How to Set Up and Use Metabase Metabot

Metabase Metabot translates plain English into data queries, SQL, and chart summaries without requiring users to write code or navigate a query builder. This guide covers everything needed to purchase the add-on, configure it correctly, prepare your data model, and use it effectively.

What Metabase Metabot Does

Metabase Metabot is an AI assistant built into Metabase that converts natural language prompts into SQL queries, charts, and data summaries. It addresses the most common blocker in self-service analytics: most team members need data insights but cannot write SQL or build complex queries.

When a user asks a question, Metabot first searches existing dashboards and saved questions for a matching answer. If no match exists, it builds a new chart using the query builder, or in the SQL editor, writes a query inline. It can also analyze existing charts and produce written trend summaries, and it can debug failing SQL queries automatically.

Metabase reports that over 90,000 companies use the platform. Metabot is specifically designed for organizations where most users are not data engineers but need regular access to operational and business data.

Requirements Before Starting

Metabot runs on Metabase Cloud only. Self-hosted Metabase instances do not support the full Metabot experience.

If you run a self-hosted instance, you can still access SQL generation by connecting your own OpenAI or Anthropic API key in Admin > AI. This enables natural-language-to-SQL in the SQL editor but does not include the chat interface, chart creation from natural language, or AI exploration mode.

For the full Metabot experience, you need:

• A Metabase Cloud subscription on any paid plan
• Admin access to your Metabase instance
• At least one connected database

Step 1: Add Metabot to Your Plan

Metabot is a paid add-on, not included in base Metabase Cloud pricing. Purchase it through the Metabase Store.

1. Go to store.metabase.com and sign in. Note this uses a separate account from your main Metabase login.
2. Go to the Instances tab and find your instance.
3. Click "Add Metabot AI."
4. Choose a plan based on expected monthly request volume. Each message any user sends to Metabot counts as one request.
5. Complete checkout. Metabot activates on your instance within a few minutes.

After activation, go to Admin > AI to confirm the Metabot toggles are visible and enabled.

Step 2: Configure Metabot in Admin Settings

In Admin > AI, three settings have the largest impact on result quality.

Internal vs. embedded toggle. Both default to enabled. If you have Metabase embedded in a product and are not ready for end customers to access Metabot, disable the embedded toggle while you run internal tests.

Verified content restriction. Available on Pro and Enterprise plans. When enabled, Metabot only draws from verified models and metrics rather than raw database tables. This substantially improves answer accuracy for teams with a well-structured data model. Enable this after your key models are verified.

Collection scope. Controls which collections Metabot searches by default when a user opens an AI exploration session. Useful for keeping development or staging collections out of the default scope. Users can still access any collection by @-mentioning it directly in a prompt.

A fourth setting worth configuring is prompt curation. In collection settings, you can add suggested questions that Metabot surfaces to users when they open a new AI exploration. This reduces friction for non-technical users who are unsure what to ask.

Step 3: Prepare Your Data Model

This step is optional for getting started but is the single biggest factor in Metabot answer quality. Teams that invest an hour or two here before rollout see noticeably better results than those relying on prompt quality alone.

In the Metabase data model editor, work through each connected model and add the following:

Field descriptions. Any field with a non-obvious name needs a description. "rev_amt" alone tells Metabot nothing. "Total order revenue in USD, after discounts and before tax" gives it real context to work from.

Semantic types. Set the correct semantic type for each field: Date, Currency, Latitude, Longitude, ID, and so on. Metabot uses these types to form correct aggregations, apply appropriate filters, and avoid grouping numeric IDs as if they were metrics.

Model-level descriptions. Add a short description to each model explaining what it represents and how rows are defined. For example: "One row per confirmed order. Excludes drafts and abandoned carts. Refreshes nightly from the orders service."

Glossary terms. In Admin > AI > Glossary, define any company-specific terms Metabot would not know from context. If "CAC" means customer acquisition cost from paid channels only at your company, define it there.

A practical approach is to prioritize the five to ten models that non-technical users ask about most often. Start there before working through lower-priority tables.

Step 4: Use Metabot in Practice

Metabot integrates into four places in the Metabase interface.

AI exploration. Go to + New > AI exploration to open a blank chat session. Ask any data question in plain English. Metabot returns a chart, table, or written answer depending on what the question calls for. Ask follow-up questions in the same thread to refine results. This is the recommended starting point for non-technical users.

SQL editor. Open the SQL editor and press Cmd+E (Mac) or Ctrl+E (Windows) to open the Metabot sidebar. Describe the query you need and Metabot writes the SQL inline. You can edit before running. Useful for analysts who know SQL but want to speed up the initial draft.

Error debugging. When a query fails, Metabot shows a "Fix with Metabot" option alongside the error message. It explains what went wrong and rewrites the query. This is particularly useful for non-technical users who have copied a query from elsewhere.

Chart summaries. Open any chart and click the Metabot icon in the top right corner. Metabot generates a written paragraph describing what the chart shows, any notable trends, and visible anomalies. No data is sent outside Metabase for this feature; summaries run against the already-loaded chart data.

Getting Better Results

A few practical patterns produce more accurate answers.

Name the model or table in your prompt. "Show me orders by region" produces less reliable output than "Show me the Orders model grouped by shipping region for 2025." Specificity reduces incorrect table guesses.

Use @-mentions to anchor prompts. In AI exploration or the SQL editor, type @ followed by a model or table name to reference it directly. This bypasses the default 100-table scan Metabot performs when no table is specified, which matters on large-schema databases.

Break complex questions into parts. Metabot handles single-level aggregation and grouping reliably. For multi-step analyses such as funnel calculations or cohort retention, ask for each component separately and combine them manually.

Use English prompts. Metabot is optimized for English and produces less reliable output in other languages.

Limitations to Know

Metabot does not support window functions, CTEs, or complex subqueries. Any query requiring those patterns needs to be written or extended manually in the SQL editor.

When no table is specified, Metabot scans the first 100 tables in the connected database. Instances with more than 100 tables should use explicit table names or @-mentions to avoid mismatches.

Treat all Metabot output as a starting point. Verify results against known figures before using them in reports or decisions, especially in regulated industries.

A Note on Teams Without an Existing Data Stack

The setup above assumes a working data warehouse and a Metabase Cloud subscription. For teams starting from a spreadsheet or file upload, VSLZ AI lets users query data in natural language without any warehouse setup or model configuration required.

Practical Summary

Metabase Metabot is a capable AI layer for teams already running Metabase Cloud. The full setup involves four steps: purchasing the add-on through the Metabase Store, enabling and scoping it in Admin > AI, adding field descriptions and semantic types to your data model, and testing with a sample of prompts before wider rollout. The model preparation step is the highest-leverage investment. Teams with well-described data models consistently get better Metabot answers than those that skip it.