Skip to main content
At the end of this page, you can find your agent’s most costly issues on the Improve tab and trace each one to the conversations behind it.

Before you start

Before you start, sign in to the dashboard and select the organization you want to work in.
  • You need a role with analytics access.
  • You need conversations in the selected range. Without them the ranked issues card shows its empty state.
  • Read Analytics overview for the scope bar, the freshness stamp, and the Conversations drawer. This page does not repeat them.

Open the Improve tab

1

Open Analytics from the sidebar

Select the Improve tab. It sits between Goals and Team.
2

Set the scope

Pick a range such as 28d, pick a deployment or keep All deployments, and choose Live or Test.
3

Read the headline

The 01 Ranked issues card opens first. Its headline reads Issues to review with a count.

What counts as an issue

The Improve tab does not run a separate report. It reads the same data as the other tabs, then compares the selected range with the previous window of the same length. Five kinds of signal become issue cards, and each kind needs a small minimum count before it appears: The no-answer reasons are No knowledge match, Low confidence, Out of scope, and Explicit can’t help. Tool failures count failed calls and timeouts together.

Read the ranked issues

Under the count, a narration line states how many issues are high-severity and explains the ranking. Issues rank by frequency and by how sharply each moved versus the previous window. Only the top three issues show by default. Select Show all to list the rest, and Show fewer to collapse the list again. Each issue card carries: The change phrase in the detail line is one of four. It reads new this window, a multiplier such as 3× the 4 last window, down from 12 last window, or steady vs 10 last window. When nothing crossed the threshold, the card reads No anomalies this window with the hint that no tool failures, no-answer moments, or escalation surges crossed it. If part of the feed fails, the card reads Issues feed is partially degraded, names the missing piece, and offers Retry. If nothing loads, the card reads Couldn’t load the issues feed.

Open the detailed diagnostics

Below the ranked issues sits a card titled Detailed diagnostics. Its description says to open the underlying quality, tool, handoff, and flow reports when an issue needs deeper investigation. Select Show diagnostics to expand it, and Hide diagnostics to collapse it. Six numbered cards appear.

02 No-answer reasons

Each row is one reason with a bar, a count, and the number of affected conversations. The card header shows the total as moments. Select a row to open the drawer for that reason. The empty state reads No no-answers recorded in this window and warns that an empty list can also mean incomplete reporting.

03 Tool reliability

A table with columns Tool, Calls, Call success, Business result, Failures, and Latency p50 → p95. The Business result column shows whether the external system accepted, rejected, or partly completed the action, for example 2 rejected. Select the red count in Failures to open only the failed calls, or select the row to open every conversation that called the tool. With no calls, the table reads No tool calls in this window.

04 Agent errors & negative feedback

Two daily bar charts side by side. The Agent errors chart has a View conversations link that opens Conversations with agent errors. The Thumbs down chart has an ↗ explore link that opens the same metric on the Explore tab. An empty chart reads No agent errors in this window or No thumbs down in this window.

05 Escalations

A daily bar chart of escalations, with a View conversations link in the header that opens Escalated conversations.

06 Where flows leak

A per-gambit table for one published flow, with columns Gambit, Entered, Fwd, Dropped, Back, Drop %, Escalated, and Dwell p50 → p90. Every count is a drill. Select the red Dropped count to see the conversations that were abandoned at that gambit. The card header names the flow version and the agent it belongs to. With All deployments selected the card reads Pick a deployment to see flow health, because flow health is per agent. Pick one deployment in the scope bar to populate it. A range with no flow activity reads No flow activity in this window.

07 Unanswered questions

Recurring end-user questions the agent could not answer, grouped into topics with a conversation count each. This card runs on its own schedule. Its header reads Fixed last-7-day window (ignores the date range above) · generated with a relative time, so the scope bar does not change it. Topics generate once a day. Before the first run, the card says that topics generate daily and the first run is pending. When the run found nothing, it reads No recurring unanswered questions with the time it was checked. Each topic lists Example 1, Example 2, and so on. Select one to open that conversation in the trace inspector, which needs live chat read access.

Act on what you find

Start with the top-ranked card, not the newest one. The ranking already weighs how many conversations an issue touches and whether it is growing. Then use the drill and the diagnostics to pick the fix:
  • A tool failure card points at a toolkit or a tool gambit. Open the failed calls, read the trace, and repair the connection or the mapping.
  • A No knowledge match or Low confidence card points at the knowledge base. Add the missing content that the traced conversations asked for.
  • An Out of scope card and the Unanswered questions topics show what end users expect the agent to cover. Decide whether to add a flow or a guardrail.
  • A Dropped count in Where flows leak points at one gambit. Open the abandoned conversations and simplify that step in the builder.
  • An escalation or thumbs-down spike is a signal to read transcripts before you change anything.

Verify

Open an issue card and select View conversations →. The drawer lists the matching conversations, and View trace on a row opens the step-by-step trace.