` tags, etc.) is supported. Advanced HTML, like script tags, are removed upon saving your content.
* Tables are not supported at this time and are removed upon save actions.
**Manual Knowledge Bases**
* Manual KBs are limited to 5,000 articles per Knowledge Base.
**API-Connected Knowledge Bases**
There are no unique restrictions for this type of Knowledge Base.
**Web Sync Connected Knowledge Bases**
* Web KBs are limited to 5,000 articles per Knowledge Base.
* Web Sync can ingest content using sitemap sync, Link discovery, or URL upload.
* Web Sync will only detect up to 250 sitemap URLs.
* There is a 60-second timeout when fetching sitemaps and individual pages. The AI-based crawler option has a 30-second timeout.
* The sitemap crawler employs the user-agent header “Zoom KB Crawler.” This can be used to add a crawler being blocked by your system to an approved list.
* Each article has a 250 MB ingestion size limit and a 10 MB text size limit (excluding image data).
**General Syncing**
The sync task includes built-in timeouts for reliability and performance. The longest individual task is **index creation**, which has a **30-minute timeout**. If index creation exceeds this limit, the task will fail and must be restarted.
Administrators can also configure periodic sync schedules (e.g., daily or weekly) to automatically keep the content up to date with the source system. In addition, built-in validation tools allow admins to run queries against the knowledge base before deploying it to production, helping verify the accuracy and performance of the KB component independently.
####
Skills and Tools help virtual agents deliver empathetic, contextual conversations while enabling the automated resolution of complex issues
**Introduction to the Tool Library**
The Tool Library in Zoom Virtual Agent allows the system to connect with external platforms through APIs. In large language model (LLM) terminology, which Zoom also refers to as third-party models, a Tool represents a defined method for retrieving data from or sending data to third-party systems. This enables the Virtual Agent to go beyond static knowledge bases by dynamically interacting with external services, providing more accurate and actionable responses during conversations.
**Using Tools to manage external data**
Tools are designed to provide precise control when retrieving or updating information. Instead of relying on the third-party model to parse complex or lengthy API responses, Tools allow administrators to define exactly which data fields should be returned. This approach reduces ambiguity, improves reliability, and enables the Virtual Agent to deliver consistent answers without being overloaded by unnecessary or irrelevant API payloads.
**Handling sensitive and excessive data**
When integrating with external systems, APIs may return sensitive information such as PII or large amounts of unnecessary data. The Tool Library supports the use of JavaScript to manipulate responses before passing them to the third-party model. Administrators can filter, transform, or delete fields to keep only the essential information. This helps limit the exposure of sensitive data, avoids overwhelming the model, and streamlines the interaction between the Virtual Agent and third-party services.
**Best practices for Tool configuration**
Administrators should design with simplicity and precision in mind. Use cases that involve retrieving specific values, such as a case ID or order status, benefit most from Tool-driven integrations. By narrowing the response scope, Tools can perform targeted actions while minimizing the complexity pushed into the third-party model context. Each Tool should be tailored to its use case and tested against the APIs it connects to for consistent behavior across different scenarios.
**Test cases prefill and validate tool inputs before publishing**
Administrators can validate what an API call or custom script Tool returns before wiring it into a live flow. In the Tool's preview panel, a Test cases tab lets administrators create and save named test cases — each name up to 30 characters — that store input-parameter values across all supported data types. When running the debug, administrators can select a saved test case to prefill the input parameters (the last-used test case carries over automatically) or enter values manually, then run the debug to confirm the Tool behaves as expected. This removes the need to retype sample inputs on every debug run and lets administrators check variable output against sample inputs before publishing the agent.
**The built-in knowledge base tool is visible and directly invokable in the builder**
Administrators can view and explicitly invoke the built-in knowledge base tool directly within the Zoom Virtual Agent builder. Previously, this tool operated as an internal system component with no visibility or direct control in the builder interface. It now appears alongside other built-in tools, labeled as a system-provided tool, with a read-only name and an editable description. Administrators can invoke the tool multiple times across agent guidance or skills, making knowledge base lookup a deliberate, defined step in a workflow.
**Understanding Skills in the Virtual Agent**
Skills are the foundation of how the Virtual Agent executes tasks. Each Skill is defined by two core elements: a **trigger description** that specifies when the Skill should activate and a set of **instructions** that guide the Virtual Agent on how to complete the task. Instructions may include natural language guidance as well as references to Tools, which allow the Skill to interact with external APIs and systems.
**How the agent decides when to use Skills and Tools**
Once the trigger description matches a customer request, the Virtual Agent evaluates all instructions within the Skill. The agent determines independently when to call a Tool based on the Tool’s name and description rather than explicit step-by-step directions. This means administrators must provide clear Skill instructions and Tool descriptions so that the agent can interpret them correctly. Importantly, the agent has access to all Skills and Tools, and can decide to use Tools associated with different Skills if they are relevant to the task.
**Best practices for designing Skills**
For reliable execution, administrators should write **clear, concise trigger descriptions** that accurately reflect when a Skill should activate. Instructions can be organized into multiple boxes for better Skill management; however, they are combined during execution. Administrators should avoid embedding explicit commands such as “call this Tool now” in instructions. Instead, instructions should describe the intended outcome, while Tool names and descriptions should signal their proper use. Limiting the number of Skills to around five (5) is recommended, as excessive configuration can lead to context window limitations.
