Better Every Call? Building the Voice AI Improvement Loop
Voice agents are moving from demos into real production - handling thousands or millions of conversations with real customers.
The next question is harder: once an agent is live, how does it actually get better?
Every conversation leaves signals behind. A caller repeats themselves, interrupts, rephrases, asks for a human, completes the task or gives up entirely. But understanding what went wrong, and turning that into the right improvement is far from straightforward.
Was it speech recognition? Noise? Latency? Turn-taking? Reasoning? The prompt? A tool failure? And once you make a change, how do you prove that you improved the system rather than simply moved the failure somewhere else?
Join Speechmatics and Roark for an evening on building the voice AI improvement loop: observing what happens in production, measuring quality, diagnosing failures, testing changes and safely making agents better over time.
We'll cover the questions teams are running into as voice AI reaches production scale:
What should a voice agent actually optimise for?
How do you attribute failures across the voice stack?
How much autonomy should an agent have to improve itself?
Can you learn from production calls you aren't allowed to retain?
What do benchmarks tell us, and what do they miss?
We'll also show the loop in action with live demos, followed by Q&A, drinks and conversation.
Agenda
6:30pm - Doors, food & drinks
7:00pm - Fireside: Building the Voice AI Improvement Loop
7:40pm - Live demos
8:00pm - Q&A
8:15pm - Drinks & conversation
*Hosted by Speechmatics × Roark