Skip to main content
Basedash answers questions by running real SQL against your data, and it shows you every query it ran. This page explains how answers are generated, how to check them, and what the AI can and can’t do.

​
How answers are generated

When you ask a question in chat, Basedash gives the AI:
  • The schema of each data source you can access, including table and column descriptions from your database or dbt
  • Your organization’s models, with their measures, segments, and verification status
  • AI context for your organization, your groups, and you
  • The names of your skills, which the AI loads when they’re relevant
  • Existing charts and dashboards it can reuse
The AI then writes SQL, runs it against your live data, reads the results, and repeats until it can answer. Numbers in an answer come from those query results, not from the model’s memory.

​
Check the work

Every answer shows what the AI did:
  • Progress steps show each action while the AI works, such as “Reading Revenue model” or “Reading Activation rate skill”.
  • Sources lists every query, with the data source, the row count, a preview of the results, and the full SQL.
  • Charts keep their SQL, so you can open the chart editor and change the query.
If an answer looks wrong, check the SQL in Sources first. Most wrong answers come from a wrong assumption, such as which table holds revenue or which statuses count as active. Fix it for next time with a model or AI context.

​
Use verified definitions

Models are the most reliable way to make answers consistent. When a model defines a metric like monthly recurring revenue, the AI queries the model instead of writing its own version. Admins can mark models as verified, and the AI prefers verified models when more than one could answer a question. Use AI context for terminology and rules, and skills for repeatable analysis methods. If you tell the AI something in a conversation that conflicts with a skill, your instruction wins.

​
Read-only by default

The AI’s regular queries are read-only. It can only change data in one case, and only with your approval:
  • The data source is Postgres, Redshift, or Supabase.
  • You have Can edit access to the data source. See data source access control.
  • The data source has no schema-, table-, or column-level access rules.
  • You’re chatting in the Basedash web app. Changes can’t be made from Slack, the API, or automations.
When those conditions are met, the AI can propose a single INSERT, UPDATE, or DELETE statement. Basedash checks the statement, then shows an approval card with the data source, a description of the change, and the exact SQL. Nothing runs until you click Approve and run. Every approved change is recorded in the audit log. To keep a data source strictly read-only, give members Can view access only, or connect it with a database user that has read-only permissions.

​
Your data access applies

The AI works with the same data access as the person asking: Insights and automations run without anyone present, so they use the data access of a specific member. Each page explains which one.

​
How your data is handled

Basedash sends table and column names, AI context, and query results to its AI providers to generate answers. Neither Basedash nor its AI providers use your data to train models. See Security and compliance.

​
FAQ

Every number comes from a query that Basedash ran against your data. Open Sources under the answer to see each query and its results. Basedash also evaluates every new AI model release against its own benchmark before using it.
Create a model for the metric and have an admin verify it, or add AI context to the table that explains what it contains and when to use it.
No. Regular queries are read-only. Changes need an approval in the web app, edit access to the data source, and a supported database. See Read-only by default.
No. The AI uses your data access, including object-level access rules and row-level security.