Give customers governed AI analytics

AI-powered embedded analytics is analytics inside a software product where end users can ask questions in plain English and get charts or answers back, alongside the usual dashboards. In Databrain, a customer types a question in the embedded dashboard, Databrain generates SQL against a modeled datamart and its semantic layer, runs it through the same tenant-scoped query path as any dashboard, and returns a chart the user can keep as a saved metric, available on the Enterprise plan.

Both turn a question into SQL; the difference is the context the model works from. Plain text to SQL hands a model your raw schema, while Databrain plans against a curated datamart and the semantic layer your team maintains: table and column descriptions, synonyms, relationships and examples. The generated SQL then runs through Databrain’s normal tenant-scoped query path. Answer quality still depends on how complete that semantic metadata is, so it is worth maintaining like code.

Databrain uses the LLM provider you configure with your own API key: OpenAI, Anthropic Claude, Azure OpenAI, Google Gemini, Llama or Mixtral. You pick the provider and model for each workspace, so model requests run under your account and that provider’s data terms. That helps when your customers’ procurement teams have a policy on which model may touch their questions. Chat Mode and AI metric summaries both use the connector you select.

Databrain runs AI-generated SQL through the same tenant-scoped query path as dashboards, so the model is never trusted to write tenant-safe SQL on its own. Chat and follow-up questions apply the same row-level security predicates as dashboards for database-level and table-level tenancy, and customer variables in custom SQL resolve before the query executes. You also decide where chat appears, through Data App access settings and the permissions on each guest token.

Yes, Databrain’s Chat Mode needs a datamart with semantic metadata before it can answer well. The prerequisites are an LLM connector, a workspace connected to that datamart, and descriptions, synonyms, relationships and examples for its tables and columns. A bare datamart still runs hand-built charts, but without that context the model has to guess which tables and joins you mean. Most weak AI answers trace back to thin semantic metadata.

Yes, Databrain’s chat appears inside the embedded dashboard itself, not on a separate page or URL. When you enable it for a Data App and the user’s guest token allows it, your customer can type a question, get a chart or table back, ask follow-ups that reuse recent context, and save a useful result as a metric, available on the Enterprise plan. EpochOS uses it to give mortgage brokers natural-language access to their data inside its product.

Yes, Databrain ships an MCP server for your developers, with 36 tools for tasks such as listing tenants, creating embeds, managing metrics and maintaining the semantic layer. It runs locally over stdio in clients like Cursor, Claude Code, Claude Desktop, Codex and Windsurf, and production write actions default to a dry run that needs explicit confirmation. It is a builder tool for your team, not an endpoint your customers’ agents call.