{"content_id":"2doywe7tfr","slug":"why-ai-chatbots-do-not-reduce-call-center-calls","locale":"en","schema_type":"TechArticle","category":"ai_data","category_name":"AI Data","title":"Why Call Center Calls Do Not Decline Despite AI Chatbots","summary":"Even as chatbot use grows, call center calls and operating costs may not decline if chatbots lack the authority to resolve issues, contextual understanding, or effective escalation to human agents. Rather than measuring automated resolution rates, organizations should assess whether customers' problems are solved and design the entire journey so AI and agents share context.","sponsorship_disclosure":null,"author":{"name":"Injoys Editorial Team","url":"https://injoys.com/ko/about"},"key_points":["Increased chatbot use may indicate that inquiry demand and customer touchpoints have grown, so it does not directly prove that call center calls have declined.","The ability to explain information and the authority to actually process refunds, cancellations, or account changes are different capabilities.","Calls made after chatbot failures tend to involve complex cases, so even if call volume remains unchanged, average handling time and emotional labor may increase.","An effective handoff provides the agent with a conversation summary, the customer's authentication status, actions already taken, and the reasons for failure.","Key performance indicators should focus on journey completion rates, repeat inquiry rates, total resolution time, and customer effort rather than simple deflection rates."],"content_markdown":"A large increase in AI chatbot conversation volume does not automatically reduce call center calls or support costs. Answering a question and resolving a customer’s problem through to completion are different tasks, and failed automation can instead push more complex and emotionally charged inquiries to human agents.\n\nThe key paradox observed in operations is this: chatbot usage rises while total call volume remains nearly unchanged, and the difficulty of inquiries handled by agents and the average handling time increase. To understand this, organizations must examine the customer’s entire resolution journey rather than usage by channel.\n\n## Why Chatbot Usage and Call Reduction Do Not Align\n\nChatbot usage is closer to a measure of touchpoint activity than an outcome metric. Conversation volume can increase simply because of more visitors, broader chatbot exposure, a change in its entry point within an app, or existing FAQ users switching channels.\n\nFor example, even if chatbot conversations increase by 300%, call volume will not decrease in the following situations.\n\n- Orders and subscriber numbers both increased, raising overall inquiry demand.\n- Existing search and FAQ users moved to the chatbot, while phone users remained unchanged.\n- A customer used the chatbot and then called again about the same issue.\n- The chatbot became a new entry point for inquiries, prompting even customers who would previously have given up to seek support.\n- The proportion of complex issues requiring a phone call did not decrease.\n\nTherefore, analysis should separate the following metrics alongside the simple number of conversations.\n\n| Metric | Calculation or meaning | Points to note |\n|---|---|---|\n| Phone contact rate | Number of phone inquiries ÷ number of orders, subscribers, or active customers | Must adjust for increased inquiry volume caused by business growth. |\n| Handover rate | Chatbot sessions transferred to a human agent ÷ total chatbot sessions | A low rate is not necessarily good. It can also be lowered by hiding the connect button. |\n| Repeat contact rate | Percentage of customers who make contact again about the same issue within a given period | Must identify whether it is the same issue even across different channels. |\n| Journey completion rate | Customers who actually completed the target task ÷ customers who attempted that task | Must distinguish providing an answer from actually completing the task. |\n| First-contact resolution rate | Percentage resolved during the first interaction without additional contact | Distorted if agents arbitrarily mark cases complete without customer confirmation. |\n| Total resolution time | Time elapsed from first contact to final resolution | Must include waiting, channel switching, and reauthentication time. |\n| Average handling time | Total time spent on support work ÷ number of cases handled | May rise if only difficult cases remain after automation. |\n\n## Three Structural Limitations That Prevent AI Answers From Resolving Problems\n\n### 1. The Gap Between the Ability to Explain and the Authority to Act\n\nGenerative AI can quickly explain information contained in documents, such as policies, instructions, and product information. However, a customer’s actual request often involves completing a multistep business process.\n\nFor example, if a customer says, “I canceled my subscription, but I was charged again, so please issue a refund,” the system may need to perform the following tasks.\n\n1. Authenticate the customer and account.\n2. Retrieve the cancellation date and payment history.\n3. Determine whether it was a duplicate charge or whether the refund conditions are met.\n4. Check approval authority and exception policies.\n5. Issue the refund and record the outcome.\n\nIf the chatbot is not connected to internal payment and order systems or lacks execution authority, it can only provide instructions. If the customer must find the relevant menu and complete the process again, the company may regard the “answer” as complete, but the customer sees the issue as unresolved.\n\nUnconditionally granting broad authority for sensitive tasks is not a solution either. Functions such as refunds, personal information changes, and account recovery require identity verification, least-privilege access, monetary limits, approval procedures, audit logs, and recovery mechanisms for failures.\n\n### 2. Lack of Context Not Revealed in the Question\n\nIt is difficult to determine an appropriate answer from the sentence, “Please recommend a resort destination for a trip with my child,” alone. Constraints such as the child’s age, travel time, budget, food allergies, pool safety, and access to medical care can change the outcome.