**Saving and managing variables with Skills**
Skills themselves can't directly save data to global variables. Instead, administrators must combine instructions with the use of built-in or custom Tools. For example, a Skill may include instructions for collecting a customer’s account number, but the actual saving of that data to a variable must occur through a Tool such as **collect\_save\_response** or a custom integration. This separation enables consistent and secure handling of sensitive data within the Virtual Agent framework.
**Example scenarios of Skill execution**
The interplay between Skills and Tools mirrors how a trained human agent applies their knowledge. For example, when a customer reports a lost credit card, the Virtual Agent initiates the “Lost their credit card” Skill. The instructions require member verification, prompting the agent to use the “Verify the member” Skill before continuing. After verification, the agent returns to deactivate the card, offers to check recent transactions, and, if appropriate, transfers the caller to a live agent. Similarly, in a “Check recent transactions” scenario, the Virtual Agent first confirms member identity before retrieving transaction details. These scenarios illustrate how Skills provide structure, while the agent automatically selects Tools and sequences tasks to complete the work.
**Instructions can include clickable hyperlinks with custom label text**
Administrators can send clickable hyperlinks with customizable display text through the Instructions feature in Zoom Virtual Agent chat. Previously, URLs appeared as plain text with no way to define a label, which made it harder for end users to know where a link led. Now administrators can assign descriptive label text to any hyperlink, so users see clear context about the destination before they click. Clickable hyperlinks are supported across web chat, WhatsApp, desktop apps, mobile apps, and additional channels.
**An email tool can send email from within a Skill**
Account owners and administrators can add an email tool directly in the agent skill builder by typing a forward slash (/). The tool exposes configurable attributes — tool name, description, subject, body, sender address, receiver address, and a variable for storing the email. Both the subject and body accept static text, concatenated variables, and URLs, and all variables resolve to their actual values before the email is sent. The email tool is available across chat and voice deployments, including Zoom Contact Center and Zoom Phone.
**The email tool can run in any flow, including post-engagement steps**
Beyond the skill builder, the email tool is available as a step in any flow — conversational, curated, or transfer-to-live-agent — and in post-engagement actions that run after a session closes. This lets an agent send a summary or follow-up email at any point in a conversation, not only at hand-off. For example, an after-hours flow can email a case summary when no live agent is available. The recipient can be a collected session variable, such as the user's email from chat, or a static address, and the subject and body support both static text and variables, with multi-language support.
**Form and support channel setup are built into the agent workflow**
Form creation for voice agents is now part of the agent build workflow, so administrators no longer have to navigate away from the current tab to create one. Support channel creation for chat agents is embedded in the same build experience. The result is a consistent setup process across both agent types and fewer steps to configure and deploy a new use case.
**Custom Action runs LLM-driven logic and saves the result to a variable**
Custom Action is a built-in tool available when building Skills for chat agents. Administrators provide custom instructions that define the purpose and behavior of the tool's LLM execution, and the output saves to a variable — String, Numeric, JSON, or Array. This supports post-engagement tasks such as generating summaries, creating tags, or preparing notes to pass to CRM systems, ticketing platforms, or human agents.
{% hint style="info" %}
This feature requires activation by Zoom. Contact your Zoom account team to enable this capability.
{% endhint %}
**A Skill library provides prebuilt, reusable Skills**
Administrators can browse a Skill library — a searchable catalog of prebuilt, reusable Skills that speeds up agent creation. Each library Skill combines one or more Tools with embedded instructions and guardrails tailored to common use cases such as appointment scheduling, case or lead creation, account lookups, and password resets. Administrators can add a Skill directly to an agent flow, edit it to create a new copy without changing the original, and save newly created Skills back to the library for reuse across agents.
{% hint style="info" %}
This feature requires activation by Zoom. Contact your Zoom account team to enable this capability.
{% endhint %}
####
Integrations connect to third-party services like CRM tools, e-commerce providers, or customer data platforms
Zoom Virtual Agent integrations connect voice and chat functionality to third-party services for additional chatbot functionality. Integrations let ZVA both push and pull data from third-party sources. For example, querying a customer’s data platform or e-commerce provider to retrieve the information needed to automate self-service while writing data back to those same systems. Integrations also connect ZVA to third-party CCaaS (Contact Center as a Service) vendors to help escalate conversations to a human agent when needed.
**Microsoft Teams integration for Zoom Virtual Agent campaigns**
The Microsoft Teams integration enables organizations to embed Zoom Virtual Agent campaigns directly inside Microsoft Teams as a native app. This brings agent chat capabilities into the Teams workspace, allowing internal helpdesk, HR, or IT support bots to serve employees directly within Teams channels or chats without switching between platforms. The integration supports both Teams desktop and web versions.
**WhatsApp and Facebook Messenger integration for Zoom Virtual Agent**
Zoom Virtual Agent chat agents can now be deployed on WhatsApp and Facebook Messenger through Zoom Contact Center, extending AI-powered support to more of customers' preferred messaging platforms.
Administrators configure this integration through the ZCC Flow Builder using the Bot Widget. Supported capabilities include:
* Text messaging
* Sending small media files
* Launching surveys
* Transferring to human agents
Currently unsupported features include:
* Large images and videos
* Curated carousels
{% hint style="info" %}
This feature requires activation by Zoom. Contact your Zoom account team to enable this capability.
{% endhint %}
**Chat agents can connect to SMS and Zoom Chat messaging channels**
Zoom Virtual Agent and Zoom Contact Center administrators can connect chat agents to supported messaging channels, including SMS and Zoom Chat, through ZCC's existing messaging channel framework. This extends chat agent capabilities beyond the current web chat and social messaging integrations, allowing organizations to deliver consistent AI-powered support across additional customer-preferred channels.
