AI Studio

By Zaid Chaudhry

• 7 articles

What is Agent Studio?

Agent Studio is where you build and manage AI agents for your support operation. An AI agent is an automated assistant that can handle incoming customer conversations on your behalf. It answers questions, looks up information, completes tasks, and hands off to a human agent when needed. What an AI agent can do - Respond to customer messages automatically, 24/7 - Answer questions using your Knowledge Base (your website content, uploaded files, or help center articles) - Use tools to take actions like looking up order details, checking status, or calling external APIs - Detect when a conversation needs a human and escalate automatically - Communicate in multiple languages How Agent Studio is organized Agent Studio has four sections, accessible from the sidebar. Knowledge Base Where you add the content your agent will use to answer questions. You can add your website by URL or upload files directly. The agent only knows what you put here. The more complete your Knowledge Base, the better the agent performs. Tools Actions your agent can take beyond answering questions. Tools can call external APIs, look up data from other systems, or run built-in platform actions. You configure what tools are available, and the agent decides when to use them. Agents Where you create and configure individual AI agents. Each agent has a name, a personality, a set of guidelines, and a selection of Knowledge Base content and tools it is allowed to use. Workflows Scripted conversation paths for scenarios that need a fixed structure. Workflows send pre-written messages, present choices and buttons, and follow a set path. Unlike free responses from the agent, workflows do not use AI to generate their messages. How it works in a conversation When a customer sends a message to an inbox that has an agent assigned, the agent responds automatically. It reads the message, searches its Knowledge Base if needed, uses any relevant tools, and replies. If the agent determines it cannot help, or hits one of your defined escalation conditions, it hands the conversation to your human support team. Before you start To get the most out of Agent Studio, have the following ready: - Your website URL, or files (PDFs, text documents) containing your product and support content - A clear idea of what questions your customers ask most often - The inboxes you want the agent to handle

Last updated on Apr 12, 2026

Creating and Configuring an AI Agent

Once you have content in your Knowledge Base, you are ready to build an agent. This article walks through every field in the agent configuration form. Getting started Navigate to Agent Studio in the sidebar, then select Agents. Click Create Agent in the top right of the panel. Basic information Name (required) A name for your agent. This is visible to your team, not to customers. Use something descriptive, for example "Support Agent" or "Billing Assistant". Description (required) A short summary of what this agent does. This helps your team understand the agent's purpose at a glance, for example "Handles general support questions for the web widget inbox." Persona Voice (required) How the agent sounds when it talks to customers. Choose one: - Friendly - Warm and approachable tone - Professional - Business-like and formal - Helpful - Supportive and solution-oriented Language settings Supported languages Select the languages this agent can respond in. Currently supported: English, Arabic, Spanish. Default language The language the agent falls back to when it cannot detect the customer's language. Auto-detect language When enabled, the agent automatically detects the customer's language from their message and responds in kind. Disable this if you want the agent to always reply in the default language regardless of what the customer writes. Guidance Guidance rules are plain-language instructions that shape how your agent behaves. Each rule is a single statement. Examples: - "Always greet the customer by name if available" - "Never discuss competitor pricing" - "If a customer asks about a refund, collect their order number first" Click + Add guideline to add more rules. You can add up to 20. Remove individual rules using the trash icon next to each one. Tip: Be specific. "Be polite" is less useful than "Apologize when a customer expresses frustration before attempting to solve their issue." Knowledge Base Select which websites and files this agent is allowed to use when answering questions. Toggle on the content you want and the agent will search it when responding to customers. - Only content with a Complete status can be assigned - Content that is still processing, pending, or failed is shown but cannot be toggled on - Use the search bar to filter by name If the list is empty, go to the Knowledge Base section first and add your content. Tools Select which tools this agent is allowed to use. Tools let the agent take actions beyond answering questions, for example looking up data from an external system or triggering an action. Only tools configured in the Tools section of Agent Studio appear here. Workflows Workflows are scripted conversation paths you build in the Workflows section of Agent Studio. Unlike free responses, a workflow sends exact pre-written messages, presents fixed choices and buttons, and follows a set path. No AI generation is involved in those steps. Toggle on any workflows you want this agent to use. Entrypoint workflow One workflow can be marked as the entrypoint (star icon). The entrypoint workflow runs automatically at the start of every new conversation, before the agent begins responding freely. Use this for things like a welcome message, collecting a customer's account number, or routing by topic. To set an entrypoint, toggle a workflow on, then click the star that appears next to it. Only one workflow can be the entrypoint at a time. If no entrypoint is set, the agent starts in free-response mode immediately. Escalation Escalation triggers Conditions under which the agent hands the conversation to a human agent. Each trigger is a plain-language description of a scenario. Examples: - "Escalate when the customer mentions a complaint or refund" - "Escalate if the customer asks to speak to a manager" - "Escalate for any billing disputes" Click Add Trigger to add more. If no triggers are defined, the agent uses default escalation behavior. Always Silent Escalation When enabled, the agent hands off without sending any message to the customer. The conversation simply appears in your team's queue. When disabled, the agent may send a brief handoff message before transferring. Inbox assignment Select which inboxes this agent should be active on. When a customer sends a message to an assigned inbox, the agent responds automatically. Each inbox can only have one agent assigned at a time. If an inbox is already assigned to another agent, unassign it there first. Saving your agent Click Save when you are done. Your agent will appear in the Agents panel. From there you can: - Edit to modify any configuration - Clone to duplicate the agent as a starting point for a new one - Delete to permanently remove the agent