\n\nWays to supplement context include the following elements.\n\n- Follow-up questions that confirm required conditions\n- Use of order and support history within the scope authorized by the customer\n- Knowledge graphs representing relationships among products, policies, regions, audiences, and exceptions\n- Ontologies that consistently define terms and relationships\n- Retrieval from current policy documents and real-time inventory and reservation systems\n\nOntologies and knowledge graphs are possible implementation methods, not mandatory requirements for every chatbot. For simple tasks, structured APIs and clear conversational flows may be more efficient. What matters is preventing the model from filling gaps through guesswork and instead having it ask for the necessary information or retrieve it from a reliable system.\n\n### 3. High-Difficulty Inquiries Left After Automation Fails\n\nWhen a chatbot handles easy inquiries, agents are left with a relatively larger share of complex exception cases. This can be viewed as a change in case mix. Even if the number of calls stays the same or decreases slightly, total support time and costs may not decline if the remaining calls become longer.\n\nThe emotional burden also increases if customers must repeat the same explanation to an agent from the beginning after already giving it to the chatbot. What began as a simple payment confirmation may escalate into a complaint that also includes frustration over the automation failure.\n\nHowever, an increase in average handling time should not automatically be blamed on the chatbot. Factors such as product outages, policy changes, the proportion of new agents, and seasonality also have an impact. When comparing performance before and after implementation, organizations should control for inquiry type and customer segment and separately analyze customers who called after using the chatbot and those who called directly.\n\n## The Solution Is Seamless Handover, Not Inquiry Deflection\n\nHandover is the process of transferring a conversation that AI could not resolve to a human agent. The purpose of a good handover is not to keep customers inside the chatbot for as long as possible, but to quickly detect the limits of automation and allow the next party responsible for resolution to continue the work without interruption.\n\nA connection to a human agent can be offered when signals such as the following appear.\n\n- The customer explicitly asks to speak with an agent.\n- The same or similar question is repeated.\n- Answer confidence is below the threshold, or no supporting document was found.\n- The situation is high risk, such as a payment dispute, account takeover, legal threat, or safety issue.\n- Negative emotions persist, or the customer rejects the answer.\n- An action performed by the chatbot fails or enters an exception state.\n\nThe agent’s screen should receive structured information that can be used immediately for the task, rather than just a full conversation transcript.\n\n| Information transferred | Specific details |\n|---|---|\n| Customer intent | The outcome the customer ultimately wants |\n| Key facts | Confirmed information such as order number, time of occurrence, product, and amount |\n| Authentication status | How identity verification was completed and the scope for which it remains valid |\n| Actions performed | Actions the chatbot retrieved or executed and their results |\n| Cause of failure | Insufficient authority, policy exception, API error, low confidence, and so on |\n| Conversation summary | Customer’s claims, explanations already provided, and remaining questions |\n| Emotion and risk signals | Intensity of dissatisfaction and whether security, safety, or legal risks are present |\n| Evidence | Version of the policy document used and related system records |\n\nBecause AI-generated summaries may contain errors, the original conversation must also be available for review. Authentication information and sensitive personal information should be transferred only to the extent necessary, with access permissions and retention periods managed appropriately.\n\n## A Three-Tier Operating Model for Dividing the Roles of AI and Humans\n\nIt is safer to divide the level of automation based on task risk, frequency of exceptions, and responsibility for judgment.\n\n| Area | Suitable tasks | Role of AI | Role of humans |\n|---|---|---|---|\n| Automation | Business hours information, shipment tracking, low-risk reservation changes | Retrieval, explanation, and task execution | Intervene when exceptions occur |\n| Collaboration | Complex product inquiries, review of policy exceptions, general complaints | Organize context, search for evidence, and suggest draft responses | Verify facts, make final decisions, and communicate |\n| Human-only | Legal disputes, high-value refunds, account takeovers, safety crises, serious emotional care | Search records and provide supporting materials | Exercise accountable judgment, approve actions, and restore the relationship |\n\nThe classification criterion should not be whether “AI writes well,” but the potential harm of incorrect handling and the degree to which it can be reversed. It is appropriate to retain human review and approval for tasks involving significant financial, legal, or safety risks or frequent exceptions.\n\n## Shifting KPIs From Deflection Rate to Journey Completion Rate\n\nThe deflection rate or automation rate is often used to mean the percentage of sessions not transferred to an agent. However, the metric can improve even when the path to human support is made difficult or customers give up on resolving their issue. Using it as a standalone KPI therefore risks encouraging optimization in the opposite direction from the customer experience.\n\nA journey-centered measurement framework should answer the following questions.\n\n1. Was the task the customer wanted actually completed?\n2. Did the customer make contact again about the same issue?\n3. How many channels and steps were required to reach resolution?\n4. How many times did the customer have to repeat the same information or authentication?