**Okta integration enables self-service identity verification in voice and chat interactions**
The Okta connector integrates Okta's identity platform with Zoom Virtual Agent so both voice callers and chat users can recover accounts without agent intervention. The agent identifies the user and verifies identity through Okta Verify push notifications. After successful verification, it can trigger automated actions such as password resets, account unlocks, MFA factor resets, and Temporary Access Code (TAC) generation. Okta stays invisible to the end user throughout.
{% hint style="info" %}
This feature requires activation by Zoom. Contact your Zoom account team to enable this capability.
{% endhint %}
**FedEx connector enables shipment tracking self-service**
The FedEx connector brings logistics self-service into Zoom Virtual Agent through the FedEx Basic Integrated Visibility API. Customers can track any FedEx shipment — including Express, Ground, and SmartPost — using a tracking number, and the agent returns real-time status, current location, estimated delivery time, and full scan history. Customers can also subscribe an email address to receive FedEx notifications for key events such as deliveries, exceptions, and updated delivery estimates. Those notifications are sent directly by FedEx to the recipient, so no webhook infrastructure is required on your end.
**Custom script tools can authenticate through a third-party connector**
Administrators can select and import an existing third-party connector's authentication directly within the custom script tool to make authenticated API calls to any external endpoint. There's no need to store or manage credentials manually — security and credential handling stay governed by the third-party connector framework. This lets administrators extend integrations beyond the prebuilt connector options for more flexible, custom API interactions.
### Virtual Agent Behavior
The following section details next-generation Zoom Virtual Agent behavior during a conversation with an end user.
####
Agent behavior in conversations pieces together structured prompts, Skills, Tools, and previous conversations
The Virtual Agent is powered by third-party models that predict the next word in a sequence. To offer accurate, useful responses, the system combines a structured prompt with guidance, Skills, available Tools, and prior conversation history. The Virtual Agent uses this context to decide whether to respond directly, reference the knowledge base, or activate a Skill.
**How the agent chooses between Skills and an available knowledge base**
When a user asks a question, the Virtual Agent evaluates whether the request aligns with a defined Skill or if it can be answered directly from the knowledge base. Skills are used when a query requires task execution or workflow steps, while the knowledge base is used when factual information is sufficient. The agent determines this choice automatically based on its instructions and available context.
Agent Guidance also influences the Virtual Agent, especially in cases where certain topics are off-limits. The Virtual Agent won’t perform agentic tasks such as looking for relevant Knowledge Base articles for any topics it’s been told, via guidance, to avoid. For more on Agent Guidance, see the previous section [Agent Guidance is an agent’s foundational identity, defining its tone, purpose and capabilities](#agent-guidance-is-an-agents-foundational-identity-defining-its-tone-purpose-and-capabilities).
**How Virtual Agent handles escalations and handoffs**
Zoom Virtual Agent includes an escalation threshold control that determines how many times a user must request a human agent before the system transfers the conversation. This setting allows the virtual agent additional opportunities to resolve customer inquiries before routing to live support.
Previously, any user message requesting a live agent would immediately trigger an escalation. With this update, administrators can configure a retry count in the chat agent settings that requires users to request human help multiple times before the transfer occurs. During these interim requests, the virtual agent prompts the user to restate their question, giving the AI additional context to attempt resolution. The system tracks each escalation request, and once the count reaches the configured threshold, the chat routes to a human support channel.
The default value is **1**, which preserves the previous behavior where the first request transfers immediately. Setting a higher threshold allows the virtual agent to make additional resolution attempts before handing off the conversation. This approach can improve containment and deflection metrics by reducing unnecessary handovers and increasing automation rates. The escalation threshold aligns with the existing threshold feature available for voice agents, providing consistent behavior across channels.
**Escalation-by-design provides a standardized framework for handoff analytics**
Zoom Virtual Agent analytics includes a standardized escalation-by-design framework across both chat and voice. This framework lets administrators define business conditions that classify certain escalations to a human agent as "by design,” resulting in intentional transfers.
To configure this feature, administrators must enable "Additional conditions for 'escalated by design'" in AI Studio. Once enabled, the system auto-detects design escalations upon publication. Administrators can include language in their agent guidance and skill instruction that the “escalation is by design” which will then be used by this feature to determine the conditions for when an escalation is by design.
**How the system prompt shapes behavior**
Each interaction is guided by a system prompt, which includes the current date, user identity, agent instructions, language preferences, available APIs, tool descriptions, conversation history, and the user’s latest request. Administrators directly influence agent behavior through the **agent instructions** and the **tools they configure**, while the system manages all other fields. This structured context provides the Virtual Agent with the right information to produce consistent, relevant answers.
**Flow of a typical conversation**
A typical interaction follows this predictable pattern:
{% stepper %}
{% step %}
The user submits a request.
{% endstep %}
{% step %}
The agent reviews its guidance, Skills, Tools, and Knowledge Base content.
{% endstep %}
{% step %}
The agent “thinks” through available options, identifying what information is missing and whether a Skill, Tool, or API call is required.
{% endstep %}
{% step %}
The agent either makes the call (if needed) or provides a direct response.
{% endstep %}
{% step %}
If the knowledge base returns multi-step instructions, the agent delivers only the first step, confirms user progress, and then proceeds sequentially.
