The Learning screen

Learning (Advanced drawer, admins only) is where the corrections your team makes become visible — and where you accelerate them.

The feedback table

Every re-classification and review resolution is recorded here as a row: when, from whom, “InboxAgent said” (the AI’s original read) vs “You said” (what your team decided). Filter by date, search by sender or subject, or tick Show wrong only to focus on the corrections. Clicking a row’s subject opens the underlying email’s detail.

The KPI strip along the top — Total Feedback, InboxAgent correct, InboxAgent wrong, Accuracy — gives you the running score. One reading note: review-queue resolutions record the action your team took, so a “You said” cell can show an action (like a resolution type) rather than a category, and those rows count on the “wrong” side of the tally. The accuracy trend across weeks is the meaningful signal, not the absolute number.

The Confirm button

Confirm on a row means “yes, this correction was right.” A confirmed correction counts double toward generating suggested rules (how learning works) — it’s the single cheapest way to speed up InboxAgent’s improvement. A five-minute weekly pass — confirming the corrections that were genuinely right — compounds nicely.

Corrections made via Re-classify arrive pre-confirmed (you actively fixed the record), so you won’t find a Confirm button on those — they’re already at full weight.

What this screen doesn’t do

Nothing here changes routing directly. Confirming feeds the learning signal; standing behavior changes still come from rules you create and suggestions you choose to act on — you stay in control.