\n5. How long did it take from first contact to final resolution?\n6. Did an automation error lead to financial loss or a policy violation?\n\nA recommended dashboard can be organized into four layers: outcomes, effort, operations, and risk.\n\n- **Outcomes:** Journey completion rate, first-contact resolution rate, repeat contact rate\n- **Customer effort:** Total resolution time, number of repeated explanations, number of channel switches, customer effort score\n- **Operations:** Phone contact rate, average handling time, wait time, agent occupancy rate\n- **Quality and risk:** Incorrect answer rate, rate of actions performed without approval, sensitive information exposure, number of disputes and recoveries\n\nCosts should not be calculated from call volume alone. Organizations should compare total cost per journey, including AI inference costs, system integration, quality evaluation, security controls, human review time, and error recovery costs.\n\n## Context Inference, Orchestration, and Handover Pipeline\n\nA practical customer service AI is less like a single chat window and more like a pipeline connecting multiple systems.\n\n### 1. Context Inference\n\nIdentify the customer’s intent and necessary conditions, and retrieve authorized customer information and current operational data. The system should not rely solely on the model’s memory, but should use authoritative sources such as policy documents, order systems, and account status.\n\n### 2. Orchestration\n\nDetermine which tool or person should receive the request. Before execution, verify authentication status, permissions, monetary limits, and risk level. If the action fails, do not retry indefinitely; switch to a recovery procedure or human support.\n\n### 3. Handover\n\nSummarize the customer’s goal, confirmed facts, actions performed, cause of failure, and recommended next steps, and transfer them to the agent. The agent reviews the original conversation and supporting evidence before continuing the process.\n\n### 4. Outcome Recording and Learning\n\nStore whether the issue was ultimately resolved and any corrections made by the agent. For recurring failures, do not merely retrain the model; determine whether the cause lies in policy documents, APIs, operational permissions, or screen flows, and improve the relevant area.\n\n## Experimental Design for Verifying Implementation Effects\n\nComparing only total call volume before and after implementation makes it difficult to separate the effects of business growth and seasonality. Where possible, organizations should conduct a phased rollout across comparable customer segments or inquiry types and observe the following items together.\n\n- Number of calls per 1,000 orders or active customers\n- Repeat contact for the same reason within 24 hours or 7 days after chatbot use\n- Completion rate and average handling time by inquiry type\n- Customer effort and satisfaction before and after connection to a human\n- Percentage of AI summaries corrected by agents\n- Error rate of automated actions and time required for human recovery\n- Agents’ cognitive load, emotional exhaustion, and job satisfaction\n\nThe definition of “resolution” must also be established in advance. An answer being displayed, the customer ending the conversation, a backend task succeeding, and the customer confirming the outcome are all different states. For tasks requiring follow-up, such as refunds, the actual transaction status must be verified.\n\n## Checklist for Product Teams\n\n- Have you distinguished between tasks the chatbot can answer and tasks it can actually execute?\n- Do you verify each task’s success criteria through backend outcomes?\n- Can customers clearly request human support at any time?\n- Do you use repeated questions, low confidence, and risky expressions as transfer signals?\n- Can agents view the original conversation and supporting evidence, not just a conversation summary?\n- Can customers avoid repeating authentication and explanations?\n- Are there limits and approval procedures for the amounts and permissions available to AI?\n- Do you measure repeat contact rates and total resolution time by inquiry type?\n- Does agent feedback lead to improvements in knowledge documents, tools, and business processes?\n- Do actual operations staff regularly test the entire chatbot journey as customers?\n\n## Conclusion\n\nThe success of an AI chatbot cannot be judged solely by conversation volume or the human-support deflection rate. If the customer’s objective is not actually completed, the chatbot becomes an additional step that must be passed through before calling, rather than a problem-resolution channel.\n\nEffective design does not hide the limitations of AI. It automates standardized, low-risk tasks, supports agents in work requiring judgment, and rapidly transfers high-risk situations to skilled humans. When this is combined with handover that transfers both the conversation and task status, AI can become a tool that reduces customer resolution time rather than a barrier that blocks access to the call center.","content_html":"\u003cp\u003eA large increase in AI chatbot conversation volume does not automatically reduce call center calls or support costs. Answering a question and resolving a customer’s problem through to completion are different tasks, and failed automation can instead push more complex and emotionally charged inquiries to human agents.\u003c/p\u003e\n\u003cp\u003eThe key paradox observed in operations is this: chatbot usage rises while total call volume remains nearly unchanged, and the difficulty of inquiries handled by agents and the average handling time increase. To understand this, organizations must examine the customer’s entire resolution journey rather than usage by channel.\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#why-chatbot-usage-and-call-reduction-do-not-align\" class=\"anchor\" id=\"why-chatbot-usage-and-call-reduction-do-not-align\"\u003e\u003c/a\u003eWhy Chatbot Usage and Call Reduction Do Not Align\u003c/h2\u003e\n\u003cp\u003eChatbot usage is closer to a measure of touchpoint activity than an outcome metric. Conversation volume can increase simply because of more visitors, broader chatbot exposure, a change in its entry point within an app, or existing FAQ users switching channels.