{% endstep %}
{% endstepper %}
**Dial-by-name routing for Zoom Phone**
Zoom Virtual Agent now supports dial-by-name (also called transfer-by-name) routing for Zoom Phone. When deployed in a voice workflow, the agent can collect a spoken or typed name, interpret the request, and route the call to the correct user or extension. This enhancement enables more natural voice-driven navigation and reduces manual keypad entry. Administrators can configure dial-by-name routing rules in the Zoom Phone settings to determine which users or departments are eligible for automated routing.
**Standalone Dial-by-Name directory**
For Receptionist standalone voice deployments, administrators can create and manage a local contact directory within the Transfer-by-Name tool to support name-based call transfer. Each entry holds a contact name and an external phone number. When a caller asks to be connected to a person or department, the voice agent matches the request against the directory and transfers to the associated number. Entries can be added individually or imported in bulk from a CSV, and the imported list can either append to or replace existing entries with similar names, and can be exported. Changes take effect without rebuilding the full voice agent.
The standalone directory is built for ZVA standalone deployments where the customer doesn't have a Zoom Phone directory to look up. Zoom Phone customers can also use the Transfer-by-Name tool, but because only one copy of the tool is supported per agent, they choose whether the tool draws on their own (customer) directory or the Zoom Phone directory. The two are alternatives selected per agent, not layered together.
**Extended maximum engagement duration for inactive conversations**
Administrators can set the **Auto-close inactive engagements within flows** timer to a maximum of 24 hours, up from the previous 12-hour limit. The admin interface displays engagement duration in a clear hours-and-minutes format, which aligns Zoom Contact Center engagement duration limits with Zoom Virtual Agent and helps reduce configuration errors by removing the need for manual minute calculations.
**Call summary and variable integration with Zoom Phone**
Zoom Virtual Agent can transfer conversation summaries and custom metadata to Zoom Phone agents during call handoffs through variable pass-through. Administrators configure, using the ZVA admin portal, which variables display in the Zoom Phone interface during transferred calls, helping improve data relevance by giving agents real-time access to pre-collected information. This can reduce average handle times by providing Zoom Phone users with ZVA-generated call summaries and key variables before answering a call, during the conversation, and after the call ends.
This feature requires a Zoom Phone Customer Engagement Pack license (formerly known as the Zoom Phone Power Pack).
{% hint style="info" %}
This feature requires activation by Zoom. Contact your Zoom account team to enable this capability.
{% endhint %}
**Best practices for optimizing agent behavior**
Administrators should design Skills so that each set of instructions is clear and independent, reducing overlap or conflicts. Skill ordering can affect how the agent prioritizes actions, so descriptions should be written to minimize ambiguity. When combined with well-structured knowledge base content, these practices help the agent smoothly determine when to respond directly, when to reference documentation, and when to execute a workflow.
For more information, see our Technical Library articles [Customize User Chat Support Experience](https://library.zoom.com/business-services/zoom-contact-center/expert-insights/customize-user-chat-support-experience) and [Integrate Zoom Contact Center Features Into Your Website](https://library.zoom.com/business-services/zoom-contact-center/expert-insights/integrate-zoom-contact-center-features-into-your-website).
**Virtual Agent refines its reasoning through structured self-improvement**
Zoom Virtual Agent uses a structured self-improvement framework that refines how the agent reasons through tasks and follows enterprise policies. This is automatic and doesn’t require admin configuration.
The system converts enterprise policies and procedures into structured, step-by-step decision protocols. This helps Virtual Agent follow verifiable logic paths rather than relying on implicit inference.
Self-improvement actions include:
* Generating test scenarios
* Evaluating agent responses against expected outcomes
* Identifying reasoning gaps
* Updating decision protocols
Administrators don’t interact with the self-improvement system directly. The agent becomes more consistent, policy-compliant, and accurate over time without manual retraining. Improvement is observable through Virtual Agent performance and analytics and Agent Tracing. See the [Virtual Agent Performance and Analytics](#virtual-agent-performance-and-analytics) and [Agent Tracing provides step-by-step visibility into virtual agent decision-making](#agent-tracing-provides-step-by-step-visibility-into-virtual-agent-decision-making) sections for more information.
**Virtual Agent can start automated outbound calls to customer lists**
Administrators can configure Zoom Virtual Agent to place outbound calls to customer lists based on defined business rules and scheduling parameters. The virtual agent can start calls, deliver scripted messages, answer customer questions using its knowledge bases and skills, and escalate to live agents when needed. Built-in call management controls help align outreach with customer preferences.
Administrators create outbound campaigns within AI Studio by selecting a contact list, assigning a virtual voice agent, defining a caller ID, scheduling dialing windows, and configuring retry behavior.
Outbound calling leverages Zoom Contact Center's outbound campaign infrastructure, and engagement results and call status are recorded in ZVA analytics. Outbound calling is currently available for voice-only channels.
**Preview dialer mode**
ZVA outbound now uses preview dialer mode in place of the previous agentless dialer mode. Previously, the outbound call was placed first and the bot voice-flow session was only spun up after pickup, so the callee could hear a stretch of dead air before the bot spoke. Now the bot session is initialized and held in a silent, ready state before the call is placed. The moment the call bridges, the bot is already listening and responds immediately — removing the pause callees used to hear on pickup and making outbound calls feel like a real conversation from the first hello.
**Zoom Phone customers can use SMS capabilities within Zoom Virtual Agent**
Zoom Virtual Agent and Zoom Phone customers can use Zoom Phone SMS capabilities directly within ZVA to enhance voice interactions. During a call, the virtual agent can send customizable SMS messages containing dynamic text, URLs, or survey links. The system supports variables and splits longer messages for efficient delivery. This allows customers to receive detailed information that's difficult to convey by voice alone.
