# Managing AI automations > Each AI automation has routing, guidance, and procedures that dictate how it will respond to a customer query. These settings can be managed at any time from yo Source: https://help.kustomer.com/en_us/managing-ai-automations-SkaCf6nFbg Last updated: 2026-10-19T19:12:10.022Z Each AI automation has routing, guidance, and procedures that dictate how it will respond to a customer query. These settings can be managed at any time from your AI automation's **Settings** page. ### **In this article** * [**Setting guidance**](#setting-guidance) * [**Writing AI automation guidance**](#writing-ai-automation-guidance) * [**Writing procedures**](#writing-procedures) * [**Writing custom guidance**](#writing-custom-guidance) * [**Writing guardrails**](#writing-guardrails) * [**Settings reference**](#settings-reference) * [**Deploying AI automations**](#deploying-ai-automations) ## **Setting guidance** Guidance shapes how your AI automation responds, how informed it sounds to your end user, what knowledge sources it can reference, and provides specific instruction on handling user disputes, edge cases, or questions about competitors. Good guidance is thorough and updated frequently as your AI automation performs. You will be able to thoroughly evaluate your AI automation's responses against tests you schedule, and you can analyze its interactions with end users. Options for guidance include: **Guidance** **Description** Knowledge Sources External-facing knowledge bases (Kustomer-based, other external URLs, or Guru knowledge sources via the Guru integration) your AI automation can reference and cite as it responds to users. Procedures Step-by-step processes written in natural language your AI automation will follow to respond to and diagnose issues. Response tone Define your AI automation's personality and tone when responding to the end user. Custom guidance Handle edge cases and special rules with conversational instruction. Guardrails Handle secret detection and define responses to questions about competitors or sensitive brands. ## **Writing AI automation guidance** Guidance can take the form of procedures, custom guidance, or guardrails. ### **Writing procedures** A procedure has steps and a time to use them. _When the customer asks about X, do A, then B, and then C._ Procedures must be written as a numbered list, and are designed to provide an AI automation with repeatable instructions that cannot be misunderstood. Therefore, the best instructions are: 1. Specific callouts to tools. Use the `@` symbol to bring up a list of tools your AI automation can access. 2. Structuring the response to present and cite information (for example as a numbered list or confirming with the user during each step, etc.). 3. How to handle success or failure. Escalate to a human agent or follow another specific set of procedures? For example, failing to resolve a shipping issue and switching to a procedure that troubleshoots order-related issues with the customer. ### **Writing custom guidance** Custom guidance works similarly to procedures, but is designed for specific edge cases. They do not need to be written as ordered lists and can be conversational in tone and explanation. Some examples include: * You are a customer service automation for Acme Co that handles refund requests that fall within 10 days of a refund policy. * You are a booking agent that can offer customers incentives when flights are overbooked. * You are a hospitality agent that handles VIP requests from specific customers. Using tools and step-by-step instruction, custom guidance allows your AI automation to handle these edge cases and reduce the number of human agent handoffs that might normally occur. ### **Writing guardrails** Guardrails have two functions: 1. Provide guidance when discussing a competitor. 2. Provide instruction for handling customer secret data. When a brand or product is mentioned, the AI automation can detect this mention and take direction on how to discuss this competitor. Customer service interactions often involve sensitive data used to identify users. Define the kinds of data the AI automation should consider secret, such as a user's ID number or birth date. These data values will be obfuscated or redacted during AI automation interactions, allowing users to use them safely without fear of exposing unnecessary personal details. ### **Settings reference** The **Settings** page for each AI automation includes these options. After changing any setting, click **Save Changes**. #### **General** **Setting** **What it changes** Display Name (required) The friendly name shown to end users. Description (required) The purpose and key tasks of the AI automation, for easy reference. Model The AI model the AI automation uses. See [Choosing the Concierge model](#choosing-the-concierge-model). #### **Guardrails** Guardrails are configured in two fields. See [Writing guardrails](#writing-guardrails) for how guardrails work. **Setting** **What it changes** Competitors The brands, products, or competitors you want the guardrails to detect and consider. When one is mentioned, the guardrails system rewords the message. Secrets Exceptions or additional guidance for secrets detection. Default rules always apply for passwords, API keys, tokens, and credentials. Detected secrets are obfuscated or redacted. For example, you can tell the guardrails not to flag promo codes or gift cards that start with a specific prefix. #### **Response tone** **Setting** **What it changes** Tone of voice The personality, communication style, and level of formality of the AI automation. Choose **Friendly**, **Professional**, **Casual**, **Matter of Fact**, or **Custom**. Custom tone text When **Custom** is selected, the written description of the tone your AI automation should