\u003c/p\u003e\n\u003cp\u003eFor example, even if chatbot conversations increase by 300%, call volume will not decrease in the following situations.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eOrders and subscriber numbers both increased, raising overall inquiry demand.\u003c/li\u003e\n\u003cli\u003eExisting search and FAQ users moved to the chatbot, while phone users remained unchanged.\u003c/li\u003e\n\u003cli\u003eA customer used the chatbot and then called again about the same issue.\u003c/li\u003e\n\u003cli\u003eThe chatbot became a new entry point for inquiries, prompting even customers who would previously have given up to seek support.\u003c/li\u003e\n\u003cli\u003eThe proportion of complex issues requiring a phone call did not decrease.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eTherefore, analysis should separate the following metrics alongside the simple number of conversations.\u003c/p\u003e\n\u003cdiv class=\"overflow-x-auto\"\u003e\u003ctable\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth\u003eMetric\u003c/th\u003e\n\u003cth\u003eCalculation or meaning\u003c/th\u003e\n\u003cth\u003ePoints to note\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Metric\"\u003ePhone contact rate\u003c/td\u003e\n\u003ctd data-label=\"Calculation or meaning\"\u003eNumber of phone inquiries ÷ number of orders, subscribers, or active customers\u003c/td\u003e\n\u003ctd data-label=\"Points to note\"\u003eMust adjust for increased inquiry volume caused by business growth.\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Metric\"\u003eHandover rate\u003c/td\u003e\n\u003ctd data-label=\"Calculation or meaning\"\u003eChatbot sessions transferred to a human agent ÷ total chatbot sessions\u003c/td\u003e\n\u003ctd data-label=\"Points to note\"\u003eA low rate is not necessarily good. It can also be lowered by hiding the connect button.\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Metric\"\u003eRepeat contact rate\u003c/td\u003e\n\u003ctd data-label=\"Calculation or meaning\"\u003ePercentage of customers who make contact again about the same issue within a given period\u003c/td\u003e\n\u003ctd data-label=\"Points to note\"\u003eMust identify whether it is the same issue even across different channels.\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Metric\"\u003eJourney completion rate\u003c/td\u003e\n\u003ctd data-label=\"Calculation or meaning\"\u003eCustomers who actually completed the target task ÷ customers who attempted that task\u003c/td\u003e\n\u003ctd data-label=\"Points to note\"\u003eMust distinguish providing an answer from actually completing the task.\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Metric\"\u003eFirst-contact resolution rate\u003c/td\u003e\n\u003ctd data-label=\"Calculation or meaning\"\u003ePercentage resolved during the first interaction without additional contact\u003c/td\u003e\n\u003ctd data-label=\"Points to note\"\u003eDistorted if agents arbitrarily mark cases complete without customer confirmation.\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Metric\"\u003eTotal resolution time\u003c/td\u003e\n\u003ctd data-label=\"Calculation or meaning\"\u003eTime elapsed from first contact to final resolution\u003c/td\u003e\n\u003ctd data-label=\"Points to note\"\u003eMust include waiting, channel switching, and reauthentication time.\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Metric\"\u003eAverage handling time\u003c/td\u003e\n\u003ctd data-label=\"Calculation or meaning\"\u003eTotal time spent on support work ÷ number of cases handled\u003c/td\u003e\n\u003ctd data-label=\"Points to note\"\u003eMay rise if only difficult cases remain after automation.\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\u003c/div\u003e\n\u003ch2\u003e\n\u003ca href=\"#three-structural-limitations-that-prevent-ai-answers-from-resolving-problems\" class=\"anchor\" id=\"three-structural-limitations-that-prevent-ai-answers-from-resolving-problems\"\u003e\u003c/a\u003eThree Structural Limitations That Prevent AI Answers From Resolving Problems\u003c/h2\u003e\n\u003ch3\u003e\n\u003ca href=\"#1-the-gap-between-the-ability-to-explain-and-the-authority-to-act\" class=\"anchor\" id=\"1-the-gap-between-the-ability-to-explain-and-the-authority-to-act\"\u003e\u003c/a\u003e1. The Gap Between the Ability to Explain and the Authority to Act\u003c/h3\u003e\n\u003cp\u003eGenerative AI can quickly explain information contained in documents, such as policies, instructions, and product information. However, a customer’s actual request often involves completing a multistep business process.\u003c/p\u003e\n\u003cp\u003eFor example, if a customer says, “I canceled my subscription, but I was charged again, so please issue a refund,” the system may need to perform the following tasks.\u003c/p\u003e\n\u003col\u003e\n\u003cli\u003eAuthenticate the customer and account.\u003c/li\u003e\n\u003cli\u003eRetrieve the cancellation date and payment history.\u003c/li\u003e\n\u003cli\u003eDetermine whether it was a duplicate charge or whether the refund conditions are met.\u003c/li\u003e\n\u003cli\u003eCheck approval authority and exception policies.\u003c/li\u003e\n\u003cli\u003eIssue the refund and record the outcome.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eIf the chatbot is not connected to internal payment and order systems or lacks execution authority, it can only provide instructions. If the customer must find the relevant menu and complete the process again, the company may regard the “answer” as complete, but the customer sees the issue as unresolved.\u003c/p\u003e\n\u003cp\u003eUnconditionally granting broad authority for sensitive tasks is not a solution either. Functions such as refunds, personal information changes, and account recovery require identity verification, least-privilege access, monetary limits, approval procedures, audit logs, and recovery mechanisms for failures.\u003c/p\u003e\n\u003ch3\u003e\n\u003ca href=\"#2-lack-of-context-not-revealed-in-the-question\" class=\"anchor\" id=\"2-lack-of-context-not-revealed-in-the-question\"\u003e\u003c/a\u003e2. Lack of Context Not Revealed in the Question\u003c/h3\u003e\n\u003cp\u003eIt is difficult to determine an appropriate answer from the sentence, “Please recommend a resort destination for a trip with my child,” alone. Constraints such as the child’s age, travel time, budget, food allergies, pool safety, and access to medical care can change the outcome.