# Understanding Kustomer Concierge > Source: https://kustomer.kustomer.help/en_us/understanding-kustomer-concierge-BkAOjnbKMg Last updated: 2026-09-11T17:54:57.119Z Kustomer Concierge, replaces the legacy AI for Customers 1.0, is the next generation of customer-facing AI automations. Concierge can respond to customers, search approved knowledge sources, follow your organization's procedures, use connected tools to complete actions, and transfer conversations to a human rep when needed. Concierge is a flexible model for building and managing AI-powered customer experiences. Instead of relying primarily on a fixed team of AI agents and their individual prompts, Concierge organizes behavior around reusable procedures, knowledge sources, tools, guidance, and—in advanced automations—visual routing steps. Concierge is designed for administrators and managers who want to automate repeatable customer service work while retaining control over what the AI can say, what it can do, and when it should hand off. ## In this article * [What Concierge can do](#what-concierge-can-do) * [How a Concierge automation works](#how-a-concierge-automation-works) * [Choosing between Simple and Advanced mode](#choosing-between-simple-and-advanced-mode) * [Understanding the parts of a Concierge automation](#understanding-the-parts-of-a-concierge-automation) * [Choosing between Simple and Advanced mode](#choosing-between-simple-and-advanced-mode) * [How Concierge differs from AI for Customers 1.0](#how-concierge-differs-from-ai-for-customers-10) * [Use cases](#use-cases) ## What Concierge can do A Concierge automation can handle a complete customer interaction rather than only generate a reply. Depending on how you configure it, Concierge can: * Answer questions using selected knowledge bases and data sources * Follow step-by-step procedures for common customer requests * Use tools to retrieve information or perform actions in Kustomer and connected systems * Maintain context as the customer continues the conversation * Recognize when the customer's intent changes and move the conversation to a more appropriate part of an advanced automation * Verify customer identity before using tools that require verification * Route or transfer a conversation to a human rep or queue * Apply response tone, custom guidance, and guardrails consistently * Start supported conversations with a configured greeting or respond when a customer contacts you For example, a retail Concierge automation could answer a return-policy question from the knowledge base, use an order tool to find the customer's purchase, follow a return procedure, and transfer the conversation to a human rep if the order is outside the standard policy. ## How a Concierge automation works The exact path depends on your configuration, but a typical Concierge interaction follows this pattern: 1. A customer message meets the automation's deployment conditions. 2. Concierge determines which agent or step should handle the request. 3. The agent uses the conversation, its procedures, selected knowledge sources, available tools, and your guidance to decide what to do. 4. If an action requires customer verification, Concierge completes the configured verification flow before allowing the tool to run. 5. Concierge evaluates the result of the agent's work and follows the configured success, failure, or handoff path. 6. After responding, Concierge preserves its place in the interaction. When the customer replies, it can continue from that point or re-evaluate the route if the customer's intent has changed. This combination of AI reasoning and controlled execution is central to Concierge. The AI can interpret natural customer language, while procedures, tools, verification, and routing rules constrain how it handles important work. ## Understanding the parts of a Concierge automation The following parts work together to define the customer experience. Part What it controls **Automation** The overall customer experience, including its name, deployment conditions, conversation settings, and one or more configured steps or agents. **AI agent** The part of the automation that interprets customer messages, chooses an appropriate procedure or tool, and responds. Simple automations use one Main Agent; advanced automations can use multiple specialized agents. **Procedures** Step-by-step instructions for specific customer needs, such as tracking an order, changing a reservation, or requesting a refund. Procedures help the agent perform repeatable work consistently. **Knowledge sources** The knowledge bases and data sources the agent can search for approved information. Knowledge provides facts; it does not replace instructions for what the agent should do with those facts. **Tools** Configured capabilities that let the agent retrieve data or take an action, such as looking up an order, updating a record, or routing a conversation. A tool must be available to the agent or procedure that needs it. **Response tone** The communication style and preferred answer length for customer-facing responses. **Custom guidance** Broader instructions for behavior, edge cases, and policies that apply across the agent's work. **Identity** A description