On July 8, Tavus announced that PALs can now join Google Meet. You invite one to a meeting the way you would invite a colleague, or drop it into a call already in progress, and it shows up as a participant: on camera, listening, taking its turn in the conversation.
This is a good move, and I want to talk about why it matters beyond one company’s feature launch. Because the interesting part is the premise underneath it.
The meeting is the natural place for an agent to show up
When AI agents first started entering the workplace, most of them arrived through phone calls. That made sense. Telephony is old, well understood, and every business already has a number. But phone calls are also a narrow channel. No screen, no faces, no shared documents, and increasingly, not where the actual work happens.
Slowly, people are realizing that Google Meet and Zoom are better places to introduce an agent. The reasoning is practical:
- The infrastructure is reliable. Meet and Zoom have spent a decade hardening real-time video at scale. You inherit all of that for free.
- Everyone already knows how to use it. There is no new app to install, no new habit to build. A meeting link is the most familiar object in modern work.
- No new permissions to negotiate. Your IT team already approved these tools. An agent that joins as a participant slots into existing access patterns instead of requesting new ones.
- Deployment is low friction. Sending a calendar invite is the entire rollout plan. Compare that to embedding a widget, provisioning phone numbers, or asking customers to visit a new site.
An agent joining your Google Meet is a natural extension of how work already flows. That is why this has been one of our main areas of focus for a long time, and it is why more and more players are now moving into this space. Tavus joining is a signal that the category is real.
What Tavus shipped
The feature itself is straightforward and well executed. A PAL, which is Tavus’s conversational persona built on their Phoenix-4 rendering, Raven-1 perception, and Sparrow-1 dialogue models, can now enter a Google Meet. It sees who is speaking, hears the room, responds in real time, and participates like anyone else on the call. Their announcement walks through use cases like meeting notes, sales call support, technical walkthroughs, and even hosting a standup.
Google Meet is the first platform they support, with Zoom and Microsoft Teams described as the next doors they are working to open. If you are a long-time Tavus user, this is genuinely great. Their rendering research is some of the deepest in the avatar space, and having that quality of face show up inside a Meet is a real step forward.
We have been living in this space for a while
Meeting agents are not new territory for us. We published a full walkthrough of building a voice AI agent that joins Google Meet and Zoom back in March, and a deeper guide on how to build an AI meeting agent in June. The Google Meet agent documentation covers the operational details: calendar integration, keyword filters so the right agent joins the right meeting, and auto-join for private calls.
That time in production has shaped what the platform supports today:
- Both Google Meet and Zoom, including private, authenticated meetings.
- Screen share and agentic tools. Agents can present slides, use whiteboards, drive browser automation, and call webhooks mid-conversation.
- Two agent types. Streaming agents show an avatar and actively participate. Notetaker agents join silently, observe, and score the conversation against a rubric you define.
- Recording, transcripts, and multimodal analysis. Every session produces a transcript and a structured report with dimension scores and evidence quotes, delivered to your dashboard and inbox.
- Calendar-driven deployment. Connect a calendar, set keyword filters, and the agent joins matching meetings on its own.
If your use case is learning and training, or running sales conversations, these interaction and analysis layers are where the value concentrates. An agent that joins a call is table stakes. An agent that shares its screen to demo your product, scores the conversation afterward, and writes the results somewhere useful is a workflow.
Use cases we see deployed today
A few patterns from actual usage, to make this concrete.
First-round screening interviews. The agent joins the Meet, conducts the interview from your question set, probes on follow-ups, and scores the candidate against your rubric. The hiring manager reviews a structured report instead of sitting through every screen.
Sales discovery and demos. A streaming agent runs the discovery call, presents slides, handles common objections, and books the follow-up with a human rep. The rep walks into the second call with full context instead of starting cold.
Sales call coaching. A silent notetaker rides along on live discovery calls. After the call, the rep gets scores on opening, discovery depth, talk-to-listen ratio, and closing, each backed by quotes from the transcript. No manager shadowing required.
Training and roleplay at scale. Teams rehearse difficult conversations, negotiations, and client calls against an agent that pushes back realistically, inside the same Meet interface they use every day.
Each of these works because the meeting is already where these conversations happen. The agent goes to the work instead of asking the work to come to it.
If you are thinking about building this yourself
Getting an agent into a meeting reliably is harder than it looks. You need a bot that can join calls, survive waiting rooms, and hold a stable connection with browser automation. You need process automation running behind the scenes to schedule joins, manage sessions, and recover from failures. You need calendar permissions, authenticated bot accounts, and a real-time voice pipeline with acceptable latency. Then you still have recording, transcription, and analysis to build.
Realistically, this is two to three months of developer work before you have something dependable, and that is before you have written a single line of what your agent should actually do. Using infrastructure that already solved these problems lets you spend that time on the part that differentiates you: the agent’s job, its context, and its evaluation criteria.
Avatar provider versus end-to-end platform
Here is the distinction worth understanding when you evaluate this space. Tavus is, at its core, an avatar provider with excellent research behind it. The face, the rendering, the perception models. That layer matters, and they do it well.
But the avatar is one layer of the stack. We wrote a full comparison of the avatar landscape in our guide to the best virtual avatar solutions in 2026, and a closer look at the latency leader in Anam vs ElevenLabs Avatars. The short version: avatar quality is improving fast across the board, and the layer is commoditizing the way text-to-speech did.
Tough Tongue AI sits above that layer, end to end. You can bring your own keys for the avatar provider you prefer. If you want Tavus’s replica quality, bring your Tavus account and deploy it through our platform. If you want Anam, which is very low latency and pleasant to use in live conversation, bring those keys instead. Pick the models you want, switch providers when a better one ships, and keep everything else constant. The meeting integrations, the screen share, the agentic tools, the recording, and the multimodal analysis all come out of the box, with no markup on top of what you pay your avatar provider.
This matters because the avatar you choose today is unlikely to be the avatar you want in a year. The agent around it, its scenarios, its accumulated context, its evaluation history, is what you actually invest in. Keeping that portable across providers is the practical win of the end-to-end approach.
See it working before your next meeting
The fastest way to understand what a meeting agent feels like in practice is to watch one get built and deployed. This walkthrough covers the whole loop in a few minutes: creating a scenario, deploying the bot into a Google Meet, and reviewing the analysis afterward. If you only have time for one thing from this post, make it this video. It will do more for your intuition than another thousand words from me.
Where this goes
More agents are coming to the meeting, from more companies, and that is healthy. The meeting is where decisions get made, deals get worked, candidates get evaluated, and teams get trained. Any serious agent platform will end up there, and Tavus arriving confirms the direction.
If you want to try this today, on both Google Meet and Zoom, with screen share, tools, recording, and analysis included, Tough Tongue AI is ready for you. Describe what the agent should do in plain language, send it a meeting link, and see what it comes back with. The first deployment usually takes minutes, and it tends to change how you think about what a meeting can be.
