10 min read
Published on October 29, 2026
For twenty-five years, search has been built around one assumption: a person types a few words, scans a list of results, and decides what to click.
That assumption is about to break.
Earlier this year, on Zoom's 15th anniversary, our CEO, Eric Yuan, wrote that the future of daily work is about "turning every conversation into completed work," with meetings, chats, and emails all connecting to outcomes. As he put it, "the goal isn't just communication, it's completion" (15 years, 15 perspectives). Completion requires AI agents to act, and to do this, each agent must find the right information before it can act. They'll ask clarifying questions. As AI usage grows, more of the questions that reach search will come from agents acting on people's behalf, and search will need to be built for them.
That shift is happening in stages. Here's how enterprise search is moving from people finding information, to AI agents answering questions for them, to AI agents retrieving context on their own to act.
Find: You type, you read. You remember a phrase, type it in, and sift through results. The burden of locating the answer sits with you.
Answer: You ask, an AI agent reads. You ask a question in plain language, and an AI agent searches, processes, and writes a cited answer for you. This is where most of us are today. The human is still the one asking.
Act: The agent asks, and nobody reads. A planning agent needs the status of a deal, the decision from last Tuesday's meeting, and the latest information along with recent project specifications prior to generating a response before drafting a follow-up. It never shows a result list to anyone. It calls retrieval the way software calls an API, many times per task, often in parallel.
Most search products were designed for Find and are being stretched into Answer. Act is where the design assumptions change.
If every app builds its own search, an agent has to guess which one to call, and the answer quality depends on that guess. A better model is to retrieve globally and act locally: one retrieval layer that spans meetings, chat, docs, and connected systems with unified ranking, and app-level tools that focus on taking action.
For years, "enterprise search" meant copying everything into one index. That still matters, and hybrid approaches that combine semantic embeddings with keyword matching remain hard to beat. But indexing isn't always possible or permitted, and as agent harnesses improve, live access through protocols like MCP can match index quality for many tasks. The winning system won't be the one that indexes the most. It will be the one that, for each question, decides what to pull from an index, what to fetch live, and how to merge the results.
This is exactly how we designed the Ask API in Zoom context intelligence. The calling agent doesn't have to choose a source, write a query for each system, or know where an answer lives. It asks a question in natural language, and Ask decides where and how to retrieve the most effective information: Zoom's own meetings, chats, and docs through the index, and connected apps like Slack and Google Drive, plus other sources, through live access where that's the better or only option. It then reasons over what comes back, so that permissions are designed to be enforced and sources are cited. When an agent already knows what it's looking for, the Search API returns the matching records directly.
A person can sanity-check a result. An agent chaining ten retrieval steps can't. If retrieval doesn't enforce who is allowed to see what, and show where each claim came from, errors compound silently. Governance is moving to the primary product criterion. Zoom's Search and Ask tools use OAuth to authenticate a user's access to their Zoom data, so an agent's access to context is limited to what the user has authorized.
Click-through rate means little when nobody clicks. The question becomes: did the agent get what it needed to finish the job, and was it right? That pushes ranking, query rewriting, and routing toward reasoning about intent instead of matching terms.
Many enterprise knowledge systems were built around documents. But the real record of how work happens, including decisions, commitments, objections, and context, lives in conversations: meetings, calls, and chats. That's the material agents most need and have had the least access to.
Search that treats meetings and conversations as equally important sources alongside documents and business systems gives agents access to a closer approximation of institutional knowledge. This is the bet we're making at Zoom: that the retrieval layer for agents should be conversation-aware from the start, not retrofitted.
That's what Zoom context intelligence is designed to do. It is built to bring together what's said in meetings and chats, what's written in docs, and what lives in connected apps such as Slack and Google Drive, and returns answers that respect permissions and cite their sources. And because agents won't live in any single product, we expose it through two simple APIs available over MCP:
Connect them once, and any MCP-capable agent can use them, whether it runs in Claude, Copilot, Codex, Muse, or whatever platform your teams adopt next. Your agents get Zoom context wherever they work, and your people never have to move the conversation.
Here's what that requires:
The products that win the Act stage will be designed for agents: built to retrieve information across systems, respect access permissions, cite sources, and search across conversations as well as documents.
The enterprises that get this right won't just have better search. They'll have agents that can actually finish work, because those agents will have access to relevant records of how decisions were reached: the meeting where the direction changed, the chat thread where the details were settled, the document where the outcome was recorded.
The opportunity for every enterprise is clear: when your agents have the right context, they deliver more relevant, grounded responses — and better outcomes.
Zoom Decisions APIs turn conversations into completed work, evaluating text and images against defined outcomes for fast, reliable AI agent decisions.
6 min read
Every AI conversation begins with a machine trying to understand a person. I have spent most of my career on that problem, and I still find it the most honest test of whether a system is ready for real work.
12 min read
Most business meetings lose momentum the moment they end. Learn how to prepare, facilitate, and document every business meeting so your work moves forward.