\u003c/p\u003e\n\u003cp\u003eWays to supplement context include the following elements.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eFollow-up questions that confirm required conditions\u003c/li\u003e\n\u003cli\u003eUse of order and support history within the scope authorized by the customer\u003c/li\u003e\n\u003cli\u003eKnowledge graphs representing relationships among products, policies, regions, audiences, and exceptions\u003c/li\u003e\n\u003cli\u003eOntologies that consistently define terms and relationships\u003c/li\u003e\n\u003cli\u003eRetrieval from current policy documents and real-time inventory and reservation systems\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eOntologies and knowledge graphs are possible implementation methods, not mandatory requirements for every chatbot. For simple tasks, structured APIs and clear conversational flows may be more efficient. What matters is preventing the model from filling gaps through guesswork and instead having it ask for the necessary information or retrieve it from a reliable system.\u003c/p\u003e\n\u003ch3\u003e\n\u003ca href=\"#3-high-difficulty-inquiries-left-after-automation-fails\" class=\"anchor\" id=\"3-high-difficulty-inquiries-left-after-automation-fails\"\u003e\u003c/a\u003e3. High-Difficulty Inquiries Left After Automation Fails\u003c/h3\u003e\n\u003cp\u003eWhen a chatbot handles easy inquiries, agents are left with a relatively larger share of complex exception cases. This can be viewed as a change in case mix. Even if the number of calls stays the same or decreases slightly, total support time and costs may not decline if the remaining calls become longer.\u003c/p\u003e\n\u003cp\u003eThe emotional burden also increases if customers must repeat the same explanation to an agent from the beginning after already giving it to the chatbot. What began as a simple payment confirmation may escalate into a complaint that also includes frustration over the automation failure.\u003c/p\u003e\n\u003cp\u003eHowever, an increase in average handling time should not automatically be blamed on the chatbot. Factors such as product outages, policy changes, the proportion of new agents, and seasonality also have an impact. When comparing performance before and after implementation, organizations should control for inquiry type and customer segment and separately analyze customers who called after using the chatbot and those who called directly.\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#the-solution-is-seamless-handover-not-inquiry-deflection\" class=\"anchor\" id=\"the-solution-is-seamless-handover-not-inquiry-deflection\"\u003e\u003c/a\u003eThe Solution Is Seamless Handover, Not Inquiry Deflection\u003c/h2\u003e\n\u003cp\u003eHandover is the process of transferring a conversation that AI could not resolve to a human agent. The purpose of a good handover is not to keep customers inside the chatbot for as long as possible, but to quickly detect the limits of automation and allow the next party responsible for resolution to continue the work without interruption.\u003c/p\u003e\n\u003cp\u003eA connection to a human agent can be offered when signals such as the following appear.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eThe customer explicitly asks to speak with an agent.\u003c/li\u003e\n\u003cli\u003eThe same or similar question is repeated.\u003c/li\u003e\n\u003cli\u003eAnswer confidence is below the threshold, or no supporting document was found.\u003c/li\u003e\n\u003cli\u003eThe situation is high risk, such as a payment dispute, account takeover, legal threat, or safety issue.\u003c/li\u003e\n\u003cli\u003eNegative emotions persist, or the customer rejects the answer.\u003c/li\u003e\n\u003cli\u003eAn action performed by the chatbot fails or enters an exception state.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe agent’s screen should receive structured information that can be used immediately for the task, rather than just a full conversation transcript.\u003c/p\u003e\n\u003cdiv class=\"overflow-x-auto\"\u003e\u003ctable\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth\u003eInformation transferred\u003c/th\u003e\n\u003cth\u003eSpecific details\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Information transferred\"\u003eCustomer intent\u003c/td\u003e\n\u003ctd data-label=\"Specific details\"\u003eThe outcome the customer ultimately wants\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Information transferred\"\u003eKey facts\u003c/td\u003e\n\u003ctd data-label=\"Specific details\"\u003eConfirmed information such as order number, time of occurrence, product, and amount\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Information transferred\"\u003eAuthentication status\u003c/td\u003e\n\u003ctd data-label=\"Specific details\"\u003eHow identity verification was completed and the scope for which it remains valid\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Information transferred\"\u003eActions performed\u003c/td\u003e\n\u003ctd data-label=\"Specific details\"\u003eActions the chatbot retrieved or executed and their results\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Information transferred\"\u003eCause of failure\u003c/td\u003e\n\u003ctd data-label=\"Specific details\"\u003eInsufficient authority, policy exception, API error, low confidence, and so on\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Information transferred\"\u003eConversation summary\u003c/td\u003e\n\u003ctd data-label=\"Specific details\"\u003eCustomer’s claims, explanations already provided, and remaining questions\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Information transferred\"\u003eEmotion and risk signals\u003c/td\u003e\n\u003ctd data-label=\"Specific details\"\u003eIntensity of dissatisfaction and whether security, safety, or legal risks are present\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Information transferred\"\u003eEvidence\u003c/td\u003e\n\u003ctd data-label=\"Specific details\"\u003eVersion of the policy document used and related system records\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\u003c/div\u003e\n\u003cp\u003eBecause AI-generated summaries may contain errors, the original conversation must also be available for review. Authentication information and sensitive personal information should be transferred only to the extent necessary, with access permissions and retention periods managed appropriately.