What shipped since our last update
Your bot keeps running when a model goes down. VoiceBot now runs on more than one model and switches to a backup automatically if the one, a bot is using becomes unavailable. Live calls continue and new calls keep connecting.
Every save creates a new version, and old ones stay put. Editing a bot no longer overwrites its configuration in place; each save is a separate, immutable version. Changing a live bot is now a reversible move: if an edit makes things worse, you go back to the version that worked instead of reconstructing it.
Setting up a new bot flow is one click. From the Bots page, configuring a flow drops you straight into a pre-filled flow editor. The account-level workaround is gone, so whoever builds the bot can finish it in one sitting without waiting on someone with admin access.
Unused bots, tools, and knowledge bases can be archived, not just left cluttering your account. Archived resources drop out of listings and stop being callable; the system also now blocks archiving something that’s still actively in use, instead of silently letting you break it.
Concurrency is visible while it’s happening. A new page shows your configured call limit, live usage, and how close the two are. You see you’re near the ceiling before a caller hits a failed call, so more capacity is something you ask for ahead of a campaign instead of during one.
Each bot can have its own nudge. What the bot says when a caller goes quiet is now set per bot rather than one fixed line across all of them. A collections bot and a support bot can each break a silence in the tone that fits, instead of sharing a generic prompt written for neither.
Calls can be capped at a maximum length. Set a limit up to 20 minutes and the call ends cleanly at it. A call that gets stuck now has a predictable ending and a predictable cost.
A new voice option for Indic languages. Maya Research is now available as a TTS provider. It widens the set of Indian languages where the bot sounds native to the person on the line, so language coverage stops being the thing that decides where you can deploy.
Exolite trial rollout is complete — VoiceBot is live by default on all Exolite accounts now, so a new user can start testing without a setup call.
Custom integrations are now routes on a managed integration engine, not code we deploy per customer. Tool calls and webhooks used to mean hand-writing and deploying a wrapper API for each customer, so every integration was a separate service to keep alive, and the ones nobody used stayed running anyway. These are now declarative routes on a shared Apache Camel engine: built visually, versioned and managed centrally, retired centrally when a customer stops using them. Invalid input is rejected at entry instead of partially saved. For you that means a new integration is configuration rather than a deployment, and it goes live without waiting on our engineering queue.
Bugs you might have hit — now fixed
- A TTS connection issue could mean up to 10 seconds of dead air. A slow handshake with the voice provider used to hang the call in silence; it now times out and retries in about 2 seconds instead, with metrics in place to catch it early if it recurs.
- A bot with a mismatched voice and language could produce a dead call with no explanation — the call would end after about 29 seconds with zero audio and a vague error. That configuration can no longer even be saved, and there’s a runtime check as a second layer of protection.
- Recorded call audio could come out distorted and sped up, from a wrong sample-rate setting in how audio was logged. Corrected.
- Call-outcome reporting could be wrong — some sessions were mislabeled as “completed by bot” when the caller had actually gone idle or the session had errored. Fixed.
- Bots and call records could appear more than once when scrolling through lists — the bots list and interaction-history table could both show duplicate rows. Fixed on the list views; if you still notice a bot or record duplicated in the underlying data itself, that deeper fix is still in progress.
- Pulling an interaction-history report over a long date range could fail outright. Fixed.
- Report timestamps showed only a time, not a date, making them hard to sort. Fixed.
What’s next for VoiceBot: a bot that tells you what went wrong before you ask
Right now, knowing how a bot actually performed on a given day means someone sitting down and sampling transcripts by hand. That doesn’t scale past a handful of bots, and it means problems get caught days after they started, not the morning after.
We’re building the fix: a call evaluator paired with a recommender agent, running on an agent harness so it can review every call a bot handles, not a sample of them. It’s already being tested against real production call traffic today, not just internally-generated test calls. The goal is simple – open the console in the morning and it already knows which calls had problems and what to do about them, instead of you going looking for the problems yourself.
