Skip to main content
ON DEMANDEveryone builds. IT approved.Watch the replay
Sigma Computing
Fundamentals

AI dashboard generator: How to build dashboards and apps by describing them

TJ Wells
TJ WellsProduct Advocate
November 20, 2026
12 min read
AI dashboard generator: How to build dashboards and apps by describing them

Many analytics platforms now ship an AI dashboard generator, so you can describe a dashboard in plain language and get a working one back. General AI assistants such as Claude also build dashboards inside a chat. The options differ in where the data comes from and what happens after the first draft. Generators built on uploaded files or extracts often show stale data, and some can't change one chart without regenerating the rest.

Sigma is the AI apps and analytics platform that lets users build agentic analyses and applications directly on live cloud data warehouse data. Dashboards generated in Sigma run on current warehouse data under the permissions your data team sets. Good results, ones that you can share with your org, come from a short prompting loop and four tests. Sigma offers two paths that end in a governed workbook you can edit element by element.

Key takeaways

  • Dashboards built from uploaded files or extracts go stale at export and leave the warehouse's access controls behind. Building on live warehouse data lets the dashboard apply each viewer's permissions.
  • Review the generator's plan before it builds, then check each chart's query against the metric definitions finance already uses.
  • Edit one chart at a time, by prompt or directly on the canvas, so charts you already approved stay as they are.
  • Prototyping in a general AI assistant works for quick questions, as long as the dashboard can move into a governed platform before other people rely on it.

What is an AI dashboard generator?

An AI dashboard generator turns a plain-language description into a working dashboard. It selects the data, writes the queries, picks the chart types, and lays out the filters. Some generators go further and build simple apps, with forms that capture input and write it back to the warehouse.

Generators come in two forms. Some sit inside analytics platforms and run against the platform's governed data and metadata. Others live inside general AI assistants. Claude Dashboards, for example, connects to warehouses such as Amazon Redshift, BigQuery, ClickHouse, Databricks, and Snowflake. It then writes a SQL query for each chart and runs it against the source.

AI dashboard generators vs dashboard templates

A template is a fixed layout someone designed in advance. You copy it and fit your data to it. A generator starts from your question and your schema, then builds the layout, metrics, and filters to fit. Templates cover the questions their designer anticipated, and a generator builds around the one you're asking today.

AI dashboard generators vs AI chat answers

A text-to-SQL chat answer usually returns a number or a chart for one question inside a conversation. A generated dashboard persists, with filters, multiple views, and a link a colleague can open next week. Because other people will rely on it, it needs the same governance controls as any shared report.

Where generated dashboards hit limits

Many generators can produce a convincing first draft. Problems appear when that draft relies on an uploaded file, an extract, or a one-off query outside the governed stack.

Data copied out of the warehouse

An uploaded CSV starts going stale the moment you export it. Keeping it current means exporting again, and every export leaves another ungoverned copy outside the warehouse. Each copy is one more place sensitive rows can leak.

Permissions that don't follow the viewer

A file has no idea who is looking at it. Warehouse row-level security doesn't travel with a copy, so unless the dashboard tool adds its own rules, a regional manager and a contractor see the same rows. Teams work around this by building a separate dashboard for each audience, and those copies drift out of sync.

Regenerating instead of editing

When the only way to change a chart is to prompt again, fixing one axis can mean rebuilding the page. Each regeneration can also shift charts you already checked, so someone has to review the whole dashboard again. A dashboard you can't edit precisely is hard to maintain past its first week.

How building a dashboard from a description works

Prompt-to-dashboard building works best as a loop. You describe, review, generate, and refine, then extend the dashboard into an app when it needs to capture input.

Describe the outcome and the data

Start with the question you want answered. Then name the table or data model, the metric definition, the time range, and the grain, the same way you'd label columns before handing someone a spreadsheet.

"Build a dashboard of this quarter's revenue by region" works as a start. Adding "use fiscal quarters, exclude one-time setup fees, and break it out weekly by sales segment" removes most of the guesswork. Spell out definitions the AI could misread, such as a fiscal calendar that doesn't run January through December.

Review a plan before the build

For complex builds, some generators propose a step-by-step plan and ask clarifying questions before they touch the canvas. Treat the plan as a cheap checkpoint. Check for the wrong source table, a join that double-counts, or a metric definition finance won't accept before any charts exist.

Refine single elements with follow-up prompts

After the first version, target one element at a time with prompts like "Add a filter for region" or "Change this to a weekly view." Open each chart's query to confirm it counts what you meant. For changes such as resizing a tile or renaming an axis, you can edit directly instead of prompting.

Add forms and write-back for a simple app

A dashboard becomes a simple app when users can act on what they see. You add an input field or editable table, a form to submit entries, and actions that route approvals or send notifications. Writes should land in governed warehouse tables.

Building in your analytics platform vs prototyping in an AI assistant

Where you build depends on how many people will use the dashboard and how long it'll live. Broad, recurring dashboards belong in your analytics platform, and quick exploratory questions can start in an AI assistant.

When to build directly in the platform

Weekly pipeline reviews, P&L and board reports, and any view touching sensitive customer or financial data belong in the platform from the start. So do dashboards whose numbers must reconcile with what finance reports. The platform supplies permissions and certified metrics from the first build.

