Activity Analytics: overview tiles, the AI Insight Yield section, and thumbs-down reasons linked to the answers that earned them. With narration.
Six months into an AI rollout, someone senior asks whether it is working, and the room reaches for anecdotes. The demo above shows what it looks like when the answer is a number instead.
The feedback loop, end to end
In SWIRL, every AI Summary and every Assistant message carries rating controls. A thumbs up is one click; a thumbs down asks for a reason. Those ratings are not telemetry that disappears into a log. They persist as first-class records, joined to the exact generated answer, the model that produced it, and the query that started it.
From ratings to yield
The Activity Analytics page aggregates the ratings into AI Insight Yield: the percentage of AI answers that earned a positive rating. It is split by RAG versus chat, trended daily, and sits alongside the operational numbers you already watch, searches, unique searchers, and errors, so AI quality is reviewed the same way uptime is.
The part that makes it actionable
A percentage alone tells you that something is wrong, not what. The thumbs-down table closes that gap: each negative rating shows its reason and links back to the specific answer that earned it. "The AI is wrong sometimes" becomes a triage queue where each item has a concrete fix, correct the source document, adjust the prompt, or swap the provider.
Watch the number move
In the demo, ratings from that morning's sessions land while the dashboard is open, and overall yield climbs from 20% to 43%. That is the operating loop in miniature: measure, diagnose, fix, re-measure. Teams that run this loop get a defensible answer to "is it working?" and, more importantly, a system that is actually improving week over week.
Activity Analytics is covered in the Admin Guide. To see your own yield number, talk to us.