use. Limited to 5,000 characters. #### **Email** **Setting** **What it changes** Template The email template the AI automation uses when sending emails to customers. Include a disclaimer in the template to let customers know they are not talking to a person. Reply to all recipients on email Whether the AI automation replies to everyone on the inbound email or only the sender. See [Reply to all recipients on email](#reply-to-all-recipients-on-email). #### **Reply to all recipients on email** This setting controls who receives the AI automation's email replies. **Setting state** **Behavior** On The AI automation CCs everyone on the inbound email, except your support mailboxes. Off The AI automation replies to the sender only. #### **Handoff** **Setting** **What it changes** Enable human-to-AI handoff Whether human users can hand off a conversation to the AI automation. #### **Memory** **Setting** **What it changes** Enable Personalized Customer Memory Whether the AI automation keeps memory of customer preferences and important aspects of historical interactions, to create a more personalized experience. #### **Conversation start** These settings control how conversations begin. You can start the conversation with a proactive greeting, or simply respond to incoming messages from customers. **Setting** **What it changes** Chat greeting message The message sent at the start of a chat conversation. Use it to provide any disclaimers or disclosures before the customer interacts with AI. SMS/WhatsApp greeting message The message sent when a customer starts a conversation via SMS or WhatsApp. #### **Verification** Select the outbound account for each channel that the AI automation uses to authenticate and verify a customer. **Setting** **What it changes** Email Address The email address the AI automation uses for verification. SMS The phone number the AI automation uses for verification. #### **Conversation abandonment and fallback** **Setting** **What it changes** Enable conversation abandonment handling Whether the AI automation handles conversations where the customer stops responding. Email abandonment threshold The number of days without a customer response before an email conversation is considered abandoned. Live conversation abandonment threshold The number of minutes without a customer response before a live conversation is considered abandoned. Live channels include voice, chat, SMS, and WhatsApp. Abandonment message The message sent when the customer doesn't respond within the set wait time. AI fallback behavior If the AI automation fails to respond due to an error or timeout, the conversation is reassigned using queue rules. AI failure message The message sent when the AI automation can't respond and the conversation is transferred. #### **Routing** **Setting** **What it changes** Skip Sending Customer Message after Routing The channels where no message is sent when a conversation is routed to a human agent. By default, messages are sent on all channels. Business hours Whether and how conversations are routed or unassigned outside business hours. See [Business hours](#business-hours). #### **Business hours** Turn on business hours awareness to control how the AI automation handles conversations outside your support hours. **Setting** **What it changes** Turn on business hours awareness Whether the AI automation follows a business schedule. Business schedule (required) The business schedule the AI automation follows. Outside business hours message (required) The message sent when a conversation is routed or unassigned after business hours. Routing behavior (required) Whether to route the conversation to a specific queue or unassign it and let your routing rules take over. Queue (required) When routing to a queue, the queue that receives conversations outside business hours. #### Choosing the Concierge model The _Model_ dropdown controls which AI model your AI automation uses to respond to customers. When Kustomer releases a new model, you can test it on a draft of your AI automation before it reaches customers. Your choice applies to the draft and goes live only when you deploy. Option Behavior Current model The model the AI automation uses today. Labeled _Deprecating Oct 26_. This model will be removed on October 26, 2026, and AI automations still using it will be upgraded automatically. Next-generation model Faster and more capable, with more natural conversations. Labeled _Recommended_. Latest model Shown when the AI automation already uses the latest model. No other option is available. To choose the Concierge model: 1. In the left navigation panel, click the **Kustomer AI** (sparkle) icon, then select your AI automation. 2. Click **Settings**; the Settings page opens. 3. In the General section, select **Current model** or **Next-generation model** from the _Model_ dropdown. 4. Click **Save Changes**. 5. Test the AI automation to confirm its responses meet your expectations. 6. Click **Deploy** to put the selected model live for customers. See [Deploying AI automations](#deploying-ai-automations). ## **Deploying AI automations** When you are ready to take your AI automation live and have it respond to customers, click **Deploy**. To deploy AI automations from Kustomer: 1. Define smart routing conditions. See Defining smart routing for more information. 1. Smart routing rules are brief, for example: 1. Product returns 2. Shipping issues 3. Changes to bookings, etc. 2. When these rules match, your AI automation will handle the conversation. 3. Routing will send conversations that do not match this criteria to a human agent. 4. Smart routing can cause the AI automation to ask up to three follow up questions to clarify a customer's intent before routing will occur. 2. Click **Deploy** to deploy your AI automation, or **Discard Changes**.