To use this feature, the following prerequisites must be met:
* Assign a valid number to the Auto Receptionist (AR)
* Enable the SMS policy
* Add at least one AR member to the SMS/Allowed Member List
{% hint style="info" %}
This feature requires activation by Zoom. Contact your Zoom account team to enable this capability.
{% endhint %}
**A Business Hours tool enables conditional routing and messaging based on schedules**
Administrators can configure a Business Hours tool that enables customized greetings, instructions, and routing rules based on business hours and holiday schedules. The tool supports custom time zones, business hours, and holiday configurations with options for recurring schedules. Each Business Hours tool can be referenced in Skills to create conditional logic.
For example, it can route calls to live agents during open hours or display closure messages during off-hours. The tool applies to both inbound and outbound scenarios.
{% hint style="info" %}
This feature requires activation by Zoom. Contact your Zoom account team to enable this capability.
{% endhint %}
####
Multi-instance deployment for multiple locations and departments
The Zoom Virtual Agent multi-instance platform with Zoom Phone delivers scalable, personalized voice experiences across many locations or departments. Organizations manage brand-level configurations centrally while allowing store- or branch-specific customization for greetings, knowledge bases, and skills. By using Zoom Phone's routing data, Zoom Virtual Agent determines the correct department or store extension automatically, which removes the need to set up per-store routing by hand. Hierarchical role management, version control, and selective publishing simplify administration, speed up deployment, and keep the customer experience consistent across all sites.
**Note:** Instance and instance-group filters let administrators scope engagements to a single instance (for example, one store location) or an instance group (for example, a regional group of stores), alongside the existing virtual agent and version filters. These filters apply across the Overview, Virtual Agent Performance, and Conversation Details reports.
{% hint style="info" %}
This feature requires activation by Zoom. Contact your Zoom account team to enable this capability.
{% endhint %}
### Virtual Agent Performance and Analytics
The following sections describe the reports that Zoom Virtual Agent offers.
Administrators can create a single subscription containing the different reports described in this section. They can filter subscriptions by specific agents, agent groups, or report types such as Overview, Virtual Agent Performance, Knowledge Performance, Query Insight, Conversation Details, and Consumer Engagement Surveys. Recipients receive one email with all selected reports, labeled by report type and agent scope.
Reporting subscriptions now offer expanded delivery cadence, dashboard coverage, and email content. Admins can choose any day of the week for weekly delivery or switch to daily delivery. A broader selection of entitled dashboards is available to subscribe to, and the report email template has been revised to include more metrics and an AI-generated summary.
####
Data Dictionary provides transparent metric definitions for Virtual Agent analytics
Zoom Virtual Agent analytics includes an integrated Data Dictionary. Administrators can browse datasets and reports to understand metric definitions, calculations, filters, and interpretation guidance. Definitions can be exported for sharing, helping clarify KPIs and encourage adoption across teams.
####
Dashboard metrics support in-place drill-down and column tooltips
Administrators can click any metric on the Zoom Virtual Agent dashboard to drill into the conversations behind it. Active dashboard filters carry through automatically, so you can move from the conversation list to an individual transcript and back without losing context. Every report table also adds tooltips on its column headers. Hover over a metric to see its definition and a direct link to the full Data Dictionary entry.
####
The Virtual Agent Performance report details Tool, Skill, and latency outcomes during end-user engagement
The Virtual Agent Performance reporting provides data for key AI agent performance indicators.
* **Key Metrics:** Unique and total tools invoked, unique and total skills invoked, latency metrics (P90/P50 first-token latency), latency trend, containment rate, escalation rate, engagement usage, and cost savings.
* **Initiated metric**: A new enhanced metric that measures when and how engagements are initiated, providing clearer visibility into how users enter the virtual agent experience. This metric helps administrators assess entry-point performance, evaluate routing strategies, and understand how changes to workflows or entry surfaces influence overall engagement behavior.
* **Total message metrics**: Provides aggregated counts of all messages exchanged across Zoom Virtual Agent and Zoom AI Expert Assist engagements. These metrics help administrators analyze conversational volume, compare message patterns across channels, and better understand the multi-modal load on AI-driven support workflows.
* **Guardrails**: An aggregate view of back-end safety guardrail activity, including total breaches, breach rate per 1,000 engagements, a category breakdown, and a trend over time. Drilling into the aggregate opens the affected conversations in the Conversation Details report, pre-filtered by the selected guardrail category. Guardrail reporting covers both voice and chat.
* **Filters:** Date range, agent type (voice, chat, chatbot), and specific AI agent.
* **Topic Groups** are also available as a filter and display dimension across analytics reports, enabling supervisors to view performance data at the business-category level.
* **Use Case:** Helps administrators evaluate AI agent efficiency, monitor tool and skill utilization, and track latency and containment trends to improve performance and ROI.
{% hint style="info" %}
The AI Agent report has a cost savings setting that affects virtual agent performance based on a preset token cost.
{% endhint %}
####
The Knowledge Performance report provides details on how knowledge articles are used during engagements
This report offers insight to how your Knowledge Base articles contribute to resolving your customers' issues.
* **Key Metrics**: Knowledge-assisted engagements, knowledge-assisted engagement rate, articles referenced, and average articles referenced per engagement.
* **Filters**: Date range, agent type (voice, chat, chatbot), and specific AI agent, and article source.
* The article source filter has two values — AI-suggested and Manually added — and each article row shows a source label, so administrators can audit how much the recommendation engine is contributing.
