Your guide to getting started with ZoomMate
Everything you need to start using ZoomMate: notes, search, workflows, and AI tools — all inside Zoom Workplace.
Select a language
Zoom AI Services bring speech recognition, enterprise context, and decision intelligence together to help developers build AI agents that get work done.
Published on October 19, 2026
Xuedong Huang is the Chief Technology Officer (CTO). Prior to Zoom, he was at Microsoft where he served as Azure AI CTO and Technical Fellow. His career is illustrious in the AI space: he began Microsoft’s speech technology group in 1993, led Microsoft’s AI teams to achieve several of the industry’s first human parity milestones in speech recognition, machine translation, natural language understanding, and computer vision, is an IEEE and ACM Fellow and an elected member of the National Academy of Engineering and the American Academy of Arts and Sciences.
Xuedong received his Ph.D. in EE from the University of Edinburgh in 1989 (sponsored by the British ORS and Edinburgh University Scholarship), his MS in CS from Tsinghua University in 1984, and BS in CS from Hunan University in 1982.
A voice agent can understand every word a customer says — and still do the wrong thing.
It can route the request to the wrong team, apply an outdated policy, or promise an action it cannot complete. None of these is a speech recognition error. Accurate speech recognition is essential. But reliable action requires more.
That is the next challenge we are addressing with Zoom Decisions APIs in Zoom AI Services (ZAS), our enterprise AI APIs built on the same technology that processes millions of media hours daily across Zoom's products. Building on Scribe, enterprise search/Onyx, and voice-agent capabilities, the new Zoom Decisions APIs give developers another building block for turning conversations into completed work:
Hear with precision. Decide with confidence. Get work done.
We have spent years improving how accurately machines understand human speech. The next step is making that accuracy useful throughout a workflow.
Consider a customer calling about a missing package. Scribe captures the request. Enterprise search can supply the relevant policies and supporting knowledge. Zoom Decisions can evaluate the issue and its urgency, and whether escalation is needed. The application then applies permissions and business rules, invokes the appropriate tools, and verifies the outcome.
ZAS provides building blocks for this architecture:
ZAS Diagram: Conversation (Scribe/Voice Agent) leads to Context (Enterprise Search/Onyx), then Decision (Zoom Decisions, new), then Action (your agent).
Frontier LLMs can already produce structured answers. For many application workflows, developers also need to evaluate predefined alternatives and estimate how strongly the available evidence supports each one. A valid output format makes an answer easier to use; decision scoring helps the application determine what to do next.
Zoom Decisions evaluates text and images against outcomes the developer defines. It returns typed answers with probability estimates: yes/no for a condition, choice among categories, or score against an ordered scale. Multiple questions can share the same input in one request. Because developers define the outcomes, "take no action" and "escalate to a person" can be outcomes too. The probability estimates tell the application when to act and when to hand off: in production AI, knowing when a model is unsure matters as much as the answer itself.
Zoom Decisions is designed for accuracy and cost efficiency in high-volume classification, routing, triage, and tagging. Zoom Decisions Pro is built for more demanding judgments, such as checking contract clauses against defined criteria, assessing escalation needs, and flagging potential policy exceptions. Both models share one interface, so developers can choose the model that best fits each workload.
I see Zoom Decisions APIs and frontier LLMs from providers such as Anthropic and OpenAI as complementary building blocks. Frontier models are well suited to exploring alternatives, connecting evidence, and planning across multiple steps. Zoom Decisions APIs provide an efficient way to make repeated, bounded judgments within those workflows. The right choice depends on the work required to reach the desired answer. A short answer can still require substantial reasoning. Validate decision quality and probability estimates on representative inputs before setting thresholds for automated action or human review.
| Use a decision model when... | Use a frontier LLM when... |
|---|---|
| The evidence is available, and the task is to evaluate developer-defined alternatives. | The task requires gathering evidence, developing alternatives, or deciding how to approach the problem. |
| The workflow repeatedly applies a defined rubric to classify, route, prioritize, or score. | The task benefits from extended computation, multi-step synthesis, planning, or novel problem-solving. |
| Evaluation shows that a typed judgment meets the application's quality, latency, and cost requirements. | The application needs a developed explanation, a new solution, or a plan that adapts as new information arrives. |
In our internal evaluation, we compared Zoom Decisions and Zoom Decisions Pro with TypeSafe's Jev and the decisions APIs from OpenAI and Microsoft. We used five text benchmarks: Contested Fast Decisions (routing, classification, and triage across 17 domains), JevBench (structured decisions across varying difficulty levels), RAGTruth (grounding), JudgeBench (evaluating AI-generated responses), and ContractNLI (contract inference).
Zoom Decisions Pro achieved an average accuracy of 83.9% across five text benchmarks, while Zoom Decisions averaged 81.3% accuracy. Both models outperformed Microsoft-Decision-1, TypeSafe's Jev, and OpenAI's Decisions API.
Note: These are Zoom internal results for the tested models, datasets, and configurations. They measure decision accuracy, not end-to-end agent performance.
Many decisions begin with an image: a damaged parcel, a receipt, a form, or a screenshot. Both models accept images alongside text. Across four image-decision benchmarks covering photo-based rules, counting and spatial relationships, receipt and form tasks, and meme moderation, Zoom Decisions Pro averaged 86.0% accuracy in our internal evaluation, just behind TypeSafe's Jev (86.3%) and 5.8 points ahead of OpenAI's Decisions API (80.2%). Microsoft-Decision-1 does not accept images and so was not tested.
Decisions need to feel seamlessinstant, whether confirming a return policy mid-call or flagging a risky transaction in real time. In our October 15 test, Zoom Decisions posted a 117 ms median latency, lower than the OpenAI and Microsoft decision APIs we tested. Competitor APIs were measured through OpenRouter and Zoom Decisions in internal testing, so the network paths differ.
We are encouraged by these results. But the benchmark that matters most is the work your application actually performs. Start with one workflow and compare accuracy, latency, and cost on representative inputs, including difficult exceptions.
At launch, Zoom Decisions is priced at $0.042 per million input tokens, the same as TypeSafe's Jev and Microsoft-Decision-1, and Zoom Decisions Pro at $0.10 per million input tokens, the same as OpenAI's Decisions API. In our evaluation, Zoom Decisions outscored OpenAI's Decisions API on both text and images at less than half the price.
Decisions APIs share an OpenAI-compatible request and response format. Developers can evaluate hundreds of choices and ask multiple questions about the same input in one call.
We are encouraged by these early results. But the benchmark that matters most is the work your application actually performs. Zoom Decisions APIs are one building block in a larger architecture — Scribe hears the conversation, enterprise search brings the context, Decisions makes the call, and your agent completes the work.
The direction is clear. AI becomes useful when developers can build with it without wrestling complexity, and when the economics let it reach every workflow that needs it. That is what Zoom AI Services is built for: taking even more work from conversation to completion.
Learn more about Zoom Decisions API or sign-in to Zoom.ai to try it out.
