AI assistant
Ask questions in plain English and get governed answers — grounded in discovered schemas, with every write proposed for your approval first.
The assistant is a chat panel over your organization’s data. Ask “how did revenue do last week?” and it resolves the answer from the measure catalog and the dataset schemas it has actually read — then shows its working. It is generally available on every tier.
What it can read
The assistant works through a fixed set of read tools, and only those:
| Tool | What it does |
|---|---|
| List datasets | The datasets your permissions let you see |
| Get dataset schema | Column names and types, read from the real table |
| Run SQL | A single SELECT over the lakehouse — never DML, never DDL |
| Chart preview | A chart computed from governed measures, rendered in the reply |
| Quality status | The current pass/fail state of a dataset’s quality rules |
| List saved charts | Your existing charts, so answers can reference them |
| Pipeline node types | The task types available when it drafts a pipeline |
Two properties follow from this design:
- No fabricated names. The assistant queries schemas it has discovered through these tools. It cannot invent a dataset or column and query it — the query would simply fail, and the system prompt forbids guessing.
- Engine numbers, model prose. Numeric answers come from queries the engine actually ran. The model shapes the sentence around the numbers; it does not produce them.
Propose, then confirm
The assistant cannot write anything directly. When you ask it to create something, it calls a proposal tool and the result appears as a card in the chat for you to review:
| Proposal | What you approve |
|---|---|
| Save chart | A chart config over governed measures |
| Create dashboard | A dashboard composed from saved charts |
| Set quality rules | Validation rules on a dataset |
| Create quality schedule | When those rules run |
| Create pipeline | A DAG built from the real task types |
Nothing is applied until you confirm it. Declining a proposal discards it with no side effects.
Scope and safety
- Your permissions, not more. The assistant acts as you, inside your organization. It sees exactly the datasets, charts and rules your role allows.
- Tool results are data, never instructions. Content coming back from tools — including data someone else loaded — is treated as material to analyze, not commands to follow.
- Conversations stay in your org. Chat history is stored with the same org isolation as everything else.
Related AI features
- AI text cards — dashboard cards whose numbers are engine-computed measure values; the model writes only the narrative around them.
- AI summary pipeline task — writes a summary of a chart’s or dashboard’s current data as a pipeline step, exposed to later tasks as
{{meta.summary}}.
Billing
AI usage is included with an allowance on every tier. On Professional you can bring your own OpenAI API key, which makes assistant usage unmetered.