* **Use Case**: Helps administrators assess the effectiveness of knowledge content and optimize it for improved resolution rates.
* **KB Active/Inactive Badging**: This feature adds a badge-style icon to KB articles that were replaced or updated. Both the old and new versions remain visible in the ZVA Report with clear status indicators showing whether each article is active or inactive.
####
The Query Insight report offers a detailed view of the consumer queries handled by virtual agents
This report provides you with a bird’s-eye view of query data, including query volume, match coverage, engagement linkage, and topic-level outcomes.
* **Key Metrics:** Total queries, average queries per engagement, popular topics, solved topics, escalated topics, and query match trends.
* **Filters:** Date range, agent type (voice, chat, chatbot), AI agent, query match type, article, topic, engagement outcome, knowledge base, and language.
* **Use Case:** Helps administrators and analysts understand query behavior, assess topic coverage, and identify areas for improving response quality and automation performance.
Zoom Virtual Agent Analytics also offers Topic Groups, a two-level hierarchy that automatically organizes granular topics into business-aligned categories like "Billing and Payments" or "Shipping and Delivery." Supervisors can identify key contact drivers at a glance through the category level while retaining the ability to more deeply analyze specific topics.
Redaction rules configured in Zoom Virtual Agent tenant preferences now apply uniformly across the Query Insights dashboard, including the query list and drill-down views. These surfaces previously rendered raw utterances regardless of the configured ruleset; they now honor the same redaction ruleset as the rest of the product.
####
Comparison analytics show two data slices side by side
Administrators can compare two slices of data directly within the Zoom Virtual Agent dashboards, without exporting or switching screens. From any report, click Compare, select a set of filters, and the dashboard re-renders in a two-cohort layout so you can read the two slices against each other. Comparison analytics work alongside the dashboard's drill-down and cross-filtering features.
####
Cross-filtering recalculates dashboard visualizations by dimension
Supervisors can click a dimensional value across the Query Insights, Conversation Details, Knowledge Performance, and Agent Performance pages to cross-filter every visualization on the page. The clickable values include topics, agents, tools, or chart segments.
A filter chip shows the active selection and can be cleared by clicking **x**, selecting the value again, or clicking outside it. Multiple cross-filters can be combined with each other and with the existing filter-bar settings to show intersecting results. Hovering over a filterable element shows a pointer cursor and a tooltip previewing the action.
####
The Conversation Details report provides engagement-based conversation insights for virtual agents
The Conversation Details reporting provides data for key AI agent engagement indicators.
* **Key Metrics:** Total engagements, multi-turn engagements, average and total interaction duration, and engagement logs with metadata (topics, tools, skills, escalation reasons, and responders). For multi-instance deployments, the report also adds an **Instance** column that identifies which instance handled each engagement.
* **Filters:** Date range, agent type (voice, chat, chatbot), AI agent, tool, engagement outcome, last responder, skill, topic, transfer accepted (voice only), and AI agent ended (voice only) guardrail trigger, and — for multi-instance deployments — instance and instance group.
* **Use Case:** Enables administrators to analyze engagement quality, track AI agent and skill performance, and review conversation-level behavior for optimization, operational oversight, and internal reporting.
When a conversation trips a back-end safety guardrail, the Conversation Details report flags it with a guardrail badge on the affected engagement and a panel showing which guardrail categories fired and on which turn. Administrators can filter the engagement list by breach category — for example, violent crimes, code-interpreter abuse, privacy, or defamation — to review only the conversations that tripped a rule. These are Zoom's back-end safety guardrails, not guardrails an administrator writes into Agent Guidance.
Conversation Details also includes a configurable visualization strip at the top of the dashboard. You can build your own visualizations for variables — both system and custom — choosing between pie, bar, and line charts. Visualizations can be added, moved, and resized, and your settings persist at the account level for your next session. Pinned charts respect the page's active filters, giving you a clear view of variable distribution right where you need it.
####
The Consumer Engagement Surveys report presents consumer feedback collected from post-engagement surveys in Zoom Virtual Agent
This report collates post-engagement survey feedback to measure consumer satisfaction, resolution success, and opportunities for improvement.
* **Key Metrics:** Assigned to agent, launches, responses, response rate, completion rate, launches trend, survey question responses, and free-text feedback distribution.
* **Filters:** Date range, AI agent, and bot (including variants if multiple exist).
* **Use Case:** Enables administrators to assess consumer sentiment, measure chatbot performance, and identify opportunities for improving engagement quality and user satisfaction.
####
Implicit resolution metrics infer customer outcomes from engagement behavior
Administrators can gauge virtual agent performance with implicit resolution, which automatically infers user resolution and satisfaction based on real behavior, such as workflow completion and engagement levels. AI-based behavioral analysis acts as a judge.
This metric provides an alternative to traditional surveys when explicit CSAT response rates are low, and it works alongside existing survey-based metrics rather than replacing them. Implicit resolution data is available in the Zoom web portal, API, or downloadable reports, helping administrators optimize Zoom Virtual Agent effectiveness.
The logic that determines implicit resolution has been enhanced to read conversations more precisely, reducing misclassifications in both directions. As a result, the Resolution Rate more accurately reflects actual outcomes. Administrators may see a slight metric shift as the improvement takes effect.
####
The Engagement Log provides a detailed record of each AI and voice agent interaction, including outcomes, duration, and transcripts for review
The engagement log provides a detailed view of every interaction handled by your AI or voice agents. It helps you analyze how each conversation was managed, whether it was contained by the AI or escalated to a live agent, and the key performance metrics of that engagement.