Last updated on Apr 12, 2026

Setting up your Knowledge Base

Your Knowledge Base is the content your AI agent uses to answer customer questions. The agent can only respond based on what you put here, so the more complete and accurate your content, the better the agent performs. Accessing the Knowledge Base Navigate to Agent Studio in the sidebar. The Knowledge Base panel is the middle panel on desktop, or the second section on mobile. Two ways to add content You can add content by uploading a file or by adding a website. Adding a website This is the recommended approach if your help center or product documentation is already published online. 1. Click Add Website at the top of the Knowledge Base panel. 2. Enter your website URL, for example https://help.yourcompany.com. If you leave off https://, it will be added automatically. 3. Click Discover URLs. The system scans your website and returns a list of all the pages it found. This may take a moment depending on the size of the site. 4. Review the list of discovered pages. Each one shows the page title and a short description. Select the pages you want the agent to learn from. You do not have to include everything. 5. Optionally give it a custom Title (defaults to the domain name). 6. Click Add Website. The selected pages are now being processed in the background. You will see the status change in the Knowledge Base panel. You can close the dialog and keep working while it finishes. You can add multiple websites. Each domain is added separately, so if your content lives across help.yourcompany.com and docs.yourcompany.com, add each one on its own. Uploading a file For content that is not published online, like internal guides, product manuals, or PDFs. 1. Click Upload Document at the top of the Knowledge Base panel. 2. Select a file from your computer. 3. The file is uploaded and processing starts automatically. Supported formats include PDF and text-based files. Processing time depends on file size. Understanding the status badges After you add a website or upload a file, you will see a colored status badge next to it. - Processing (blue, pulsing) means content is being read and indexed. Wait for this to finish. - Complete (green) means it is ready to use and can be assigned to an agent. - Failed (red) means something went wrong. You can retry. - Partial (amber) means some pages succeeded and others failed. The content that was processed is still usable. - Pending (yellow) means it is queued and will begin processing shortly. Managing your websites Retry failed pages If some pages failed during processing, you can retry just the ones that failed without re-processing everything else. Refresh If your website content has changed since you last added it, you can refresh to re-read all the pages and update the agent's knowledge. Rediscover pages If new pages have been added to your website since you originally added it, use this to find them. You can then choose which new pages to include. Add pages Add specific new URLs to an existing website without running a full rediscovery. Cancel If processing is taking too long, you can cancel it. Any pages already processed will be kept. Delete Remove a website or file from your Knowledge Base entirely. If it is currently assigned to any agents, you will be warned before deletion. Assigning content to an agent Adding a website or file to the Knowledge Base does not automatically make it available to any agent. You must assign it. 1. Go to Agents, select the agent you want, and click Edit. 2. Scroll to the Knowledge Base section. 3. Toggle on the websites or files this agent should use. 4. Save. Only content with a Complete status can be assigned. Tips for a good Knowledge Base - Start with your most common questions. Look at your conversation history and add the content that answers what customers ask about most. - Keep content focused. One topic per page or article works better than long pages covering everything. - Use clear headings. Well-structured content with headings produces better results when the agent searches for answers. - Update when things change. If your website content changes, refresh the website in the Knowledge Base. Outdated content leads to wrong answers. - Start small and expand. Begin with your most important content, make sure the agent handles it well, then add more.