\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#a-three-tier-operating-model-for-dividing-the-roles-of-ai-and-humans\" class=\"anchor\" id=\"a-three-tier-operating-model-for-dividing-the-roles-of-ai-and-humans\"\u003e\u003c/a\u003eA Three-Tier Operating Model for Dividing the Roles of AI and Humans\u003c/h2\u003e\n\u003cp\u003eIt is safer to divide the level of automation based on task risk, frequency of exceptions, and responsibility for judgment.\u003c/p\u003e\n\u003cdiv class=\"overflow-x-auto\"\u003e\u003ctable\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth\u003eArea\u003c/th\u003e\n\u003cth\u003eSuitable tasks\u003c/th\u003e\n\u003cth\u003eRole of AI\u003c/th\u003e\n\u003cth\u003eRole of humans\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Area\"\u003eAutomation\u003c/td\u003e\n\u003ctd data-label=\"Suitable tasks\"\u003eBusiness hours information, shipment tracking, low-risk reservation changes\u003c/td\u003e\n\u003ctd data-label=\"Role of AI\"\u003eRetrieval, explanation, and task execution\u003c/td\u003e\n\u003ctd data-label=\"Role of humans\"\u003eIntervene when exceptions occur\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Area\"\u003eCollaboration\u003c/td\u003e\n\u003ctd data-label=\"Suitable tasks\"\u003eComplex product inquiries, review of policy exceptions, general complaints\u003c/td\u003e\n\u003ctd data-label=\"Role of AI\"\u003eOrganize context, search for evidence, and suggest draft responses\u003c/td\u003e\n\u003ctd data-label=\"Role of humans\"\u003eVerify facts, make final decisions, and communicate\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Area\"\u003eHuman-only\u003c/td\u003e\n\u003ctd data-label=\"Suitable tasks\"\u003eLegal disputes, high-value refunds, account takeovers, safety crises, serious emotional care\u003c/td\u003e\n\u003ctd data-label=\"Role of AI\"\u003eSearch records and provide supporting materials\u003c/td\u003e\n\u003ctd data-label=\"Role of humans\"\u003eExercise accountable judgment, approve actions, and restore the relationship\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\u003c/div\u003e\n\u003cp\u003eThe classification criterion should not be whether “AI writes well,” but the potential harm of incorrect handling and the degree to which it can be reversed. It is appropriate to retain human review and approval for tasks involving significant financial, legal, or safety risks or frequent exceptions.\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#shifting-kpis-from-deflection-rate-to-journey-completion-rate\" class=\"anchor\" id=\"shifting-kpis-from-deflection-rate-to-journey-completion-rate\"\u003e\u003c/a\u003eShifting KPIs From Deflection Rate to Journey Completion Rate\u003c/h2\u003e\n\u003cp\u003eThe deflection rate or automation rate is often used to mean the percentage of sessions not transferred to an agent. However, the metric can improve even when the path to human support is made difficult or customers give up on resolving their issue. Using it as a standalone KPI therefore risks encouraging optimization in the opposite direction from the customer experience.\u003c/p\u003e\n\u003cp\u003eA journey-centered measurement framework should answer the following questions.\u003c/p\u003e\n\u003col\u003e\n\u003cli\u003eWas the task the customer wanted actually completed?\u003c/li\u003e\n\u003cli\u003eDid the customer make contact again about the same issue?\u003c/li\u003e\n\u003cli\u003eHow many channels and steps were required to reach resolution?\u003c/li\u003e\n\u003cli\u003eHow many times did the customer have to repeat the same information or authentication?\u003c/li\u003e\n\u003cli\u003eHow long did it take from first contact to final resolution?\u003c/li\u003e\n\u003cli\u003eDid an automation error lead to financial loss or a policy violation?\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eA recommended dashboard can be organized into four layers: outcomes, effort, operations, and risk.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cstrong\u003eOutcomes:\u003c/strong\u003e Journey completion rate, first-contact resolution rate, repeat contact rate\u003c/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eCustomer effort:\u003c/strong\u003e Total resolution time, number of repeated explanations, number of channel switches, customer effort score\u003c/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eOperations:\u003c/strong\u003e Phone contact rate, average handling time, wait time, agent occupancy rate\u003c/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eQuality and risk:\u003c/strong\u003e Incorrect answer rate, rate of actions performed without approval, sensitive information exposure, number of disputes and recoveries\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eCosts should not be calculated from call volume alone. Organizations should compare total cost per journey, including AI inference costs, system integration, quality evaluation, security controls, human review time, and error recovery costs.\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#context-inference-orchestration-and-handover-pipeline\" class=\"anchor\" id=\"context-inference-orchestration-and-handover-pipeline\"\u003e\u003c/a\u003eContext Inference, Orchestration, and Handover Pipeline\u003c/h2\u003e\n\u003cp\u003eA practical customer service AI is less like a single chat window and more like a pipeline connecting multiple systems.\u003c/p\u003e\n\u003ch3\u003e\n\u003ca href=\"#1-context-inference\" class=\"anchor\" id=\"1-context-inference\"\u003e\u003c/a\u003e1. Context Inference\u003c/h3\u003e\n\u003cp\u003eIdentify the customer’s intent and necessary conditions, and retrieve authorized customer information and current operational data. The system should not rely solely on the model’s memory, but should use authoritative sources such as policy documents, order systems, and account status.