When to prototype in an AI assistant first

Claude Dashboards, in beta on paid Claude plans, is a good fit for a one-off question you want to explore. Each chart shows its SQL query and when its data last refreshed, which makes a prototype easy to check. Once a prototype needs deeper analysis or other people start relying on it, it needs a path into a platform with governed data and permissions.

What to look for in an AI dashboard generator

Run these four tests on any AI dashboard generator during a proof of concept. Together they decide whether you can share and maintain a generated dashboard.

Live, governed warehouse data

Generated queries should run inside your cloud data warehouse, against current tables, with no extract step. To test it, update a row in the warehouse, refresh the dashboard, and confirm the change appears. A generator working from an extract won't show the change until the next export.

Permissions applied to every viewer

Open the same dashboard as two users whose permissions should produce different results. Each should see only the rows and columns their permissions allow. If both see the same restricted rows, check whether queries run on the builder's credentials, a shared service account, or a copy.

Builds from your data models and metrics

Choose a generator that starts from the data models and governed metrics your team already trusts. A generator that writes raw SQL against undocumented tables has to guess what "revenue" means, and defined metrics remove much of that guessing.

In a small, vendor-run semantic layer benchmark, semantic layer configurations scored 100% on in-scope questions. Raw text-to-SQL scored about 51% to 63% on the same questions, and adding data models raised accuracy for both approaches.

Output you can keep editing

You should be able to trace each generated element to the SQL behind it and edit it on its own. Check whether you can change one chart's query, move a filter, or switch to manual editing without losing the rest of the build. During the proof of concept, ask for a one-chart change and confirm nothing else on the page moves.

How Sigma supports AI dashboard generation

Sigma Assistant builds dashboards inside Sigma, and Send to Sigma brings in dashboards prototyped in Claude. Both paths end in an editable Sigma workbook that queries your warehouse live.

Sigma Assistant in plan and build modes

In plan mode, you describe the dashboard and Sigma Assistant proposes a plan you can adjust. Build mode selects the right data sources and builds charts, tables, KPIs, input tables, and actions inside a workbook. It works in one shot or turn by turn, applies your row-level and column-level security, and validates queries before they execute. Your organization needs a supported AI provider configured, and your account needs permission to use Assistant and edit workbooks.

The output consists of standard workbook elements, so you can keep prompting the assistant or switch to the drag-and-drop UI at any point.

Prototype in Claude and send to Sigma

If a question starts in Claude, Send to Sigma moves the dashboard into a workbook with its SQL and your existing permissions intact. It requires a Sigma Build license and a paid Claude plan, and it runs through the Sigma MCP Server, which lets Claude and ChatGPT query governed data. Once in the workbook, you can change one chart at a time and extend the dashboard into an app.

Teams that manage workbooks in code can follow the same prototype-then-refine pattern with Workbooks as Code (Beta). Treasury4 used it with AI to cut dashboard prototype time by over 90%.

Generate dashboards with the Sigma MCP server and Workbooks as Code

If your team already works in a coding agent, you can build Sigma dashboards from there too. The Sigma MCP server lets Claude, Cursor, or Codex search your Sigma content, read your data models, and build a workbook from a plain-language request. In Sigma’s own walkthrough, Claude designed a data model, built the workbook, and added an input table for approvals in about seven minutes.

Workbooks as Code is what makes that possible. It describes a workbook as a structured spec that an agent can read and write. When I built a scenario planning app this way, a Sigma Skill had the agent ask me about the app type, the editable fields, and the approvals before it drafted a plan. The whole build took eight minutes, and I covered the details in a post on building governed apps with a coding agent.

What comes out is a normal workbook. It queries your warehouse live, applies your permissions, and opens in the Sigma editor, so you can run the same one-chart-change test from earlier. Treasury4 used Workbooks as Code with AI to cut dashboard prototype time by over 90%.

Migrate an existing dashboard with Sigma Migration Skills

The same approach works for dashboards you already have. Point Claude Code or Codex at a Tableau workbook, and the Sigma Migration Skills read its data model and calculations, rebuild it in Sigma through Workbooks as Code, and check the results against the source value by value. Skills exist for Power BI, Qlik, ThoughtSpot, and other tools too. The agent stops to ask for your input at key decision points, so review its plan the same way you’d review any generator’s.

Simple apps with input tables, forms, and actions

Sigma calls the simple apps you build this way AI apps. Users can edit data in workbook input tables and write it back to the warehouse, or submit entries through forms inside the workbook. An audit trail logs each edit and its author. New data writes to a schema your admin designates, so source data isn't overwritten. After a submission, actions chain approvals and notifications.

Generate your first AI dashboard with Sigma

Start with one recurring question your team still answers with a spreadsheet export, then follow these steps:

  1. Pick a data model your data team already endorses, so the metrics match what finance and leadership see.
  2. Open Sigma Assistant in plan mode and describe the dashboard, including the metric definitions, time range, grain, and filters.
  3. Review the plan, correct any wrong source or definition, and then build.
  4. Refine individual charts with follow-up prompts or by hand.
  5. Before you share it, have a colleague with narrower permissions open the workbook and confirm they see only their slice.

From there, the dashboard can grow into an app that writes decisions back to the warehouse. Get a demo to see both paths on your own warehouse.

FOLLOW SIGMA

Related articles

Get started