If you have ever needed a quick number from your company data, you know the routine. You file a ticket, wait for an analyst, or open a spreadsheet export and hope the formulas still work. On 8 October 2026 Anthropic put Claude Dashboards into beta, and it removes most of that waiting: you ask a question in plain English, and Claude builds a live dashboard that keeps itself current.
The catch is that a dashboard built in thirty seconds can be wrong in thirty seconds too. This guide covers what Dashboards does, who can use it, a copy-paste "metric brief" prompt that gets the right chart first time, and the two checks I run before I share any dashboard Claude builds.
What is Claude Dashboards?
Claude Dashboards is a beta feature, launched on 8 October 2026, that turns a plain-language question into a live dashboard built from your connected data. Claude writes the queries, picks the charts and lays out the page, then refreshes the numbers as the data changes. Every number links to the query behind it.
According to Anthropic's announcement, Dashboards connects to data platforms such as Amazon Redshift, BigQuery, ClickHouse, Databricks and Snowflake, and also works with business-app connectors. Anthropic's own example is a dashboard of Salesforce opportunities.
Three design choices matter for everyday users:
--- Inspectable numbers. Click any figure to see the query that produced it, or ask Claude to explain it in plain English.
--- Visible freshness. Each chart shows when its data was last refreshed, so you know whether you are looking at this morning or last week.
--- A handoff, not a replacement. When a question needs deeper analysis, you can send the dashboard to Amplitude, Grafana, Hex, Mixpanel, Omni, Perplexity, PostHog or Sigma. Looker, monday.com and Tableau are listed as coming soon.
Anthropic describes Dashboards as best for quick, exploratory questions, such as how this week's sign-ups compare with last month's. That framing is honest, and it is the right way to use it.
Which Claude plans include Dashboards?
Dashboards is in beta on paid Claude plans: Pro, Max, Team and Enterprise. Claude Motion, the animation tool launched the same day, is limited to Team and Enterprise. Docs, Slides and Design left beta on the same date and now work on every plan, including Free. Enterprise admins must switch Dashboards on themselves.
The availability details, all from the 8 October announcement:
--- Paid individual plans: Pro and Max users can try Dashboards now, as long as they can connect a supported data source.
--- Enterprise: Dashboards and Motion are off by default. Admins turn them on in Organization settings, under Artifacts.
--- Free plan: no Dashboards. You can still ask Claude to turn a pasted table or uploaded CSV into a static chart, but it will not refresh itself.
--- Sharing: dashboards can be shared with people outside your organisation, or with anyone who has the link, only if your admin allows it.
One practical note for Hong Kong teams: if your sales data lives in Excel files on a shared drive rather than in a warehouse or CRM, Dashboards has nothing live to connect to yet. Start with the system your team already treats as the source of truth.
How do you write a dashboard prompt Claude gets right?
Write a metric brief, not a wish. A metric brief names five things: the exact metric and how it is calculated, the source table or app, the time window and grain, the filters, and one number you already know. Vague prompts such as "show me sales" force Claude to guess definitions, and guessed definitions produce believable but wrong charts.
Anthropic's sample prompt is "Build a dashboard of this quarter's revenue by region." It works for a demo, but in a real company "revenue" might mean invoiced, paid or booked, and "this quarter" might be calendar or fiscal. Here is the brief I use instead.
Try this prompt:
Build a dashboard from our Salesforce connector.
METRIC: Closed-won revenue = sum of Opportunity Amount where Stage = "Closed Won". Count each opportunity once, using Close Date.
SOURCE: Salesforce Opportunities only. Do not join other tables unless you tell me why first.
WINDOW AND GRAIN: Last 12 full weeks, weekly, weeks starting Monday, Hong Kong time.
FILTERS: Exclude opportunities owned by the test user "Sandbox Admin". Show the top 5 regions, group the rest as "Other".
KNOWN ANSWER: Total closed-won revenue for September 2026 should be about HK$1.8 million. If your number is more than 5% away, stop and show me the query before building charts.
Layout: one headline number for the last full week, one weekly trend line, one bar chart by region. Under each chart, write one sentence describing exactly what it counts.
