You can also browse, search and filter all release notes directly in the application → Open release notes in TAI.
Analytics: see how much of your agent usage is chat versus reports
New Feature.
A new Channel filter on Analytics narrows the summary cards, the over-time charts, and every breakdown to the channels you pick — chat, reports, the API, or imported legacy Copilot activity.
A new Usage by channel breakdown ranks those channels by how much activity came through each, so you can see at a glance where your agent runs are coming from.
Preview Knowledge search as one specific agent would run it
Improvement.
Knowledge > Search can now be scoped before you run it: the whole catalogue (as before), a single knowledge collection, or one agent — which searches exactly the sources that agent can reach, so you can check what it would actually find.
Each result now shows how relevant it is and whether it matched on meaning (Semantic), wording (Keyword), or both. An empty result explains itself too, distinguishing an empty scope, sources still waiting to be indexed, and sources that are indexed but hold nothing relevant — naming the closest content it found in that last case.
Analytics now counts only usage TAI meters
Fix. Only Company Admins.
Planner Copilot and Customer Care Copilot activity briefly appeared in your Analytics. Those Copilots run outside TAI and their usage is not yet covered by TAI's consumption accounting, so your Analytics figures included activity your credits never did. It is hidden again until the two agree, and will return once they do — with the full history, not just from that point on.
Report runs no longer hang at Running when you close the tab
Fix.
Closing the tab or navigating away while a report run was streaming could leave it stuck showing Running, even though it had actually stopped — it only cleared up on its own half an hour later. That run now resolves to Failed immediately, so the library reflects what really happened. Also fixed: a report, run, or schedule URL ending in a non-id segment like /reports/all used to render "not found" — it now takes you to the matching list instead: the report library, or the schedules list for a scheduled-report URL.
Report delivery now warns when an empty allow list blocks everyone
Fix. Only Company Admins.
An empty allow list looked harmless before. It now warns that no one can receive scheduled reports until you add an entry.
Knowledge search now returns only genuinely relevant matches
Fix.
Knowledge search always returned its nearest matches, however distant — so a query with no real connection to your content still came back with something. A relevance threshold now applies, and a search that truly finds nothing says so.
This covers both places knowledge is searched: the preview under Knowledge > Search, and the retrieval your agents do while answering. Agents therefore stop quoting a loosely related source when nothing in their knowledge actually addresses the question — they say they don't know instead. If an agent now misses something you expected it to find, the preview will show you how close the nearest content was.
Advisors: clearer error states and a recoverable Apply
Fix. Only Company Admins.
A round of fixes to the Advisor surfaces, all in the same direction — the screen should never tell you something it does not know.
Watching a running evaluation survives a network blip instead of freezing.
Applying a proposal that does not go through can be retried, and Apply is offered only where it works.
A run that failed to load, a partly-unreadable reply, and an agent whose runs or conversations could not be fetched each say so, rather than presenting as empty or complete.
A failed run shows why it failed, not where it had got to.
Analyzing a report run is offered once the run has finished.
Agents no longer present a cut-off row as complete
Fix.
When an agent checks a list of records, a row whose answer ran long used to arrive with its last columns silently empty — it looked complete. Now the agent retries the row, and if it still does not fit it marks the row as cut off, keeps what it did retrieve, names the records affected and offers to re-check them.