of who the agent is, what it represents, and the purpose it serves. Availability may depend on your organization's configuration. **Guardrails** Controls for sensitive subjects, including how the agent discusses competitors and what information it must not reveal. **Conversation settings** Settings such as greeting messages, fallback messaging, human handoff, verification, business hours, and supported channel behavior. Available settings can vary by organization and channel. ### Procedures, knowledge, and tools have different jobs These three parts are closely related, but they are not interchangeable: * A **procedure** tells Concierge how to handle a type of request. * A **knowledge source** supplies information Concierge may use in its answer. * A **tool** lets Concierge retrieve live data or perform an action. For an order-status request, the procedure might tell Concierge to confirm which order the customer means, use an order lookup tool, explain the current status, and offer a human handoff if the lookup fails. A shipping article in the knowledge base could provide general carrier or delivery guidance, but it would not contain the customer's live order status. ## Choosing between Simple and Advanced mode Concierge supports two ways to organize an automation. ### Simple mode Simple mode uses one Main Agent. You add the knowledge sources, procedures, tools, tone, guidance, and guardrails that the agent needs, then configure what should happen when the agent succeeds or cannot complete the request. Simple mode is a good fit when: * One agent can cover the automation's related use cases * Procedures provide enough structure for the work * The AI can decide which procedure or tool is appropriate * You do not need a visual sequence of required routing or branching steps For many automations, Simple mode is the best place to start. ### Advanced mode Advanced mode provides a visual automation graph. In addition to AI agents, you can add deterministic blocks for tasks that must follow a specific route or outcome. Available building blocks include: * AI intent classification * Option menus * Exact or templated messages * Business-hours checks * Conversation routing * Existing agents * Configured tools Advanced mode is useful when you need guaranteed routing, fixed choices, required actions, explicit failure paths, or multiple specialist agents. For example, you might classify billing and technical-support requests before sending them to separate agents, then route any unmatched request to a human queue. > **Note:** Switching an automation from Simple mode to Advanced mode is a one-way change. Review the automation's requirements before switching. ## How Concierge differs from AI for Customers 1.0 AI for Customers 1.0 and Concierge both automate external customer conversations, but they use different configuration and execution models. Area AI for Customers 1.0 Concierge **Product generation** Legacy customer automation Current customer automation **Primary structure** A fixed team structure with a Supervisor that delegates work to child agents One Main Agent in Simple mode, or a visual graph with agents and deterministic blocks in Advanced mode **Instructions** Behavior is defined primarily through prompts and instructions on individual agents Behavior is separated into procedures, knowledge, tools, tone, custom guidance, identity, and guardrails **Procedures** No dedicated procedure object Reusable, agent-scoped procedures with ordered steps and tool references **Conversation execution** Each event runs through the legacy automation and completes that run; conversation history may still be available to the agents The automation pauses after an agent response, stores its place, and resumes when the customer replies **Changes in customer intent** The Supervisor and child-agent tool configuration determine delegation Advanced automations can classify intent and re-evaluate the route when the customer changes topics **Action validation** Tool success or failure is handled by the legacy execution path Concierge can evaluate whether the agent completed its goal and use the result to continue, fail, or hand off **Verification** Verification is handled outside the agent's own execution flow Verification can run inline before a protected tool is used **Guardrails** Does not use Concierge's built-in conversational guardrail model Supports default conversational protections and configurable competitor and secrets guidance **Building experience** Configure an AI Agent Team and its agents Build a Simple or Advanced automation, with Architect available to help create, analyze, and improve the configuration **Reporting** Measured in the legacy AI for Customers reporting experience Measured in Concierge reporting and observability experiences The practical difference is that Concierge gives each type of instruction a clearer home. A migration is therefore a translation, not a direct copy: a long legacy prompt may need to become several procedures, selected knowledge sources, tool assignments, tone settings, and custom guidance in Concierge. ### Use cases Concierge works best on repeatable customer requests that follow a predictable path and can be resolved