This report is valuable for supervisors, administrators, and analysts who need visibility into agent effectiveness, customer experience, and system performance.
**Key columns in the engagement log**
* **Engagement ID**: A unique identifier for each interaction. You can use this ID to search for and review specific conversations.
* **Channel**: Indicates whether the engagement came through a chat agent or a voice agent.
* **Source**: Shows the entry point, such as Zoom Contact Center, where the interaction was initiated.
* **Consumer**: The customer or end user who initiated the engagement. This may appear as a phone number for voice calls or a username for chat sessions.
* **AI Agent**: Identifies which AI agent handled the interaction.
* **Tool / Skill / Topic / Article**: Displays which tools, skills, or knowledge resources the AI agent used during the engagement, such as "get ticket status" or "Agent Transfer Request."
* **Engagement outcome**: The final result of the engagement. "Containment" means the AI agent resolved the issue without escalation. "Escalation" indicates the interaction was transferred to a live agent. Escalation reasons are shown when applicable, such as "Escalated by design" or "Escalation requested."
* **Start time / End time**: The timestamps of when the engagement began and ended.
* **Last responder**: Indicates who completed the engagement, either the AI agent or a human agent after escalation.
* **Engagement duration**: The total time the engagement lasted.
* **Interaction turns**: The number of back-and-forth exchanges between the AI agent and the customer.
* **Interactive duration**: How long the customer was engaged in the interaction, which is especially useful for voice calls.
* **Transfer accepted / AI agent exited**: Shows whether the transfer to a live agent was successful and whether the AI agent left the engagement.
This reporting supports auditability, review, and training by providing audit trails that can improve both AI and human agent performance.
For example, if the IT Helpdesk Agent shows a high number of escalations for ticket creation tasks, supervisors can review transcripts to identify why users are requesting a human agent and update the AI’s configuration or training data.
**Agent Tracing provides step-by-step visibility into virtual agent decision-making**
When reviewing an engagement, administrators can use Agent Tracing to see the full decision chain, from input processing through the final response. This supports troubleshooting, Quality Assurance review, and performance optimization.
Agent Tracing provides a unified, human-readable trace linked to a specific engagement. It includes a full customer transcript alongside system actions. Each pipeline step is time-stamped with start and end durations.
Agent Tracing includes:
* Input capture
* NLU/intent classification
* Search and retrieval steps (i.e., query rewriting and source selection)
* LLM inference
* Tool/function calls
* Response rendering
Agent Tracing also provides engagement-level summaries that speed up troubleshooting. A tool-call summary lists every tool invoked during the engagement, each with its success or failure status, so administrators no longer have to scan the trace turn by turn to find a specific call. A variable summary gives a consolidated view of every variable set during the same engagement. Together, these summaries surface script executions, skill triggers, and the overall interaction path at a glance.
Agent Tracing also captures detailed tracing and debug information for built-in tools, including the SMS tool, email tool, save response, and end conversation. For each tool invocation, the trace records the reason the tool was selected, the input parameters passed, variable resolution output, execution status, and any errors returned. The trace also shows why a conversation ended, including whether it was triggered by an instruction or by a timeout following nudge retries.
{% hint style="info" %}
You can find Agent Tracing within Virtual Agent analytics. It’s accessible from the Engagement Log when reviewing individual interactions.
{% endhint %}
For validating agent changes before deployment, see the Evaluation section below.
####
Evaluation validates agent changes before deployment through scenario testing and version comparison
When changing an agent's configuration, guidance, Skills, or Tools, administrators need to verify those changes won't degrade performance before publishing to production. Evaluation provides a structured testing environment for this.
Admins can run evaluations against any published or historical agent version, not just the current draft. Evaluations run as read-only against the selected version snapshot, so the active agent isn't modified. This anchors results to a specific, known configuration.
You can then evaluate two agent versions side by side using the same scenario set. Results display per-version metrics (such as scenario success rates, behavior match scores, and more) so admins can see exactly where performance changed between versions. A detailed transcript view lets admins drill into individual test cases to see how responses differ.
Before an evaluation executes, the system validates test case integrity. That includes checking for missing tools, undefined variables, or schema mismatches. This prevents wasted evaluation runs and eliminates false negatives caused by configuration errors rather than actual agent behavior.
Evaluation supports administrators' efforts to make production-ready agent changes, which can help speed up iteration cycles and manage risk. Combined with Agent Tracing, it creates a robust diagnostic and validation workflow.
####
The ZVA Receptionist dashboard gives smaller deployments a single-screen performance view
For smaller ZVA Receptionist deployments that don't need the full analytics suite, a single-screen dashboard presents performance in plain language. Every value is rolled up from existing metrics, so no new instrumentation is required. A hero row shows calls handled, phone time handled (expressed as hours), and cost saved. A containment ring pairs with an escalations count and a "Read these conversations" link that drills straight into the underlying engagements. The dashboard also surfaces the top reasons people called and a resolved/unresolved outcomes breakdown, and a per-agent table and simplified conversation log round out the screen — all scoped by a single filter row.
### Supplemental Details
The following sections detail miscellaneous information related to Zoom Virtual Agent.
####
Privacy and transparency are at the core of Zoom AI
Zoom AI follows a federated-model approach that places user data privacy and transparency at its core. Zoom does not use customer content (such as audio, video, chat, screen-sharing, attachments, whiteboards, and reactions) to train Zoom’s or its third-party AI models.