Last updated on Apr 12, 2026

Escalation: Handing Off to a Human Agent

Not every conversation can be resolved by an AI agent. Some situations need a human. Escalation is how the agent recognizes this and transfers the conversation to your support team. How escalation works When the agent detects that a conversation needs human attention, it stops responding and makes the conversation available to your team. Your agents will see it in their queue like any other conversation. There are two ways escalation can happen: Automatic escalation The agent uses its own judgment to decide when it cannot help. For example, if a customer is clearly frustrated and the agent has no useful answer, it may escalate on its own. Trigger-based escalation You define specific scenarios where the agent should always escalate. These are configured in the agent's settings under Escalation. Setting up escalation triggers 1. Go to Agent Studio in the sidebar. 2. Select Agents and click Edit on the agent you want to configure. 3. Scroll down to the Escalation section. 4. Click Add Trigger. 5. Write a plain-language description of when the agent should escalate. Each trigger is a sentence describing a scenario. Examples: - "Escalate when the customer mentions a complaint or refund" - "Escalate if the customer asks to speak to a manager" - "Escalate for any billing disputes" - "Escalate when the customer provides a legal threat" - "Escalate if the customer has asked the same question more than twice" You can add as many triggers as you need. Remove any trigger using the trash icon next to it. Tip: Write triggers the way you would explain them to a new team member. The agent interprets them as instructions, so clear and specific descriptions work best. Silent vs. visible escalation When the agent escalates, it can either tell the customer or stay quiet about it. Always Silent Escalation (toggle) When this is turned on, the agent hands off the conversation without saying anything to the customer. The customer does not see a message like "Let me transfer you to a team member." The conversation simply appears in your team's queue and a human can pick it up. When this is turned off, the agent may send a short message to the customer letting them know a human will be taking over. To configure this: 1. Go to Agent Studio, select Agents, and click Edit on your agent. 2. Scroll to the Escalation section. 3. Toggle Always Silent Escalation on or off. When to use silent escalation Silent escalation works well when: - You want the handoff to feel seamless to the customer - Your team picks up conversations quickly and the customer will not notice a gap - The agent has already collected enough context that the human can continue naturally Visible escalation works better when: - Response times from your team may be slower, and you want the customer to know a human is coming - You want to set expectations, for example "A team member will be with you shortly" What happens after escalation Once the agent escalates, it stops responding to that conversation. Your team sees the conversation in their queue with the full message history, including everything the agent said and any information it collected. If no escalation triggers are defined, the agent still uses its own judgment to escalate when it cannot help. Adding triggers gives you more control over exactly when that happens.

