AI, agents and insights
AI features that work on governed data, with people deciding what the AI may change.
For the status of each feature, see Capabilities at a glance.
Features#
Agents have a goal, a set of tools and an autonomy level. Their tools read, create and update records, draft documents, notify record owners and send email.
People stay in control: Agents are read only by default. They can ask before each change, or act on their own where their owner allows it. Email to an outside address always waits for a person, and a request left unanswered for 15 minutes is declined.
Discover shows how a workflow actually ran. Impact previews how a change spreads to related records. Process mining shows paths and variants. Agent runtime replays a mined workflow as an agent.
People stay in control: The runtime does a dry run by default, and a policy blocks any action that isn't explicitly allowed.
A panel on every console page. It answers questions and can act as you on records, documents and runs.
People stay in control: Deletes wait for your confirmation. An admin can switch off its access to workspace tools.
Describe a workspace view or form in plain language and the AI builds it, with preview, undo and redo.
People stay in control: Nothing is saved until you apply it. The AI never sees record values, and cannot publish, share or run business actions.
The AI generates dashboard screens from a description. They show supplied data, not live queries yet.
People stay in control: New dashboards stay switched off until someone previews and enables them.
Computes metrics, anomalies and forecasts from approved calculation plans.
People stay in control: A second admin must approve each plan before it runs.
Who controls agents#
| Action | Roles |
|---|---|
| Run agents | Operator, contributor, admin |
| Approve agent actions | Approver, contributor, admin |
| Create and manage agents | Contributor, admin |
Model choice#
An admin sets one AI model for the whole deployment under Settings → LLM: OpenAI, Anthropic, Ollama, or any OpenAI-compatible endpoint, including a self-hosted model.
Without a model, AI features say so and change nothing. Mapping suggestions fall back to name matching. Discover and process mining don't need a model.