https://www.computerweekly.com/news/366634053/How-Zoom-is-approaching-agentic-AI
Zoom is looking to turn its artificial intelligence (AI)-powered digital assistant into an AI agent that can grasp the context of meetings and autonomously complete tasks.
Speaking to Computer Weekly on a recent visit to Singapore, Zoom’s chief technology officer, Xuedong Huang, said the latest iteration of the company’s digital assistant, AI Companion 3.0, is built on its agentic AI architecture defined by four key attributes.
The first attribute is persistent, long-term memory that goes beyond a single conversation. Unlike chatbots that often forget the context after a few queries, an AI agent must remember past interactions, meetings and decisions to understand the bigger picture. Huang said this is required for AI agents to be useful. “We all have long-term memory, so AI must have good memory,” he said, adding that Zoom is well-positioned to provide this capability as it hosts business discussions where context is created.
The second attribute is the ability to perform deep reasoning so that AI agents can understand intent in conversations and complete the required tasks in what Huang deemed as “conversation to completion”.
“Conversation to completion requires good reasoning to identify the most important tasks you need to accomplish,” said Huang. This means the AI can listen to a discussion and infer that a particular decision requires a follow-up email, a new entry in a project plan, or a scheduled meeting, without being explicitly told to do so.
As no single tool can do everything, the third attribute of agentic AI is the ability to act as a conductor, orchestrating multiple tools and other specialised agents to accomplish a complex goal. This means it can integrate different functions to execute a multi-step task.
For example, if a user asks an AI agent to organise a follow-up to a quarterly planning meeting, the agent might need to access the meeting transcript for action items, use a word processor to draft a summary, integrate it with a calendar to schedule follow-up sessions, and use an email client to send out communications as part of one workflow.
Finally, agentic AI needs to be proactive and autonomous. Rather than wait for commands, it can anticipate user needs and take initiative. Huang compared this with the current state of popular AI models.
“ChatGPT, as we understand today, is passive,” he said. “It’s not executing tasks on your behalf. You had a Zoom meeting, but ChatGPT doesn’t know what was discussed.”
By contrast, an AI agent would not only know what was discussed, but would suggest and even begin executing the next steps.
Powering Zoom’s AI agents is the company’s federated approach to AI that it has implemented since the first version of AI Companion. Instead of using just one large language model (LLM), Zoom uses models from different companies, like OpenAI and Anthropic, along with its own small language models (SLMs) that are designed for specific jobs such as summarising or translating.
However, the benefit of model federation isn’t just about the models, but also how they interact through a process called inference scaling. For example, to summarise a meeting, AI Companion would use Zoom’s own SLM to perform the initial heavy lifting, creating a first-pass summary quickly and at a low computational cost.
It then passes this initial summary to a more powerful LLM such as GPT-4 to polish and refine it, improving the language, structure and nuance. This ensures that expensive, powerful models are only used when necessary, while cheaper, specialised models handle the majority of the processing.
Huang pointed out that this “committee of models” also enables oversight and control of AI outputs. “Because of the power of how we federate, we have a committee that can detect risks and safety issues better than one single member can,” he said, adding that the diversity of models inherently makes the system robust and safer.
For those who need agentic AI to be more suited to their needs, Custom AI Companion also offers the ability to create and deploy custom AI agents using a low-code tool, with access to a tooling library and pre-built templates for various workflows. Additionally, administrators can connect to pre-built, third-party agents and use the Agent2Agent (A2A) protocol to get work done across applications such as ServiceNow.
More recently, Zoom teamed up with Nvidia to further bolster its agentic AI capabilities. This includes adding Nemotron reasoning models to its federated model architecture, using Nvidia’s graphics chips and AI software stack to optimise AI Companion’s core capabilities, and building its new 49-billion-parameter LLM based on Nemotron.
03 Nov 2025