Release notes

August 2025

Real-time AI assistant switching

Make your AI more adaptive than ever. Based on detected intent and your routing rules, conversations are sent to the right assistant at the right moment – without changing channels. This supports specialised departmental assistants (e.g., billing, support, technical) and cross-organisation handoffs for unified experiences.

When a switch occurs:

The new assistant takes over with full context
The previous assistant steps back to keep the flow clean
On webchat, the assistant’s name and avatar update dynamically

Training workflow reliability and controls

A major upgrade to make training clearer and more dependable:

Resolved stuck training states – short snippets that could remain pending now complete or fail deterministically, with actionable error messages when something goes wrong (e.g., site scraping issues).
Download original training files – each uploaded file now has a Download button in its settings so you can retrieve the exact source.
View and edit text snippets – open a training text snippet to review and edit it; you’ll be prompted to retrain so changes take effect.
Clickthrough on training URLs – open referenced URLs directly from the training item for quick inspections.

AI timeout trigger

Ensure no messages go unanswered. In rare cases where an automated reply cannot be produced – for example, the LLM is temporarily unavailable or a response is otherwise unobtainable – the AI timeout event triggers as a final safety net. Choose how to handle it such event – escalate to a human, switch to another assistant, or open a ticket.

Balanced round-robin chat assignment

Workloads are balanced more fairly. Agents with fewer active chats are prioritised; when workloads match, assignment follows round-robin order based on the last allocation. This prevents overloads and remains consistent even when agents go Away or return.

Smarter agent availability controls

Agent status management now includes Offline. Agents can stay logged in while clearly marked as off-shift, helping managers distinguish short breaks from end-of-shift unavailability. This improves coverage planning and ensures chats are only routed to available agents.

Faster canned replies with hotkeys

Type “/” to open canned responses instantly – speeding up common replies and reducing agent effort.

Other improvements

Ticket filters now show email delivery status (pending, sending, sent, send failed, delivered, delivery failed) to surface and redeliver failures quickly.
Resend button for emails with pending status; automatic resend for pending and send failed.
Assigned chats open automatically in multi-chat layouts for quicker handling.
Automatic chat session closure keeps inboxes tidy by closing inactive threads after a set period.
Expanded thumbs up/down feedback lets customers rate both agent and AI assistant messages for richer insight.
Bulk ticket export supports up to 90 days per export, and analytics queries are refined for more accurate reporting.
Soft-delete users and AI assistants – preserves audit history while revoking access.
Added close icon to confirmation messages for swift manual dismissal.
Improved bulk ticket export reliability for large datasets.
Infrastructure upgrades improve performance and reliability in high-volume periods.
Security updates strengthen data protection and platform resilience.

Bug fixes

Cleaned up terminology and consistency across the platform.
Fixed occasional export failures and inbox errors.
Improved conversation sorting and create-ticket stability.
Resolved duplication and display issues in system messages.
Corrected occasional analytics assistant reporting errors.
Addressed errors when opening exports or updating inbox views.
Fixed inbox conversation sorting edge cases and create-ticket-action errors.
Resolved subscription error log issues.
Improved multi-chat layout handling with clearer plan limits.
Fixed real-time inbox message rendering across channels – removed delays, duplicates, and ordering issues.
Corrected UI to reflect user permissions accurately

550K