Zoom AI Studio Tour: Building and Testing Virtual Agents (Full Transcript)

A walkthrough of Zoom AI Studio’s shared AI platform, agent configuration, knowledge, skills, diagnostics, escalations, and pre-production testing.
Download Transcript (DOCX)
Transcription editor
How to Build an AI Voice Agent with Zoom Virtual Agent Knowledge, Skills Testing
Added on 11/17/2026
Speakers
add Add new speaker

[00:00:02] Speaker 1: Here we are in the Zoom Virtual Agent Admin Portal inside AI Studio. You'll notice two agents already built here. Customer Service Main, which handles our primary inbound calls, and Customer Service Outbound, a dedicated agent for outbound campaign work. We'll stay on Main today. On the left menu, under AI Studio, you'll see Virtual Agent and, just underneath it, AI Expert Assist. Virtual Agent just answered Amy's call. AI Expert Assist supports a live agent, which we'll see once this call escalates. Both draw from the same foundation, one knowledge library, skill library, tool templates for connecting to external systems, and integrations to make it easy. Build something once in any of those areas, then decide where it runs, in your virtual agent, for your live agents, or both. That's a single platform built from the ground up, not two products bolted together. Now, back to the virtual agent. Let's drill into the Customer Service Main agent and see how it was configured. At the top, under Properties, this is where the agent's voice lives. We're using Hope today, one of several voice actors across genders, conversational styles, and accents. And if your brand needs its own voice, you can upload one directly here. Below that is the greeting, which can also trigger a lookup action. For example, performing a lookup on the caller's phone number before the virtual agent even speaks. Now, we chose not to do that here. We'd rather greet Amy, ask her how we can help, and then personalize the response once we know what Amy needs. Scrolling down, there are three foundational areas to every virtual agent. Agent guidance, knowledge, and skills. Agent guidance is where you set the ground rules, what the agent should do, what it should never do, references to system variables, and the conditions under which it needs to hand off to a live agent. Knowledge is next, and this is exactly what powered Amy's troubleshooting a moment ago. You'll see two knowledge bases here, and that's intentional. The first is general. Warranty policies, return policies, anything appropriate for any caller. Let's take a look. Here's a list of the articles. Now, it looks like this whether you build the knowledge base natively right here in Zoom or whether it's synced in from an external knowledge source. Either way, once it's in here, it's usable the same way. And going back to the lens I asked you to keep earlier, a knowledge base like this one isn't locked to just a virtual agent. The same articles can ground Zoom virtual agent's answer on a call like Amy's or show up as a suggested answer for a live agent inside AI Expert Assist. Speaking of live agent, later we're going to see what happens when the caller dials in and asks a question to the virtual agent where there's not an answer in the knowledge base, but when the call escalates to the live agent, they have the answer. We'll show you how that live agent's answer can be used to add a new article back to this knowledge base. The second knowledge base is our product knowledge base, and this is the one that mattered for Amy's call. Notice the segments column. Every article in here is filtered by the product SKU variable, which is a tag on each article. Because troubleshooting steps for a Pro 100 look nothing like the steps for a Pro 250 or 300, before the agent can pull a useful answer from this knowledge base, it has to know which SKU it's dealing with, which is exactly what happened on the call you heard with Amy. And that brings us to skills, where that SKU identification actually lives. You'll see five skills configured here. Whatever you want your virtual agent to own, whether it's order status, warranties, returns, whatever it is, you build a skill for it, and you can jumpstart any skill from the skill library, which ships with pre-built patterns for things like human escalation, one-time password verification, intelligent routing, email collection, and more. Now, the skill that I want to open and show you is product support and troubleshooting, because that's the skill Amy's call just ran. Every skill has three parts, the name shown at the top, the trigger description, which tells the system when to reach for this skill. It's a good idea to include examples. So for this skill, including phrases like, my machine isn't working, helps the virtual agent select this skill correctly. And the instruction section at the bottom, which contains the actual business logic, written using prompt engineering, you'll see tool references in blue. These are the tools that the virtual agent uses to interact with the external retail system. This is a custom JavaScript tool. In addition to this type of tool, we also have no-code tools, pre-built connector tools, the ability to connect to an MCP server. As long as your external system provides an API, Zoom CX can connect to it. In this example, there are three tools connecting out to external systems and a fourth tool, which is a built-in SMS tool that lets the virtual agent text the customer directly using the Zoom Contact Center phone number that the customer dialed. This is the logic you just heard execute end-to-end. Look up the caller by phone number, find the order, confirm the product with the customer, and once the SKU is confirmed, that's the segment filter that goes straight into the product knowledge base I just showed you to pull the right troubleshooting article for that exact model. Worth pausing on one thing here. What I've shown you is built to make the mechanism of how ZVA works easy to see. A clean handful of steps with one clear path. A production build carries more instruction depth than this, and that's not a gap to shore up later. It's just the nature of real conversations. Real callers don't stick to the one phrasing you had in mind. They combine two issues in one breath. They describe the same problem five different ways. They go somewhere you didn't plan for. So the real discipline in building a skill like this is writing instructions that hold up across the full range of what a caller might actually say, not just the handful of scenarios that are easiest to demo. This is precisely what agent performance, which I'll get to in a moment, is built for. Testing a configuration against a wide set of real-world phrasing before it ever reaches a live customer. Now, before we get to that kind of automated testing, there's a simpler step every person building a virtual agent reaches for first. Up in the top right is the preview mode, a manual calling tool. Anyone building this agent can place a live test call against the current configuration. One quick note before I dial in. Preview stores a test phone number as a variable so it can simulate a real caller ID lookup, the same identity check that ran on Amy's call without needing an actual phone call to trigger it. Let's start the preview.

