Version 5.8

AI Agents Version 5.8

17 September, 2026

• Custom Dashboards:
  Charts can now be built by describing them to an AI agent and arranged on a dashboard of your own, over conversation data, AI Insight and Custom Attribute values, data table records and flow execution results. They live under Dashboards > Custom (the existing dashboard becomes Overview), where a date range and channel filter apply to every chart, charts are dragged and resized into place, and hovering a chart offers "Configure Chart" to revise it with the agent instead of rebuilding it. Chart types include a Sankey, which shows the volume flowing between two fields such as topic to resolution. Custom Dashboards needs its LLM settings filled in under tenant settings before charts can be created, and reporting data runs one day behind, so today is not offered in the date filter.

• Knowledge Base Tags:
  Documents, websites and FAQ rows can now be tagged and retrieval narrowed to those tags, so one knowledge base can serve several brands, products or regions instead of being split into separate ones. Tags are defined per knowledge base in its Tags section and applied with "Add tag" on each row of the Add Data Source screen, while FAQ files carry a tag per extra column and offer a downloadable template that already has the correct column headings. A "Tags" section in the Knowledge Retrieval and RAG Agent nodes then filters retrieval on fixed values or values taken from the conversation, and a tag that is in use by a document or a flow cannot be edited or deleted until it is removed from them.

• Static or AI-Generated Guardrail Block Messages:
  A guardrail that blocks a turn no longer has to reply with the same fixed sentence every time, for content filter blocks and for the PII rule in blocking mode alike. In the guardrail wizard's General Settings step, "Input Block Message Mode" and "Output Block Message Mode" each offer Static or LLM and are set independently; choosing LLM replaces the block message with an optional instruction field that steers the wording. Guardrails created before this release stay on Static in both directions, and a generated response uses the AI agent's own model and provider settings rather than the configuration chosen for content filters.

• Collecting Multiple Values in a Custom Attribute:
  A Set Attribute row for a key whose output type holds multiple values can now add to what the conversation has already collected instead of replacing it, so two values set at two points in one conversation both reach the Custom Attributes screen, the exports and the reports. An "Append" / "Overwrite" switch appears beside the Value heading when a list type key is selected, and it opens on Overwrite for new rows and for flows built before this release, so existing behaviour is unchanged. An attribute keeps at most 10 values per conversation, and a key with more than three values is shown as the first three with a "+N" control that expands the rest.

• AI Insights:

  • Re-running Insight Generation After a Failure:
    When insight generation fails for a conversation, it can now be started again from the conversation itself, so a conversation is no longer left out of analysis because of a transient error. A "Generate AI Insights" button appears under the AI Insights heading on the Conversations detail page, only for conversations whose generation failed, and the values are listed in place as soon as the run succeeds. The button is unavailable while redaction is running or while the add-on AI settings it needs are missing.
  • Insights and Custom Attributes in the Daily Conversation Report:
    The Daily Conversation Report API now returns each session's AI Insight and Custom Attribute values, so conversations can be reported in your own warehouse at the richness the product shows, without a second call. The two arrive as separate lists, each entry carrying a name, an output type and its values, and as empty lists when nothing was produced for that session. The payload also adds session fields that were stored but never reported (conversation topic, end time, live agent handover count and client id), a readable name beside each coded value such as the resolution type, and the AI agent id and name on each bot message, filled for new messages only.

• No Repeated Questions After an Interruption:
  When an end user talks over a voice agent, the interrupted answer is now carried into the next request instead of being dropped from it, so the agent keeps what it was saying when it was cut off. Together with the user's next message, that lets it work out how far it got, and it stops asking the customer a question it had already asked before the interruption. Nothing needs to be configured, and conversations without an interruption are unaffected.

• Clearer Prebuilt Tool Descriptions:
  The descriptions of the prebuilt agent tools have been rewritten so an agent picks the right tool more often, and the change reaches existing agents without any edit. The call and chat transfer tools are now described as escalating the conversation to a human agent, and the end call and end chat tools state that the conversation ends completely once they are used.

• More Realistic Simulated Users in AI Testing:
  The simulated user in AI Testing now carries its persona across the whole conversation instead of exaggerating it on every turn, so a test run reads like a real conversation and its results are easier to judge. Scenarios and personas do not need to be changed; existing ones pick up the new behaviour on their next run.

• Minor UI Improvements:
  Several minor improvements have been made to the user interface to
  enhance user experience and usability.

• Minor Bug Fixes:
  Various minor bugs have been addressed to improve overall system
  stability and performance.