Three details do most of the work:
--- The definition line. Spelling out how the metric is calculated removes the single biggest source of error, which is Claude choosing a sensible but different definition from your finance team.
--- The join rule. Asking Claude to explain any join before it makes one stops double counting, the most common silent mistake in generated SQL.
--- The known answer. Giving Claude a number to hit turns it into its own first reviewer. If it cannot match your September total, you find out before anyone sees a chart.
How do you check a Claude dashboard before sharing it?
Run two checks: a known-answer test and a query read-through. First compare one headline number against a report you already trust. Then open the query behind each chart and ask Claude to explain its joins, date filters and exclusions in plain English. Together they take about ten minutes and catch most wrong dashboards.
Check 1: the known-answer test
Pick a number that someone in your company has already signed off: last month's total revenue, last quarter's new customers, or the number of tickets closed in September. Ask the dashboard for the same figure. If it matches within a few percent, the definition is probably right. If not, the gap tells you where to look. A total that is roughly double usually means a join is counting rows twice. A total that is slightly low usually means a date filter or a time zone is cutting off the edge of the period.
Check 2: the query read-through
You do not need to read SQL to do this. Click the number, open the query, and paste this follow-up:
Try this prompt:
Explain the query behind this chart for someone who does not write SQL.
1. Which tables does it read, and does it join any of them? If yes, could the join count any record twice?
2. What exact date range does it use, and in which time zone?
3. What does it exclude, and why?
4. List any assumption you made that I did not state in my brief.
Answer in short numbered lines. Do not change the dashboard yet.
The fourth question is the useful one. Claude is good at telling you what it assumed when you ask directly, and those assumptions are exactly what you need to confirm with the person who owns the data.
Where does Claude Dashboards fall short?
Dashboards is a beta built for quick questions, not a governed reporting system. It depends on clean, connected data, it can write plausible but wrong queries, and it has no shared metric layer, so two people asking the same question can get different definitions. Use it to explore, then move recurring reports to your BI tool.
The limits worth knowing before you rely on it:
--- No connector, no dashboard. It only sees what your connectors expose. Messy spreadsheets on a shared drive are not a live source.
--- Definitions are per prompt. Without a semantic layer, "active customer" means whatever the person asking said it means. This is why the metric brief matters.
--- Permissions follow the connector. Use a read-only account for the data connection, and check who can see a dashboard before you share the link outside your team.
--- Refreshes are not alerts. A dashboard that updates is not the same as a system that warns you when a number moves. For alerting and cohort analysis, use the handoff to a dedicated analytics tool.
--- It is a beta. Anthropic may change features, connectors and limits. Treat anything you build now as a working draft.
None of this makes Dashboards less useful. It means the person asking still owns the definition, the check and the decision to share.
Try it now: your first 20-minute Claude dashboard
Pick one question you already know the answer to, write a metric brief for it, build the dashboard, then run both checks. Starting with a known answer is the fastest way to learn how Claude interprets your data, and it gives you a tested brief you can reuse for every dashboard after it.
--- Minutes 0 to 5: choose a question with a signed-off answer, such as last month's revenue or orders. Confirm your data source is connected in Claude.
--- Minutes 5 to 10: fill in the metric brief above. Be strict about the definition, window and known answer.
--- Minutes 10 to 15: let Claude build the dashboard, then run the known-answer test against your trusted report.
--- Minutes 15 to 20: paste the query read-through prompt for each chart. Note every assumption Claude lists, and fix the brief where it guessed.
Save the corrected brief in a note or a Claude Project. Next time someone asks for a similar view, you start from a definition that already passed a check. That is the real productivity gain here: not a faster chart, but a faster chart you can defend in a meeting. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team. Feature and plan details reflect Anthropic's Claude Dashboards and Motion announcement of 8 October 2026; beta features, connectors and plan access may change. Related reading: how Claude Docs, Slides and connectors work in one chat and output contracts for consistent AI results.
Is your team ready to trust AI with its data?
A dashboard is only as good as the data and habits behind it. Take UD's free AI Ready Check to see how prepared your business is, and when you want AI built into your reporting and daily workflows, we'll walk you through every step, from tool setup to workflow design and deployment.