[00:07:37] Speaker 2: Hi, thanks for calling Zoomily. I'm Hope, your AI virtual assistant. How can I help you today?

[00:07:42] Speaker 1: Hi, I ordered a coffee machine earlier this year, and I'm having some trouble with it.

[00:07:46] Speaker 2: Hi, Amy Martin. I was able to pull up your account using your phone number, and I see your PRO100 order.

[00:07:52] Speaker 1: Stopping here. Notice the agent recognized this test caller off the phone number and pulled up the PRO100 order on the account, the skill that we reviewed earlier running live in the builder's own browser. The wrench icon next to the transcript opens details, the builder's diagnostic view into this call. What you're looking at here is the chain of thought, the agent's reasoning laid out step-by-step so you can see exactly how it got from what the customer said to what it did. Just above that is agent tracing, a lower-level trace of every LLM call and every tool call behind that reasoning. Agent tracing is there any time you need to go a level deeper than chain of thought to confirm the agent is doing what you actually built it to do. Next, let's go into the settings cogwheel. Here we can see agent-level settings, such as whether the agent is voice-only, which is our current configuration, or have this agent also handle the chat channel. This can be useful if you want the same functionality on both channels, but you only want to configure it once. We also support a growing list of languages. There is also an agent behavior section, a handful of optional dials worth knowing about. Fill-a-words, for one, if a back-end system takes a second to respond, the agent doesn't just go silent. It'll say something like, let me check that for you, the same way a person naturally fills the gap while they look something up. There's an inactivity timeout and a max call duration, so a call that goes quiet or runs long doesn't just sit open indefinitely. And a custom dictionary, so brand terms and product names get recognized and pronounced the way your business actually says them, not, however, a generic model guesses. One more piece here, and this one matters for reporting, escalation and handover. This is where you can configure, at the agent level, whether a handoff counts as escalated by design, meaning this scenario was always meant to route to a live agent, versus an escalation that happens because the agent hits something it couldn't resolve. Last stop, and the one I mentioned earlier, agent performance. This is where you build test cases and run simulations against your agent before it ever touches production traffic. If you caught the earlier session with Arvind, he covered this, so I'll keep it brief. Here on the screen, there are two different numbers, and they're answering two different questions. Pass rate is test-based. Of all the test cases we ran, what percentage passed outright? Performance score is more holistic. It's measuring the agent's overall ability to actually get the user's task done, even accounting for the tests that didn't pass cleanly. Below the scores is a summary describing what the tests covered, and underneath a full breakdown of exactly which test cases passed and which failed. So you're never just looking at one number and wondering what's behind it. Getting an agent to 50 or 60 percent is quick. Getting past 90 is where the real work happens. Mapping business requirements into prompt guidance until it holds. That's the benefit of having this test suite built natively into the product. It understands your configuration and every version of it. So before you publish a change to production, you're running it against real test cases and can see exactly what changed and whether it still holds, not just hoping it does. It'll even suggest test cases on its own, catching gaps you might not have thought to write. http://www.kenhub.com

ai AI Insights
Arow Summary
The demo tours Zoom AI Studio’s Virtual Agent Admin Portal using the Customer Service Main agent. It explains that Zoom Virtual Agent and AI Expert Assist share the same knowledge libraries, skills, tools, and integrations, allowing content and capabilities to be built once and used for self-service, live-agent assistance, or both. The presenter reviews agent properties such as voice, greetings, caller lookup, guidance, segmented knowledge bases, and reusable skills. A product-support skill identifies a caller and product SKU, uses external-system tools to retrieve order data, filters product knowledge for the relevant model, and can send SMS messages. The presentation also covers manual preview testing, diagnostic chain-of-thought and agent tracing, channel and behavior settings, handoff reporting, and Agent Performance simulations for validating real-world test cases before production. An unrelated URL appears at the end of the transcript.
Arow Title
Zoom AI Studio Virtual Agent Admin Portal Demo
Arow Keywords
Zoom AI Studio Remove
Zoom Virtual Agent Remove
AI Expert Assist Remove
knowledge base Remove
skills Remove
agent guidance Remove
product SKU segmentation Remove
external integrations Remove
preview mode Remove
agent tracing Remove
call escalation Remove
agent performance Remove
test simulations Remove
Arow Key Takeaways
  • Virtual Agent and AI Expert Assist use a shared platform foundation of knowledge, skills, tools, and integrations.
  • Knowledge bases can be native or externally synced, and their articles can serve both virtual and live agents.
  • Segmenting product knowledge by SKU helps the agent retrieve model-specific troubleshooting guidance.
  • Skills combine trigger descriptions, business-logic instructions, and tools for external-system actions such as caller and order lookups.
  • Preview mode supports live manual testing with simulated caller identity, while diagnostics expose reasoning and lower-level tool/LLM traces.
  • Agent Performance provides pass rates, holistic performance scores, test-level results, and suggested test cases to validate changes before production.
  • Escalation settings distinguish intended live-agent transfers from unresolved virtual-agent escalations for reporting.
Arow Sentiments
Positive: The tone is confident, instructional, and product-focused, emphasizing unified capabilities, flexibility, diagnostic visibility, and rigorous pre-production testing.
Arow Enter your query
{{ secondsToHumanTime(time) }}
Back
Forward
{{ Math.round(speed * 100) / 100 }}x
{{ secondsToHumanTime(duration) }}
close
New speaker
Add speaker
close
Edit speaker
Save changes
close
Share Transcript