The learning engine analyses completed conversations, identifies what's working and what's not, and proposes specific changes — with supporting evidence. Nothing changes until you approve it.
The pattern detector analyses completed conversations and classifies findings by evidence strength before generating any proposal. Low-confidence findings are labelled clearly — no finding is dressed up as more certain than it is.
Approving a proposal creates a new campaign version automatically. All new conversations run under the new version. The prior version is preserved — one click to revert.
Every campaign change — manual or from a proposal — creates a versioned record with a timestamp, the user who made it, and a change summary. The full history is visible from the campaign detail page.
How it works: Conversations started before a change are linked to the prior version — attribution is always accurate.
If a change makes performance worse, any prior version can be restored instantly. Reverting creates a new version record so the revert action itself is auditable.
How it works: All new conversations switch to the reverted config immediately on revert.
Versions created from proposals include a link back to the originating proposal. You can see the evidence and reasoning that justified any change — even after the proposal is marked applied.
How it works: The full proposal detail — evidence class, metric projection, diff — is always accessible from version history.
The learning engine reads from the database and writes only to the proposals table with status PENDING. It has no write access to campaigns or live conversations — it cannot affect anything in flight.
How it works: Approval is the only action that updates campaign configuration — and only a human can approve.
We'll show you a real proposal — evidence, diff, and the version it produced.