For more details on how data flows, is processed, stored, and retained, and how feature controls work, see the [Zoom AI Security and Privacy Whitepaper](https://www.zoom.com/en/products/ai-assistant/resources/privacy-security/#ai-companion-features).
####
Users must be granted role access to modify Zoom Virtual Agent settings
Users must be granted role access to Zoom Virtual Agent before configuring or modifying any settings.
{% stepper %}
{% step %}
As a Zoom account owner or admin authorized to edit roles, navigate to the [Role Management](https://zoom.us/role#/) page on the web portal.
{% endstep %}
{% step %}
Click the pencil icon to the right of the role you are editing.
{% endstep %}
{% step %}
Navigate to the **AI Studio** subsection and define the role access as desired.
{% endstep %}
{% step %}
Repeat for any additional roles.
{% endstep %}
{% endstepper %}
**Zoom Virtual Agent does not support granular role access at this time**
Zoom Virtual Agent does not support granular role access for product configuration or setting changes. User roles granted access can modify all product settings and configurations without limit. Account admins should exercise caution when granting role access, but do have the ability to control view and/or edit access to the various subcomponents.
**Access to the Usage (Billing) dashboard can be granted separately**
Administrators can grant access to the Zoom Virtual Agent Usage (Billing) dashboard independently of the other Zoom Virtual Agent dashboards in AI Studio. A separate Virtual Agents permission under Usage Reports controls the billing dashboard on its own, so finance and operations users can review billing data without seeing unrelated product dashboards, and product and engineering users can use the operational dashboards without visibility into billing.
####
Messages and files from conversations can be configured for permanent or soft deletion over time
Message and file data generated through use of Zoom Virtual Agent can be configured for automatic deletion over time. With this option, data can be deleted in daily or monthly increments ranging from 1–30 days or 1–12 months.
Data marked for permanent deletion will be irretrievably removed from an account and can't be restored upon request. Data marked for soft deletion will be removed from an account but can be retrieved for up to 30 days upon request before permanent deletion.
####
Zoom Virtual Agent is available globally, but data processed by AI can't be localized
Zoom Virtual Agent is available to all global commercial customers in [unrestricted countries and regions](https://support.zoom.us/hc/en-us/articles/203806119-Restricted-countries-or-regions). Communications content, like recordings and transcripts, can be localized in admin settings via the Zoom web portal. However, data processed using AI/LLM features can't be localized to specific regions. The [Zoom AI for Contact Center and AI Expert Assist Security and Privacy Whitepaper](https://www.zoom.com/en/products/ai-assistant/resources/privacy-security/zoom-contact-center/) has more information on Zoom Contact Center data handling.
####
Change logs and version comparisons are available for published Agents
Admins can view a change log when publishing agent modifications. The log shows all updates to the draft version, including additions, deletions, and modifications, with color coding and owner info for each change. Admins can review changes before deciding to publish or cancel.
Admins can also compare two published agent versions. The comparison covers all agent components, including greeting, guidance, greeting skills, skills, and tools. Visual indicators highlight any differences, replacing manual tracking using external tools.
####
Zoom Virtual Agent supports multiple languages and dialects
The next-generation Zoom Virtual Agent supports multiple languages and dialects across chat and voice. The list of currently supported languages includes:
|
Arabic
Arabic (Gulf)
Cantonese
Catalan
Chinese (Simplified)
Chinese (Traditional)
Danish
Dutch (Netherlands)
English (Australia)
English (Indian)
English (Italian)
English (New Zealand)
English (Scottish)
English (Singaporean)
English (UK)
|
English (US)
French (France)
French (Canada)
Finnish
German
English (Welsh)
Hebrew
Hindi
Indonesian
Italian
Japanese
Korean
Norwegian
Malay
|
Polish
Portuguese (Portugal)
Portuguese (Brazil)
Romanian
Russian (Russia)
Spanish (European)
Spanish (Mexico)
Spanish (US)
Swedish
Thai
Turkish
Vietnamese
|
| ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
The system displays available voice actors specific to the selected language and supports localized text-to-speech capabilities for numbers, dates, phone numbers, and other formatted content. This expansion enables organizations to provide more localized customer experiences across multiple regions.
{% hint style="info" %}
Additional languages will be added over time. This document will be updated as new language support launches. Adding new languages requires activation by Zoom. Contact your Zoom account team to enable additional language capabilities.
{% endhint %}
**Expanded multi-language support for chat agents**
Zoom Virtual Agent chat experiences now support an expanded set of languages, allowing administrators to configure a single chat agent that serves customers across multiple regions and languages. This enhancement eliminates the need to create and maintain separate agents for each supported language, returning to the streamlined multilingual approach that was available in ZVA 1.0.
When creating a chat agent, administrators can select multiple languages that the agent will support. Once configured, the system automatically determines the engagement language based on factors such as welcome screen selections, browser HTML language settings, and other contextual signals. The selected language remains consistent throughout the conversation, providing a coherent experience for the end user.
Machine translation handles the localization of greeting messages, flow content, and button text automatically. This means administrators don't need to translate each element of the chat experience when adding language support.
The current release focuses on language selection at the start of the conversation. Administrators configure supported languages during chat agent creation, and the system locks in the engagement language when the conversation begins.
####
Zoom Virtual Agent is not available to Zoom for Government customers
Zoom Virtual Agent is not available to Zoom for Government (FedRAMP) customers at this time.
---
# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.
## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation by asking a question.
Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters:
```
GET https://library.zoom.com/business-services/zoom-virtual-agent/zoom-virtual-agent-explainer.md?ask=
&goal=
```
`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case.
The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.
Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.
