Skip to main content

# CAAC AI Workflow

  • Updated

    CAAC AI Agent is an intelligent customer service assistant designed by Crescendo Lab for enterprises. It can respond to customers in real time, proactively handle repetitive issues, and automatically hand over to a human agent at the right time. This article will guide you through the AI Agent’s workflow, configuration methods, and how to publish or unpublish, as well as common troubleshooting tips, helping you maximize the service value of AI.

    In this section, we will introduce the complete process of how the AI Agent interacts with customers, including initial conversation judgment, handover to human agents, and how to use the feedback mechanism to continuously improve AI response quality.

    This flowchart demonstrates the process of how the CAAC AI Agent handles new incoming messages:

    1. New message enters CAAC: whenever a new customer message enters the system. Both text and image messages enter the flow below.
    2. Determine whether it meets handover conditions: The system will determine based on pre-set rules (such as specific keywords, sentiment analysis, or topic categories) whether the message needs to be handed over to a human agent.
      • If it meets the handover condition:
        • The conversation will be directly handed over to the team
        • The system will generate a handover note summarizing key points of the AI-customer conversation to help the agent quickly understand customer needs
        • The AI Agent will notify the customer about the handover, and the CAAC agent will take over the response
      • If it does not meet the handover condition:
        • The system will assign the message to the AI Agent
        • The AI Agent generates an instant reply from the knowledge base used by its agent type (Crescendo AI, Chatbase, or OpenAI) and sends it to the customer
    3. User feedback: Regardless of whether the message was replied to by AI or handed over to a human, users can provide feedback on AI behavior and response content. This feedback can later be used to improve the knowledge base and handover rules.
    4. AI conversation count check: If the AI has interacted with the user more than a specific(customized setting) times, it will trigger a handover to a human agent (with a handover note).
    5. User reply waiting time check:
      • If the user does not reply for more than a specific(customized setting) minutes, the system will automatically mark the conversation as resolved.
      • If the user replies within a specific(customized setting) minutes, the conversation will return to the CAAC AI workflow and continue processing.

    - Handover Note (AI handover note)

    When a conversation originally handled by AI needs to be handed over to a human agent, the system will automatically generate a “handover note.”

    This note summarizes the key points of the AI-customer conversation so that the agent can quickly understand the background and customer needs, reducing communication time and improving service efficiency.

     

    - Feedback and Optimization Mechanism (AI feedback loop / AI training)

    The feedback and optimization mechanism allows users to provide ratings for AI responses and handover decisions, helping you review AI performance and identify areas for improvement.

    • Functions and Use Cases

    Message content evaluation: If you find the AI Agent’s reply accurate and helpful, you can give a 👍. If the content is inappropriate for the customer, you can give a 👎 and specify the reason.

    Handover evaluation: Provide feedback on whether AI handed over the conversation at the right time. For example: the AI should have handed over but didn’t, or it handed over unnecessarily.

    • How to provide feedback

    1. Select 👍 or 👎 to quickly record your preference regarding the AI response.

    2. Select the “chart” icon to provide more detailed feedback. For example:

    • Suggestions for improving AI replies
    • Check whether AI referenced correct information
    • Check whether AI handed over correctly
    • Evaluate AI reply quality (completeness, accuracy, tone, etc.)

    If you expand the feedback details, and if the AI handed over the conversation, the system will also show the reason for AI handover, helping you understand the logic and context behind the decision.

    • Feedback Records and Applications
    1. All feedback will be fully recorded for later review and classification
    2. Feedback will not immediately change AI behavior, nor directly update the knowledge base
    3. You can request these records from Crescendo Lab to review AI’s replies and handover performance, and then optimize the knowledge base or handover settings

    - Message Types the AI Agent Can Handle

    The AI Agent reads the content of an image and then applies the same flow it uses for text: if the image matches one of your routing rules, the conversation is handed to the corresponding human agent; if not, the AI answers from the knowledge base or searches your product feed; if the AI cannot answer, the behavior follows the agent's AI Instruction. Your existing routing rules, keywords, and topic settings require no rebuilding or adjustment.

    Image support by AI agent type:

    AI agent type Images Behavior
    Crescendo AI ✅ Reads the image, then handles routing, knowledge base, and product recommendation
    Chatbase ❌ Hands the conversation to a human agent
    OpenAI ❌ Hands the conversation to a human agent
    Self-hosted AI ❌ Hands the conversation to a human agent

    For agents connected to Chatbase, OpenAI, or a self-hosted AI, add a line in the AI Instruction asking customers to describe their question in text. This improves AI coverage.

    ⚠️ August 2026 update｜Stickers are treated as images and are handled by the AI Agent.

