The Voice AI Coworker: When Your AI Teammate Joins the Meeting

Claude Code settled an argument this year: an AI can be a coworker. Hand it a hard refactor, walk away, come back twenty minutes later to a tested branch and a summary of what it did and why. Millions of developers have quietly reorganized their day around that fact.

A voice AI coworker is the version of Claude Code that shows up on a video call instead of in a terminal — an AI teammate you delegate a meeting to, not just a task. Not a notetaker. Not a bot in the corner. A colleague that runs the 30-minute screening call, absorbs the customer discovery interview, or holds the recurring 1:1 you keep meaning to have but never quite schedule.

Below is why an AI coworker for meetings is the highest-leverage form the technology can take next, and what changes about a team when everyone on it has one.

But notice where that coworker lives. A terminal. A repo. A world of files, diffs, and tests, a world built to be operated by text.

The most consequential work in a company doesn't live there. It happens in conversations. The brainstorm where the strategy actually takes shape. The pre-flight before a sales call. The interview. The customer who finally tells you the truth in minute twenty-two. If the AI coworker stays in the terminal, it stays adjacent to that work forever.

We started Waterr on a simple conviction: the next AI coworker doesn't print output into a terminal. It joins the call.

The meeting is the interface for real work

People ask why we insist on video calls when a chat thread would be cheaper to build and cheaper to run.

Nobody huddles because they love meetings. They huddle because explaining is how humans transfer knowledge. We think out loud. We watch faces while ideas land. We pull up a screen and say "look at this, here." That's why the Zoom call, the huddle, the Teams standup have survived every productivity tool ever aimed at killing them: for knowledge transfer, brainstorming, and getting a team aligned before real work, the medium is the work.

So an AI employee that helps with strategy, brainstorms with the team, and works real-world problems has one non-negotiable requirement: it has to show up where that work happens. In the room. Seeing what everyone sees. Speaking when it's its turn.

That sounds obvious until you try to build it. Then you learn that voice intelligence is a different discipline from text intelligence, and most of what made the terminal coworker great does not transfer.

Voice is not text with a delay

Text-based intelligence is patient. Claude Code can think for four minutes, publish its plan, call tools, fail a test, retry, and show you the reasoning trace afterward. You tolerate the wait because the medium tolerates the wait. That's the contract.

A realtime conversation offers no such contract. The agent has a few hundred milliseconds to start speaking or the conversation breaks. Silence isn't a loading state in voice; silence is the other person checking whether the line dropped. Three things follow, and they change everything about how you build:

Voice is one-shot. Text lets you regenerate a bad reply. In voice, the wrong answer has already been heard. Every turn is production.

Voice is time-aligned. A transcript has no clock, but a conversation does. The pause before an answer, the half-interruption someone pulls back: that's signal, and a model reading only words misses it.

Voice has no reasoning trace. Text intelligence gets to show its thinking and be judged on the process. Voice gets judged on one spoken sentence, with all the thinking invisible behind it. Text intelligence hides its work and shows its result. Voice intelligence has to be the work, in the moment of the result.

Long reasoning traces to close a task, the thing that makes text agents powerful, are simply not available in-band. That constraint is the design brief.

What customers taught us about trust

We've spent the past year putting AI personas into real interviews, sales roleplays, customer discovery calls, and onboarding sessions with partners across very different domains. I expected each domain to want different things. Mostly, they didn't. Two findings survived every deployment, and I heard both first from customers, not from our own metrics.

The agent has to sound relevant and smart right now. One partner put it bluntly: the moment the agent forgot something it had said two minutes earlier, the whole session "felt off", their word, and the participant checked out. Users forgive an AI for many things. They do not forgive it for contradicting itself. Self-coherence is the trust threshold, and below it nothing else about the product matters.

The user has to feel understood, not just heard. Heard is transcription. Understood is adapting to intent: noticing the candidate is nervous and slowing down, noticing the buyer is stalling and asking the direct question, taking the follow-up as a follow-up instead of restarting the topic. Watching session recordings taught me you can see the exact moment a participant decides the agent understands them. Their answers get longer. That's the moment the conversation becomes real work instead of a demo, and honestly, chasing that moment is most of why I still watch the recordings.

Everything we build gets tested against those two bars, in that order.

Where the intelligence actually lives

Here's the strategic bet underneath all of this, and it runs against the industry's instinct.

The instinct is to wait for a bigger realtime model, one that can somehow reason deeply inside a 300-millisecond window. We think the ceiling isn't model capacity at all. The frontier voice models already know the constraint the user stated in minute four. What they lack, at the moment of the next turn, is having that constraint in front of them: it's buried in context, and surfacing it costs more time than the turn allows.

So we build the layer around the model instead. A platform that watches the conversation in parallel, deliberates off the critical path, and hands the voice model exactly what matters in the gap between turns that already exists: the fact the user stated, the commitment the agent made, the correction it must not ignore. The same platform grounds the agent in the rest of the meeting: the shared screen, the camera, the documents in the room. Realtime reasoning as a harness around the model, not a bigger model.

Text intelligence solved its version of this by publishing the thinking. Voice intelligence has to solve it by hiding the thinking, and surfacing only the one right sentence, at the one right moment. A great coworker on a call does exactly that: listens more than they speak, and passes the right note across the table at the right time.

I'm deliberately not laying out the whole architecture here. What I'll say is that this layer, not the voice pipe, is where nearly all of our research effort goes, and the gap it closes is measurable.

The coworker that shows up

The next twelve months will split voice AI into two camps. One camp treats voice as text with a delay: their agents will demo beautifully at turn 3 and fall apart by turn 30, and a human will quietly take back every meeting that matters. The other camp treats the conversation as the product: time-aligned, one-shot, memory-critical, multimodal, a participant rather than a narrator.

We're building for the second camp: a coworker you bring into the brainstorm, the strategy review, the interview, the discovery call, and that earns its seat the same way a human does: by remembering what was said, understanding what was meant, and doing the work in the room.

Claude Code showed what happens when an AI coworker meets the terminal. We're building what happens when one joins the call.