from approved knowledge, a defined procedure, or a tool call. The use cases below are patterns already running in production Concierge automations. These use cases are rarely isolated. A single conversation often combines a knowledge lookup with one or more tool calls, and the same automation typically covers several of the patterns below through separate procedures. #### Answering policy and product questions from knowledge sources Concierge searches the knowledge bases and data sources you select and answers using only that approved content. This is the most common starting point for a new automation because it requires no connected systems. Common examples include: * Shipping, delivery, and carrier policy questions. * Return windows, warranty terms, and eligibility rules. * Account, billing, and product FAQs. * Store or service availability questions. Knowledge quality determines answer quality. Articles with descriptive titles, a single topic per article, and clean formatting retrieve far more reliably than long multi-topic articles or articles titled with internal codes. For guidance, see [writing articles for AI](https://help.kustomer.com/en_us/articles-best-practices-HJVnpENA). #### Resolving order and delivery inquiries Order status is the highest-volume Concierge use case. A procedure confirms which order the customer means, calls an order lookup tool, explains the current state, and offers a handoff when the lookup fails or the order falls outside standard policy. A single order procedure commonly covers: * Order status and production or fulfillment stage. * Tracking numbers, carrier, and expected delivery dates. * Payment status on an existing order. * Proof, mockup, or artwork approval for custom orders. On unauthenticated channels, configure verification on the order lookup tool and use custom guidance to describe the format of your order numbers. Without that context, the agent can confuse a verification code with an order number. #### Handling returns, exchanges, and refunds Concierge can explain return policy from knowledge, retrieve an existing return or RMA through a tool, and create or update return records in Kustomer or a connected system. Return procedures depend on knowledge more heavily than order procedures do, so confirm that every fact the procedure needs to state exists in a knowledge source. Return addresses and shipping label instructions are the most common gaps. When the information is missing, the agent has no source to draw from and may generate a plausible but incorrect answer instead of handing off. State explicitly in the procedure and in custom guidance that the agent must hand off rather than supply an address or label it cannot attribute to a source. #### Managing subscriptions and recurring orders Subscription automations pair a read tool with an update tool: the agent retrieves the subscription, then modifies it. This supports: * Changing delivery frequency, quantity, or next ship date. * Pausing or skipping a delivery. * Processing a cancellation and updating subscription status. * Offering a frequency change as an alternative before a cancellation. When a procedure chains two tools, tell the agent which identifier to pass between them. A step such as "use the ID returned by the subscription lookup" prevents the update tool from being called with the wrong record. #### Triaging inquiries and filtering non-support traffic Deployment conditions and Smart Routing decide which conversations Concierge handles at all. Two patterns are well established: * **Filtering out non-support traffic.** Define the automation to handle all inbound messages by default and exclude specific categories, such as out-of-office and auto-replies, marketing and vendor outreach, and spam or system noise. An exclusion-based definition performs better than listing every supported topic, which risks skipping legitimate requests whose phrasing does not match the list. * **Routing to specialized handling.** In Advanced mode, classify intent before handing the conversation to a specialist agent, and route unmatched requests to a human queue. Consider Advanced mode when one agent accumulates enough procedures that its responsibilities are difficult to separate. ### Use cases that need extra care Some requests are technically possible but carry more risk. Plan these deliberately: * **Irreversible actions.** Cancellations, refunds, and credits cannot be undone. Scope these tools to a single agent or procedure, keep them away from agents that also handle adjacent read-only requests, and test the confusing cases. A customer asking you to resend an order confirmation is not asking to cancel the order. * **Exact-string lookups.** Knowledge search is semantic, not literal, so codes such as error 4010 and error 5123 look similar to it. Retrieve error codes, SKUs, and reference numbers with a tool that queries your system of record rather than relying on knowledge search. * **Non-English conversations.** Concierge can hold conversations in languages other than English, but test each supported language end to end rather than testing only the greeting, and confirm that tone and custom guidance produce the response language you expect.