Last updated on Apr 12, 2026

Building Workflows

Workflows are scripted conversation paths. Every message in a workflow is pre-written by you. The agent follows the script exactly, sending fixed messages, presenting choices, and collecting input. Use workflows when you need a predictable, structured interaction, for example collecting a customer's details before looking up their order, or presenting a menu of options at the start of a conversation. Accessing workflows Navigate to Agent Studio in the sidebar. The Workflows panel is in the left panel on desktop. Click Create Workflow to open the visual editor. How the editor works The workflow editor is a visual canvas where you place nodes and connect them with lines. Each node is a step in the conversation. The flow moves from top to bottom, followingthe connections you draw between nodes. Every workflow starts with a Start node and ends with an End node. You add steps in between to build the conversation. To add a node, click the + button or right-click on the canvas and select the type of node you want. To connect two nodes, drag a line from one node's output handle to the next node's input handle. To configure a node, click on it. The settings panel will appear on the right side where you can edit its properties. Types of nodes Conversation / Collect Input This is the main building block. A conversation node can contain one or more "cards" that do different things. Talk cards send content to the customer: - Send message sends a text message you have written - Send image sends an image you upload - Send attachment sends a file (PDF, audio, video, etc.) - Listen cards wait for the customer to respond: - Collect choice presents a set of options for the customer to pick from (like buttons). You define each option and where it leads. The workflow pauses until the customer selects one. - Collect freeform input waits for the customer to type a response. You can add validation rules such as email, phone number, URL, number, date, or a custom pattern. When a listen card collects a response, it saves the value to a variable that you can use later in the workflow. Rules for cards within a single node: - Only one message card per node - Only one image card per node - Only one listen card (choice or freeform input) per node - A message or image card should come before a listen card Branch A branch node splits the flow based on conditions. You define one or more conditions that check the value of a variable, and the workflow follows the matching path. For example, if you collected the customer's issue type in a previous step, you can branch on it: - If issue type equals "billing", go to the billing flow - If issue type equals "technical", go to the technical flow - Otherwise (else), go to a general flow Available comparisons include: equals, not equals, contains, does not contain, starts with, ends with, greater than, less than, is empty, is not empty, and matches a pattern. Tool A tool node runs one of your configured tools during the workflow. For example, it can look up an order, check an account status, or call an external API. You select which tool to run, provide any required inputs (which can reference variables collected earlier), and optionally save the result to a variable for use in later steps. Attachment An attachment node sends a file to the customer. You upload the file and optionally add a caption. Supported types include images, video, audio, and documents. Trigger Workflow This node starts a different workflow from within the current one. Use this to break complex flows into smaller, reusable pieces. For example, a main welcome workflow could trigger a "collect shipping address" workflow and then continue after it completes. End The end node marks the end of the workflow. Once the workflow reaches this node, the agent returns to free-response mode and can answer further questions using its Knowledge Base and tools. Variables Variables let you store and reuse information throughout a workflow. When a listen card collects input from the customer, you save it to a variable. You can then reference that variable in messages, branch conditions, or tool inputs. To insert a variable in a text field, type {{ and select from the list. Workflow variables are values you collect or set during the flow, for example vars.email or vars.order_id. System variables are built-in values available automatically: - contact.name - the customer's name - contact.email - the customer's email address - contact.phone - the customer's phone number - conversation_id - the current conversation ID You can also use a Set Variable node to manually set a variable to a specific value at any point in the workflow. Saving your workflow When you are ready, click Save. You will be asked to give the workflow a name and an optional description. The editor will check for common issues before saving: - The Start node must be connected to at least one other node - Message cards cannot be empty - Listen cards must save their response to a variable - Choice options cannot be empty - Tool nodes must have a tool selected - Trigger Workflow nodes must have a target workflow selected If any issues are found, you will see a message explaining what needs to be fixed. Assigning a workflow to an agent After saving, the workflow appears in the Workflows panel but is not active until you assign it to an agent. 1. Go to Agents, select your agent, and click Edit. 2. Scroll to the Workflows section. 3. Toggle on the workflows you want this agent to use. 4. To make a workflow run at the start of every conversation, click the star icon next to it to mark it as the entrypoint. 5. Save. Example: Collecting customer details before starting Here is a simple workflow you could use as an entrypoint: 1. Start node connects to a Conversation node. 2. The Conversation node sends a message: "Hi! Before I help you, can I get your email address?" 3. A listen card (freeform input with email validation) waits for the response and saves it to vars.email. 4. Connect to another Conversation node that sends: "Thanks! How can I help you today?" 5. A listen card (freeform input) collects their question and saves it to vars.question. 6. Connect to an End node. After the workflow ends, the agent switches to free-response mode and answers the customer's question using its Knowledge Base, already knowing their email.

Last updated on Apr 12, 2026

How AI Works During Workflows

While the messages your workflow sends are fixed and pre-written, the AI stays active in the background whenever the customer responds. This article explains what the AI does during a workflow and how it affects your AI usage. Why the AI is involved Workflows present choices, ask questions, and collect input. But customers do not always respond in the exact format you expect. Instead of failing on anything that is not a perfect match, the AI interprets what the customer meant and handles the response naturally. What the AI does Input normalization When a customer replies to a choice or input step, they do not have to respond with the exact wording you defined. If your workflow presents the options "Yes" and "No", the customer can type "yeah", "sure", "nah", or "no thanks" and the AI will interpret their intent and match it to the correct option. The same applies to numbered options. If you present three choices, the customer can reply "2" or describe what they want in their own words. Format handling When collecting structured input like phone numbers or emails, the AI normalizes the response into the expected format. For example, if a customer types their phone number as "345 678 9012", the AI can convert it to "+393456789012" based on the validation format you set. Clarification If the customer's response is ambiguous, the AI will ask a brief clarifying question before continuing the workflow. Invalid input If the customer provides something that does not match what the workflow expects, the AI will politely ask them to try again and explain what is needed. Off-topic messages If a customer asks something unrelated to the workflow during a scripted step, the AI can choose to exit the workflow and respond freely using the assigned Knowledge Base or Tools, or gently guide the customer back to the current step. This depends on the context and what the customer asked. AI usage Because the AI processes every customer response during a workflow, workflows do use AI credits. The scripted messages themselves do not consume AI, but interpreting and normalizing customer replies does.

