Zoom Built Its Business on Capturing Your Attention. Now It Is the Plumbing AI Agents Depend On.
Zoom built its business on capturing your attention during meetings. Now AI agents are treating it as essential infrastructure — and that is awkward.
On May 18, Zoom expanded access to its Model Context Protocol server, the technical layer that lets external AI systems query Zoom meeting data — transcripts, recordings, summaries, action items — without the data leaving Zoom's control. At launch, agentic search pulls from Zoom's own tools alongside ten external platforms including Salesforce, Workday, and ServiceNow. The Anthropic Claude integration arrived in April; the OpenAI Codex plugin landed in May, giving developers a way to pull meeting context directly into coding environments where the work is happening.
The practical implication for enterprises is concrete: if you deploy an AI agent that needs to know what your team decided in a meeting, Zoom is becoming a verified source those agents query in real time. The data does not move or duplicate. The agent comes to Zoom, and Zoom's existing access controls determine what it sees.
The alternative is moving meeting data to a data warehouse or vector store, which requires duplicating the data, maintaining access controls in two places, and accepting that the copy diverges from the source over time. Querying Zoom in place avoids all three. For developers building with Zoom MCP, the workflow — described in Zoom's own developer documentation — lets an AI agent search across multiple meetings using natural language to identify key decisions, pull relevant transcripts and summaries, then chain those outputs into downstream tasks: action items in a project tracker, documentation in a wiki, follow-up calendar events, or in the case of the Codex plugin, automation scripts and code comments generated directly from meeting decisions. That is the difference between integrating a set of APIs and managing a live data sync with all its failure modes. Real-world deployment testimonials from developers building with Zoom MCP were not available for this story.
The tension is in who is doing the querying. Google, Microsoft, Slack, and Zoom all publicly frame themselves as the AI hub for work and business intelligence, according to No Jitter. But frontier AI vendors — OpenAI, Anthropic, Google — also want that same hub position, which means they need access to business communications data to make their agents useful inside enterprises. Zoom is simultaneously a hub competitor and a data supplier to those competitors.
The second-order effect is concrete: if Zoom's access-control layer becomes the ratification point AI agents use to verify decisions made in meetings, then Salesforce, Workday, ServiceNow, and every other SaaS vendor in the Zoom ecosystem face the same choice. They can open their own MCP servers and become queryable peers, or accept being downstream targets that Zoom's MCP layer routes around. At launch, Zoom's agentic search already spans its own tools plus ten external platforms. The question is whether those platforms ratify the query or get bypassed by it.
The NSA issued guidance this week flagging that MCP deployments require formal security models — zero-trust network architecture, explicit input validation, monitoring — not just better prompts. Whether that overhead actually limits enterprise adoption depends on how you frame the alternative. In-place querying means Zoom's access controls govern what agents see, which for some enterprises is simpler and more auditable than managing a separate data warehouse where meeting data has been duplicated and access is governed separately. For compliance-bound enterprises that already trust Zoom's access controls for meetings, adding an MCP query layer may be less security overhead than building a new data pipeline. For enterprises that do not already operate inside Zoom's access model, the NSA's requirements may add real deployment friction. The honest answer is that it depends on the enterprise's existing architecture, and Zoom has not disclosed MCP adoption figures or named enterprise customers using the expanded capabilities.
The My Notes standalone app, at $10 per user per month launching later in May, looks like a metered revenue play — a way to monetize the humans who access the same context the agents are querying — rather than a stated platform strategy. Zoom has not disclosed revenue attribution for MCP-based features in public filings.
What to watch is whether that position holds. If the open MCP standard matures and competitors offer equivalent query layers, Zoom's specific implementation becomes commoditized infrastructure and the enterprise relationships become the only durable differentiator. If Zoom's access control implementation and trust relationships resist commoditization, the data governance role is worth more than the hub rhetoric suggests. The meeting is no longer the product. The record of the meeting is — and whoever controls that record is sitting on something the AI industry needs.