    Message type Handled by AI Behavior
    Image ✅ —
    Sticker ✅ Treated as an image
    Video ❌ Handed to a human agent
    File ❌ Handed to a human agent
    Voice ❌ Handed to a human agent
    ⚠️ Limitations of image understanding
    ・Product recognition relies on text or a model number printed on the image. A plain product photo with no text cannot be matched to your product feed; the AI asks the customer for the product name or model instead.
    ・The AI cannot look up that customer's own orders, coupons, membership, or points from a screenshot. It provides general information only.
    ・When the image is blurred, the text is too small, or the AI cannot determine the customer's intent, the behavior follows the agent's AI Instruction: if handover to a human is stated there, the AI hands the conversation over; if not, the AI asks the customer for more detail. To have the AI hand over whenever it cannot answer, state this in the AI Instruction.

    AI Agent Setup Guide

    To set up an AI Agent, please refer to: Tutorials｜AI Agent Set Up

    How to Start/Stop Assigning Conversations to an AI Agent

    Once Crescendo Lab has helped you complete Athe I Agent setup, the AI Agent will appear in your CAAC system. You can enable the AI Agent anytime to automatically reply to conversations and hand them over to human agents when necessary. This section explains how to publish/unpublish the AI Agent, along with assignment rules and considerations.

    - Assign New Conversations to AI Agent

    Follow the steps below to set the AI Agent as the first responder for all new conversations:

    Step 1: Enter “Assignment Settings” and set the default assignment type as “General.”

    If you select “Permanent,” once a conversation is handed over to a human and marked as resolved, when the same conversation reopens, it will not be automatically reassigned to the AI Agent. For more details, refer to: Tutorial｜CAAC Assignment Types, Methods, and FAQs

    Step 2: Create an Auto-Assignment Rule

    • Go to “Assignment Settings” → “Auto-Assignment” and click “Add.”
    • Configure as follows:
      • Condition type: Any plain text message
      • Assignment target: Agent
      • Select target: Your AI Agent in CAAC

    Step 3: Move the newly created rule to the top priority. After saving, all new messages will be automatically assigned to the AI Agent for the first reply.

    - Stop Assigning Messages to AI Agent

    • Scenario 1: You only want to transfer a single chat to a human

    In that chatroom, change the assignment from AI Agent to your desired team or agent. All future messages in that chat will then be handled by the assigned human agent.

    • Scenario 2: You don’t want AI Agent to be the first responder for all messages

    Go to “Assignment Settings” and disable the setting. After disabling, new messages will no longer be automatically assigned to the AI Agent.

    Please check if there are any ongoing conversations still assigned to the AI Agent. If so, manually reassign them to the designated human team or agent to ensure the AI Agent no longer responds.

    FAQs

    Q: Why didn’t the AI Agent reply when a user sent a message (e.g., an image)?

    A: First check the agent type. For Crescendo AI, the AI reads the image and replies; when the image is blurred, the text is too small, or the AI cannot determine the customer’s intent, the behavior follows the agent’s AI Instruction. For Chatbase, OpenAI, or a self-hosted AI, the conversation is handed to a human agent when an image arrives, and subsequent messages are no longer forwarded to the AI Agent.

    If you use a self-hosted AI (for example, connected to Raccoon), your AI side stops “seeing” messages after the conversation is handed over. This is not a transmission failure, but CAAC stopping the forwarding per the rule.

    Q: What does the AI do when a customer sends an image unrelated to your service?

    A: The AI tries to understand the image. When it cannot determine the customer’s intent, the behavior follows the agent’s AI Instruction: if handover to a human is stated there, the AI hands the conversation to the corresponding human agent; if not, the AI asks the customer to describe what they need.

    Q: I’m not satisfied with the AI Agent’s replies. How can I adjust?

    A: AI Agent replies come from the connected knowledge base. If you are not satisfied, please update the knowledge base content to ensure AI can provide accurate and expected responses. If you use ChatBase as the knowledge base, refer to: Tutorial｜CAAC AI - ChatBase Knowledge Base # How to Optimize My Knowledge Base

    For agents of the Crescendo AI type, the knowledge base is configured inside CAAC. See: Tutorials｜CAAC AI - Crescendo AI.

    Q: The AI Agent’s handover timing doesn’t match my needs. What should I do?

    A: We recommend using the feedback and optimization mechanism to record AI handover actions and reasons. Then contact Crescendo Lab to obtain these records and adjust handover settings based on your needs.

    Q: Why does the AI Agent still reply to some messages after I’ve disabled “Auto-Assign to AI Agent”?

    A: Disabling “Auto-Assign to AI Agent” only stops new messages from being automatically assigned to AI.

    If conversations had already been assigned to the AI Agent before disabling, those will continue being handled by AI until the conversation ends or is manually reassigned to a team/agent.

    To immediately stop the AI Agent from replying:

    1. Locate all ongoing conversations still assigned to the AI Agent
    2. Manually reassign them to another team or agent

    This ensures the AI Agent will no longer reply to any messages.

    Related articles