\u003c/p\u003e\n\u003ch3\u003e\n\u003ca href=\"#2-orchestration\" class=\"anchor\" id=\"2-orchestration\"\u003e\u003c/a\u003e2. Orchestration\u003c/h3\u003e\n\u003cp\u003eDetermine which tool or person should receive the request. Before execution, verify authentication status, permissions, monetary limits, and risk level. If the action fails, do not retry indefinitely; switch to a recovery procedure or human support.\u003c/p\u003e\n\u003ch3\u003e\n\u003ca href=\"#3-handover\" class=\"anchor\" id=\"3-handover\"\u003e\u003c/a\u003e3. Handover\u003c/h3\u003e\n\u003cp\u003eSummarize the customer’s goal, confirmed facts, actions performed, cause of failure, and recommended next steps, and transfer them to the agent. The agent reviews the original conversation and supporting evidence before continuing the process.\u003c/p\u003e\n\u003ch3\u003e\n\u003ca href=\"#4-outcome-recording-and-learning\" class=\"anchor\" id=\"4-outcome-recording-and-learning\"\u003e\u003c/a\u003e4. Outcome Recording and Learning\u003c/h3\u003e\n\u003cp\u003eStore whether the issue was ultimately resolved and any corrections made by the agent. For recurring failures, do not merely retrain the model; determine whether the cause lies in policy documents, APIs, operational permissions, or screen flows, and improve the relevant area.\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#experimental-design-for-verifying-implementation-effects\" class=\"anchor\" id=\"experimental-design-for-verifying-implementation-effects\"\u003e\u003c/a\u003eExperimental Design for Verifying Implementation Effects\u003c/h2\u003e\n\u003cp\u003eComparing only total call volume before and after implementation makes it difficult to separate the effects of business growth and seasonality. Where possible, organizations should conduct a phased rollout across comparable customer segments or inquiry types and observe the following items together.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eNumber of calls per 1,000 orders or active customers\u003c/li\u003e\n\u003cli\u003eRepeat contact for the same reason within 24 hours or 7 days after chatbot use\u003c/li\u003e\n\u003cli\u003eCompletion rate and average handling time by inquiry type\u003c/li\u003e\n\u003cli\u003eCustomer effort and satisfaction before and after connection to a human\u003c/li\u003e\n\u003cli\u003ePercentage of AI summaries corrected by agents\u003c/li\u003e\n\u003cli\u003eError rate of automated actions and time required for human recovery\u003c/li\u003e\n\u003cli\u003eAgents’ cognitive load, emotional exhaustion, and job satisfaction\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe definition of “resolution” must also be established in advance. An answer being displayed, the customer ending the conversation, a backend task succeeding, and the customer confirming the outcome are all different states. For tasks requiring follow-up, such as refunds, the actual transaction status must be verified.\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#checklist-for-product-teams\" class=\"anchor\" id=\"checklist-for-product-teams\"\u003e\u003c/a\u003eChecklist for Product Teams\u003c/h2\u003e\n\u003cul\u003e\n\u003cli\u003eHave you distinguished between tasks the chatbot can answer and tasks it can actually execute?\u003c/li\u003e\n\u003cli\u003eDo you verify each task’s success criteria through backend outcomes?\u003c/li\u003e\n\u003cli\u003eCan customers clearly request human support at any time?\u003c/li\u003e\n\u003cli\u003eDo you use repeated questions, low confidence, and risky expressions as transfer signals?\u003c/li\u003e\n\u003cli\u003eCan agents view the original conversation and supporting evidence, not just a conversation summary?\u003c/li\u003e\n\u003cli\u003eCan customers avoid repeating authentication and explanations?\u003c/li\u003e\n\u003cli\u003eAre there limits and approval procedures for the amounts and permissions available to AI?\u003c/li\u003e\n\u003cli\u003eDo you measure repeat contact rates and total resolution time by inquiry type?\u003c/li\u003e\n\u003cli\u003eDoes agent feedback lead to improvements in knowledge documents, tools, and business processes?\u003c/li\u003e\n\u003cli\u003eDo actual operations staff regularly test the entire chatbot journey as customers?\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch2\u003e\n\u003ca href=\"#conclusion\" class=\"anchor\" id=\"conclusion\"\u003e\u003c/a\u003eConclusion\u003c/h2\u003e\n\u003cp\u003eThe success of an AI chatbot cannot be judged solely by conversation volume or the human-support deflection rate. If the customer’s objective is not actually completed, the chatbot becomes an additional step that must be passed through before calling, rather than a problem-resolution channel.\u003c/p\u003e\n\u003cp\u003eEffective design does not hide the limitations of AI. It automates standardized, low-risk tasks, supports agents in work requiring judgment, and rapidly transfers high-risk situations to skilled humans. When this is combined with handover that transfers both the conversation and task status, AI can become a tool that reduces customer resolution time rather than a barrier that blocks access to the call center.\u003c/p\u003e\n","tags":["AI chatbot","Call center","CX","Customer service automation","Handover"],"faqs":[{"question":"If the number of chatbot users increases, shouldn't call center calls decrease too?","answer":"Not necessarily. Increased chatbot usage can also result from greater exposure, customer growth, or a shift from existing FAQ channels. If the same customer calls again after using the chatbot, conversation volume may increase while call volume remains unchanged."},{"question":"Does a high chatbot containment rate mean automation has succeeded?","answer":"It is difficult to judge based on the containment rate alone. The containment rate can also increase if the option to connect to an agent is made difficult to find or if customers give up on resolving their issue. Actual task completion, whether customers make repeat inquiries, total resolution time, and customer effort should be measured together."