Last updated on Apr 12, 2026

Setting Up Tools

Tools let your AI agent take actions beyond answering questions. A tool connects to an external API so the agent can look up data, trigger actions, or retrieve information from other systems during a conversation. For example, you could create a tool that looks up a customer's order status, checks their account balance, or creates a support ticket in another system. Accessing tools Navigate to Agent Studio in the sidebar. The Tools panel is in the left panel on desktop. Click Create Tool to open the tool configuration form. Creating a tool The tool form has four sections. Basic information Name (required) A unique name for the tool. This is how the tool appears in your agent configuration and in the Tools panel. Names must be unique across all your tools. Description (required) Describe what this tool does. This is important because the AI reads this description to decide when to use the tool. Write it clearly, for example: "Looks up order status by order ID and returns tracking information." Category An optional label to help you organize your tools. Authentication Choose how the tool authenticates with the external API. Options: - None - no authentication needed - API Key - sends an API key as a header or query parameter. You provide the key name (for example X-API-Key) and the key value. - Bearer Token - sends a token in the Authorization header - Basic Auth - sends a username and password - OAuth 2.0 - authenticates using a token URL, client ID, client secret, and optional scope Request configuration Method The HTTP method the tool uses to call the API: GET, POST, PUT, PATCH, or DELETE. Endpoint URL (required) The full URL the tool calls, for example https://api.yourservice.com/orders/lookup. Custom Headers Any additional headers the API requires. Click + Add Header to add key-value pairs. Authentication headers are added automatically based on your authentication settings, so you do not need to add those manually. Timeout How long the tool waits for a response before giving up, in seconds. Parameters This section appears only for POST, PUT, and PATCH methods. Define the data the tool sends in the request body using a JSON schema. This tells the AI what inputs the tool needs. For example, a tool that looks up orders might need an order_id parameter. The AI will extract this from the conversation and pass it to the tool automatically. Testing a tool After saving a tool, you can test it to make sure it works before assigning it to an agent. 1. Open the tool from the Tools panel. 2. Click Test. 3. If the tool has parameters, fill in test values. 4. Click Run Test. 5. Review the response to confirm the tool is working correctly. Assigning a tool to an agent Creating a tool does not automatically make it available to any agent. You must assign it. 1. Go to Agents, select your agent, and click Edit. 2. Scroll to the Tools section. 3. Toggle on the tools this agent should be allowed to use. 4. Save. How the agent uses tools When a customer asks a question that requires looking something up or taking an action, the agent reads the descriptions of its available tools and decides which one to use. It extracts the necessary information from the conversation (like an order ID or email address), calls the tool, and uses the response to answer the customer. For example: - Customer says: "What is the status of my order #A7X-993?" - Agent recognizes this needs the order lookup tool - Agent extracts the order ID "A7X-993" from the message - Agent calls the tool with that order ID - Tool returns the order status - Agent responds: "Your order #A7X-993 is currently in transit and expected to arrive on April 15." This happens automatically. You do not need to configure trigger conditions or map inputs manually. Where tools are used Tools can be used in two places. In free-response mode When an agent is responding freely to a customer (not in a workflow), it decides on its own when to call a tool based on the conversation. You control which tools the agent has access to in the agent configuration under the Tools section. In workflows You can add a Tool node to any workflow. Unlike free-response mode, a tool node in a workflow runs at a specific step that you define. You choose which tool to run, provide the inputs (which can reference variables collected earlier in the workflow), and optionally save the result to a variable for use in later steps. For example, a workflow could collect a customer's order ID in a conversation node, then pass it to an order lookup tool node, then display the result in the next message. Tips for effective tools - Write clear descriptions. The AI uses the description to decide when to call the tool. "Looks up order status by order ID" is better than "Order tool." - Keep tools focused. One tool should do one thing. A tool that looks up orders and also processes refunds should be split into two separate tools. - Test before assigning. Always test a tool to make sure it returns the expected response before assigning it to a live agent. - Secure your credentials. API keys and tokens are stored securely, but use dedicated API keys with the minimum permissions needed for the tool's purpose.

Last updated on Apr 13, 2026