Six months of Zowie, running live
The hard part of an AI agent was never getting the first one live. It's keeping it running at scale and changing a live system safely, every couple of weeks, without breaking what already works.
That's what we spent the first half of 2026 on. In this session, Jakub Darul (Engineering Manager) and Anna Jasińska (GTM Team) walk through what shipped, show it running on real interactions, and answer the question underneath each release: why it was hard to build, and why the reason it was hard is the reason it holds up in production.
What you'll see:
- Why Zowie voice is low-latency by design. How the Decision Engine keeps business logic off the language model, so we run smaller, faster models, and what that freed us to build: full cancellations, smart insertions, and voice configuration that isn't locked to one vendor.
- Hello with custom visual aids. The agent talks while the interface shows the detail, live. Why this wins on everything voice is worst at, and the conversion and order-value numbers behind it.
- How a change reaches a live agent safely. Versioned Flows, staging before production, Collaboration mode review before anything ships, and Supervisor scoring a change against real conversation patterns before a customer ever touches it.
The takeaway: shipping fast and running safely in a regulated enterprise aren't in tension. At Zowie, they come from the same architecture.
