An AI agent joins the meeting

A design meeting often involves searching the existing code, sketching a diagram or updating the document on screen. That work helps the team make decisions, but it also takes someone's attention away from the conversation.

At QvAI, we built a voice AI agent to help a client's engineering team handle this work during their Google Meet calls. Connected to their company knowledge and tools, the agent can answer questions, draft diagrams and make authorised document updates while the team continues the discussion.

A data model the team could see and refine

One meeting focused on a complex data model: how different pieces of information should connect, and how the design would fit the software the team had already built.

As the engineers discussed the model, the agent drafted an entity–relationship diagram, or ERD, directly in their Notion technical review document. The diagram showed how the data was organised and connected, giving everyone a shared view of the proposal.

The diagram was embedded in the document using Mermaid, a tool that turns text descriptions into diagrams. The agent could edit both the Mermaid description and the surrounding explanation on the Notion page, so the visual model and the design notes could develop together.

The agent also used the existing code to inform its advice. It suggested keeping a history of status changes in a separate database table, following a pattern already used elsewhere in the system. The team had a concrete proposal to consider, based on how their own software already worked.

As the diagram took shape, the engineers could examine the relationships and spot anything missing. They could refine the draft while they were still discussing the relevant part of the design.

More attention for the design

We noticed less searching during calls and more attention on the discussion.

The visualisation work was especially valuable. Using the agent made us recognise how much of a meeting goes into making an idea understandable. Someone has to draw the relationships or write down a spoken proposal so that others can examine it. The team could now ask the agent to do that work.

Turning an idea into a diagram is also part of testing whether it makes sense. A relationship that sounds clear in conversation can raise new questions once everyone sees it. A missing connection becomes easier to point to; a proposed change becomes easier to discuss.

In this case, the agent produced the diagram more quickly. Its findings and draft gave the engineers something concrete to question and refine during the meeting.

Ask for help without leaving the conversation

Participants can address the agent by voice or chat to ask a question or assign a task. It follows the ongoing conversation, giving it context for the requests people make.

Connected to the client's knowledge base, codebase, Notion and GitLab, it can:

• Find relevant information about the existing system.
• Draft and revise diagrams and written explanations.
• Make authorised updates to documents and GitLab tickets.

The agent speaks when called on. If it identifies something relevant to add on its own, it can use Google Meet's raise-hand feature. Participants decide when to give it the floor. It can share information in chat too.

Participants can authorise changes by voice. As the team agrees what needs to change, someone can ask the agent to update the document or GitLab ticket in the same conversation.

How work continues alongside the conversation

The agent in the meeting uses a real-time voice model to process speech and produce spoken responses. When a request needs deeper investigation or work in another system, it delegates to more capable task agents running in the background. These have the tools needed to research, analyse or edit.

Those agents can search company information or work in Notion and GitLab while the voice agent continues participating. The team can use their results during the discussion.

At QvAI, we built the Google Meet connection, access to company systems and task delegation around this voice interaction. Together, these let a request made in the meeting lead to a code search or an edit on the Notion page.

For voice interaction, this implementation works with the OpenAI Realtime API or the Gemini Live API. Both support live conversation and tool use. ElevenLabs Realtime is another possible option to evaluate for future versions.

Where this approach could help your team

Our example comes from an engineering team. The same approach could support other meetings where progress depends on retrieving information or updating shared work:

• Product teams could ask the agent to consult existing product documentation and refine a specification as requirements become clearer. Participants could review the wording together while the reasons for each requirement are still being discussed.
• Operations teams could have the agent draft a process diagram during the discussion, making handovers, dependencies and exceptions easier to examine.
• Project teams could ask the agent to retrieve task information and make authorised updates to tickets and supporting documents as decisions are made.

These are potential applications. Their usefulness depends on the information available to the agent and the tools it can access. This implementation joins Google Meet; supporting other meeting platforms would require suitable integrations.

Start with the work your meetings depend on

Think about your last design review or planning session. What did someone have to look up, draw or update to help the discussion move forward?