On December 4th, we launched Voice 2.0. We posted to LinkedIn, set up a few phone numbers people could call to try it out, and watched a small number of strangers attempt to jailbreak the agent in real time. Then we mostly stopped talking about it.

That was deliberate. Voice was the channel we were most worried about getting wrong. The history of voice AI in customer support is mostly the history of demos that fall apart on the second hard question. Sounding human turns out to be the easy part. The hard part is what happens when a customer calls about a disputed transaction at 11pm and the agent needs to query four backend systems while keeping them on the line.

For five months, we did not want to claim “production-grade.” So we did not.

This week we have the data.

What changed

GiveCard used Lorikeet voice during the 2025 SNAP government shutdown to handle 60,000+ emergency calls across English, Spanish, and Mandarin. The City of San Francisco deployment went live over a single weekend. Three hundred thousand cardholders were served in total.

Flex doubled CSAT versus their prior tool, cut average call duration in half, and handled a four-times volume surge in their busiest week. Berry Street is running 500 to 1,000 outbound appointment reminders a day. Carmoola is going live this week. Across the broader portfolio of voice deployments, head-to-head win rate against Sierra and Decagon is above sixty percent.

These are not pilot numbers.

Why it worked

The thing we kept misreading about voice AI is what kind of problem it actually is. We were thinking of it as a speech problem, which is what every voice-only competitor seems to be thinking too. It is a coordination problem.

A real production voice agent has to listen to a person describing a fraud event, pull their account history from a CRM, run a fraud check, coordinate with a payment processor, log a regulatory audit trail, and respond in under a second, while sounding patient. That is not text-to-speech tuning. That is workflow execution wrapped in audio.

The architectural decision that made this work for us was treating voice as the same problem as chat. The agent that answers your phone call is the same agent that answers your email and your chat and your SMS. The workflows are the same. The guardrails are the same. The knowledge base is the same. We added an audio layer on top.

This is the bet that Pockets of Determinism encodes. Natural-language agents call structured subworkflows as tools. The structured subworkflows enforce determinism inside their scope. The natural-language agent decides which subworkflow to call. The result is an agent that can hold a flexible conversation but can never invent a procedure.

Without that architecture, voice falls apart at exactly the moment it matters. With it, voice does what GiveCard, Flex, and Berry Street do now.

What we learned

Three things, in roughly the order they hurt us.

First, voice required heavier deployment effort than we had budgeted. Customers who were live in chat in three days needed three weeks for voice. The infrastructure is meaningfully more involved - telephony, codec choice, voicemail detection, latency budgets, kill switches. We have absorbed most of that into the platform now, but we cannot pretend a fintech going from zero to tens of thousands of calls per month is a turnkey deployment. It is a real implementation project, just one that finishes in weeks instead of quarters.

Second, voice exposes data gaps that chat hides. In chat, a customer will tolerate “let me look into that.” In voice, they hear the pause and lose trust. Every voice deployment we have run has surfaced API coverage gaps in the customer’s stack - data the agent needed in real time that the customer’s backend could not return in under five hundred milliseconds. The agent is only as fast as the slowest call it has to make.

Third, the resolution-priced model matters more in voice than in chat. Per-seat or per-minute pricing creates the wrong incentives for everyone. Lorikeet voice is $1.50 per resolved call on the Start tier. If the agent does not resolve the call, we do not get paid for it. That alignment makes a difference in how customers actually use the product.

What this means for the category

Voice AI has spent most of the last three years stuck in a demo-to-pilot loop. Companies pilot a voice agent, the agent works on simple FAQ traffic, and then the pilot stalls when leadership asks whether it can handle the actual hard calls.

The way past this is not to make the agent sound more human. It is to make the agent capable of doing the work. Five months in, the production case at GiveCard, Flex, and Berry Street is what the answer looks like when an AI voice agent is genuinely capable.

If voice has been the channel on your roadmap that you have been holding off on, we think the wait is over. Talk to us when you are ready.