AI in Chartizer
What the assistant does, where it appears, and the rule it always follows.
Chartizer uses AI for the tedious parts of data work: guessing which fields mean the same thing, describing what a dataset contains, and turning a question into a chart.
The rule
AI suggests, humans decide
Nothing the assistant produces runs or lands on its own. Every query it writes is shown to you and executes only once you accept it, and mappings, schema changes, and dataset edits all wait for a person too. A wrong query wastes a source’s capacity and a wrong mapping silently corrupts every join built on it, so neither happens unreviewed.
Where it appears
Investigations
Proposes a plan and a query at each step, then — once you accept — reports a finding backed by charts and the queries behind them.
Ask your data
Ask a question in plain language, get a chart plus the query behind it.
Mapping suggestions
Proposes which fields across sources mean the same thing, with confidence and evidence.
Dataset summaries
Generates a plain-language description of a new dataset and its fields.
Anomaly notes
Flags unusual movement in a scheduled dataset — a row count that halved, a spike in nulls.
What it sees
The assistant works from metadata plus a small sample, not your full tables.
| Input | Included | Notes |
|---|---|---|
| Field names and types | Always | The primary signal |
| Dataset and source names | Always | |
| Sampled values | Up to 50 rows | Used for mapping and type inference |
| Full table contents | Never | |
| Query results | Only for Ask your data, in your session | Not retained |
Sampled values leave your database
For the Cloud runtime, samples are processed by the model provider. If that is unacceptable for a dataset, disable sampling on it — see Data handling.
Turning features off
AI is per-team and per-feature, under Settings → AI.
| Setting | Default | Effect when off |
|---|---|---|
| Mapping suggestions | On | No suggestions; mapping stays fully manual |
| Ask your data | On | The Ask box is hidden |
| Dataset summaries | On | Descriptions stay empty until written by hand |
| Anomaly notes | Off | No anomaly entries in the activity feed |
| Value sampling | On | Suggestions use names and types only, less accurately |
Team Admins control these; individual members cannot re-enable a disabled feature.
Confidence
Every suggestion carries a score and its reasoning.
| Score | Meaning | Typical action |
|---|---|---|
High (≥ 0.9) |
Names, types, and sampled values all agree | Review and accept |
Medium (0.7–0.9) |
Two of three agree | Check the evidence |
Low (< 0.7) |
Weak signal, shown for completeness | Usually reject |
Reviewing rejections is worth the time
Rejections are fed back as team-specific signal. Rejecting id → dealer_id once stops it
being suggested repeatedly.
Limits
- Suggestions are proposed when a dataset first appears and on schema change, not continuously.
- Ask your data reads datasets you already have access to — it cannot widen your permissions.
- The assistant does not write to sources. Every connector is read-only.