# Understand observability in automations > Observability refers to the ability to monitor track and understand the performance of your AI Agents whether theyre part of an AI Agent Team or used with Source: https://help.kustomer.com/en_us/understand-observability-in-automations-B1CMeXm1ye Last updated: 2026-07-17T18:15:05.535Z Observability helps you review how the Copilot AI Agent helps your human reps. Use observability to investigate AI Agent behavior, understand which tools or knowledge sources were used, and identify areas to improve before or after deployment. **Note:** Observability only tracks traces from Copilot, not from Concierge.  The main unit of observability is a **trace**. A trace shows the steps an AI Agent takes while processing a message or input. This can include tool usage, knowledge retrieval, guardrail checks, agent responses, and other operations that contributed to the final result. * For **Concierge**, traces reflect how the agent processes an inbound message from a customer.  * For **Copilot**, traces show how the agent processes an input from an agent. See [Understand observability in automations](https://help.kustomer.com/en_us/understand-observability-in-automations-B1CMeXm1ye). By analyzing traces, you can gain insight into how effectively an AI Agent handles interactions, identifies issues, and makes improvements to ensure optimal performance. **Who can access this feature?** **User types** Admins can access the Customers page. ### **In this article** * [What are traces?](#what) * [View traces for an AI Agent](#view) * [How to read a trace](#read) * [Understand guardrails](#guardrails) ### **What are traces?** Traces follow the lifecycle of a request as it moves through an AI Agent. They help you understand what happened between the original message or input and the response generated by the AI Agent. Depending on the AI Agent setup and the interaction being reviewed, a trace may show: * The original message or input * The AI Agent, supervisor, or specialist that handled the request * Knowledge sources used to support the response * Tools used during the interaction * Inputs and outputs for tool calls * Guardrail checks * Reasoning or decision points that help explain the AI Agent’s behavior * The final response generated by the AI Agent Trace details vary based on your AI Agent configuration, the channel, the tools available to the AI Agent, and the type of interaction being reviewed. ### **View traces for automations** You can view traces from AI Agent observability pages or from supported conversations. **Note:** Envoy with Copilot doesn’t support trace versioning. You’ll always see the latest version of the trace. **For Envoy:** * From the left navigation, click **Kustomer AI**\>**Envoy**. * Open the AI Agent or automation you want to review. * Click **Observe Copilot**. Depending on your workspace and AI Agent version, traces may appear in an **Observe**, **Optimize**, **Analyze**, or **View Traces** experience. The exact navigation can vary, so confirm the labels in your workspace during UI review. **From directly within a conversation:** 1. You can also open traces directly from a conversation that was handled by a human rep using Copilot. 1. Open a conversation augmented by Copilot. 2. Open the conversation options menu. 3. Select **Observe Traces**. 4. If traces are available for the conversation, Kustomer opens the trace view for that interaction. ### **How to read a trace** **Note:** The results you see in your traces may vary based on how you set up your AI Agent. A trace shows the operations that occurred while the AI Agent generated a response. Select a trace item to review more detail about that step. A trace may include the following types of activity: ![](https://cdn.kustomerhostedcontent.com/media/570fad9d9001bc1000163b28/2ff4badf9e36df5d09b89d5b44d991b3.png) **Note:** When a newer inbound message interrupts an assistant run, the earlier run appears as **Cancelled** in the trace. The trace explains that the assistant stopped the earlier response and started again using the newest message. ### Message or input The message or input that started the trace. For customer-facing AI Agents, this is usually a customer message. For Copilot, this is usually an agent input. ### Knowledge retrieval If the AI Agent used knowledge to answer the inquiry, the trace may show the knowledge source, article, or extracted content used to support the response. ### Tool usage If the AI Agent used a tool, the trace may show which tool was used and details about the tool input or output. Tools can help the AI Agent retrieve, update, or act on data while handling an interaction. ### Reasoning or decision points ![](https://cdn.kustomerhostedcontent.com/media/570fad9d9001bc1000163b28/fc1cfa282b6c892bedec656ceb53d9a7.png) Reasoning or decision points help explain why the AI Agent took a specific next step while generating a response. These trace details can show whether the AI Agent decided to respond directly, use a tool, or hand off work to another AI Agent teammate. For example, a reasoning or decision point may show: * The reasoning type, such as Response, Tool Use, or AI Agent Teammate * The AI tool or AI Agent teammate selected by the AI Agent * The arguments or message sent to the selected tool or AI Agent teammate * The response returned by the tool or AI Agent teammate * The final AI response generated after the decision point * Available tools or AI Agent teammates that the AI Agent could choose from * Model thoughts, validation thinking, or active procedures, when available for the trace Use these details to understand how the AI Agent interpreted the input and why it followed a particular path. ## Understand guardrails in traces Guardrails help control what AI Agents can say or do during an interaction. In traces, guardrail-related steps can help you understand whether a message or response triggered configured guidance or safety checks. ![](https://cdn.kustomerhostedcontent.com/media/570fad9d9001bc1000163b28/bf3a67601e858a3b53bb387ce8ce1818.png) ### Understand guardrails By setting guardrails, you can prevent your automations from mentioning or discussing competitors in their interactions. In addition to protecting against mentions of competitors, these guardrails work alongside standard safeguards within the product to ensure automations don’t disclose sensitive information. You can see if any guardrails were detected in the trace for each message. ![](https://cdn.kustomerhostedcontent.com/media/570fad9d9001bc1000163b28/ee19de6c637840ff448a98492513c158.png) Guardrail trace details may include: Guardrail Description Competitor Detects explicit mentions of competitors configured in the AI Agent’s competitor guidance. Toxicity Detects abusive, threatening, insulting, or otherwise toxic content. NSFW Detects content that may be inappropriate for work or unsafe for customer support interactions. Secrets Detects credentials or secret values, such as passwords, private keys, or tokens. Guardrails work alongside standard safeguards in Kustomer AI. If a guardrail is detected, review the trace details to understand what was detected and how the AI Agent handled the interaction.