},{"question":"Why might average handling time increase after introducing a chatbot?","answer":"If AI handles simple inquiries, complex exceptions and emotionally charged complaints may become concentrated among agents. Because of this change in case mix, handling time per agent interaction may increase even if call volume decreases."},{"question":"What should a good handoff to a human agent include?","answer":"It should include the customer's ultimate goal, authentication status, verified facts, actions performed by the chatbot, reasons for failure, relevant policies, and a conversation summary. Because AI summaries may contain errors, the agent should also be able to review the original conversation and supporting evidence."},{"question":"Does giving AI the authority to issue refunds or cancellations solve the problem?","answer":"Some journeys will improve, but granting unlimited authority is not safe. Identity verification, least privilege, amount limits, approval conditions, audit logs, and error recovery procedures should be established, and automation should be introduced gradually, starting with low-risk tasks."},{"question":"Are ontologies and knowledge graphs absolutely necessary for customer service chatbots?","answer":"They are not absolutely necessary. They are useful when dealing with complex product and policy relationships, but simple tasks can be handled with structured APIs and clear conversation flows alone. The key is to ask for the necessary context or retrieve it from a reliable system."},{"question":"How do you determine when to hand off to an agent?","answer":"It can be determined based on explicit requests for an agent, repeated questions, low answer confidence, task failures, persistent negative sentiment, and high-risk signals such as payment disputes, account takeovers, and safety issues. High-risk inquiries should not be kept in the chatbot for an extended period merely to reduce costs."},{"question":"What should be used as the chatbot's core KPIs?","answer":"Journey completion rate, first-contact resolution rate, repeat inquiry rate for the same reason, total resolution time, and the number of repeated explanations can be prioritized. Call contact rate, average handling time, error rate, and total cost per journey should also be reviewed."},{"question":"How can you verify whether a chatbot has reduced call center costs?","answer":"Rather than comparing only total call volume before and after implementation, calculate the call contact rate adjusted for the number of customers and transaction volume. It is appropriate to compare the total cost per journey, including completion rates by inquiry type, repeat inquiries, agent handling time, AI operating costs, human review, and error recovery costs."}],"sources":[{"url":"https://www.nist.gov/itl/ai-risk-management-framework","title":"NIST AI Risk Management Framework","type":"source"},{"url":"https://cloud.google.com/solutions/contact-center-ai-platform","title":"Google Cloud Contact Center AI Platform","type":"source"},{"url":"https://www.ibm.com/think/topics/ontology","title":"IBM: What is an ontology?","type":"source"},{"url":"https://learn.microsoft.com/en-us/azure/architecture/ai-ml/guide/human-in-the-loop","title":"Microsoft Azure Architecture Center: Human-in-the-loop","type":"source"}],"images":[{"id":626,"url":"https://injoys.com/rails/active_storage/blobs/proxy/eyJfcmFpbHMiOnsiZGF0YSI6NzU5NywicHVyIjoiYmxvYl9pZCJ9fQ==--317ba98209612caa9af6cdc98ef2ee488ef9e21b/ai-97159692.webp","is_representative":true,"generation_method":"ai_image","license":"ai_generated","mime_type":"image/webp","translations":{"ko":{"alt":"단순 문의는 처리하고 복잡한 문제는 대기 줄과 상담원에게 넘기는 AI 챗봇 흐름도","caption":"챗봇이 일부 요청을 해결하지만 오류와 복잡한 문의는 콜센터 상담원에게 이어진다.","description":null},"en":{"alt":"AI chatbot routing simple requests to solutions and complex issues to a call center queue","caption":"The chatbot resolves some requests while errors and complex cases flow to waiting callers and human agents.","description":null},"ja":{"alt":"簡単な問い合わせを処理し、複雑な問題を待機列とオペレーターへ回すAIチャットボット","caption":"チャットボットが一部の依頼を解決する一方、エラーや複雑な相談はコールセンターへ送られる。","description":null},"es":{"alt":"Chatbot de IA que resuelve consultas simples y deriva problemas complejos a la cola del centro de llamadas","caption":"El chatbot resuelve algunas solicitudes, mientras los errores y casos complejos pasan a agentes humanos.","description":null},"id":{"alt":"Chatbot AI menangani permintaan sederhana dan mengalihkan masalah rumit ke antrean pusat panggilan","caption":"Chatbot menyelesaikan sebagian permintaan, sedangkan galat dan kasus rumit diteruskan ke agen manusia.","description":null},"pt":{"alt":"Chatbot de IA resolve pedidos simples e encaminha problemas complexos à fila da central de atendimento","caption":"O chatbot soluciona algumas solicitações, enquanto erros e casos complexos seguem para atendentes humanos.","description":null},"zh-hant":{"alt":"AI 聊天機器人處理簡單需求，並將複雜問題轉入客服中心等候隊伍","caption":"聊天機器人解決部分請求，但錯誤與複雜案件仍會轉交真人客服。","description":null},"de":{"alt":"KI-Chatbot löst einfache Anfragen und leitet komplexe Probleme an die Warteschlange des Callcenters weiter","caption":"Der Chatbot erledigt einige Anliegen, während Fehler und komplexe Fälle an menschliche Agenten gehen.","description":null}}},{"id":627,"url":"https://injoys.com/rails/active_storage/blobs/proxy/eyJfcmFpbHMiOnsiZGF0YSI6NzYwMywicHVyIjoiYmxvYl9pZCJ9fQ==--3ee1a2d62648624b7dfb616c81456a18181d3f5c/ai-6fc792ff.webp","is_representative":false,"generation_method":"ai_image","license":"ai_generated","mime_type":"image/webp","translations":{"ko":{"alt":"고객 문의가 AI 챗봇과 전화·이메일을 거쳐 상담원에게 연결되는 흐름과 분석 대시보드","caption":"AI 챗봇과 여러 문의 채널을 상담원이 이어받아 해결하는 과정을 보여준다.","description":null},"en":{"alt":"Customer inquiry routed through an AI chatbot, phone and email to a human agent, with analytics below","caption":"The illustration shows a human agent resolving inquiries across AI and customer service channels.","description":null},"ja":{"alt":"顧客の問い合わせがAIチャットボットや電話、メールを経て担当者につながる流れと分析画面","caption":"AIと複数の問い合わせ窓口を人間の担当者が引き継いで解決する流れを示している。","description":null},"es":{"alt":"Consulta de una clienta canalizada por chatbot, teléfono y correo hacia una agente, con panel de análisis","caption":"La ilustración muestra a una agente resolviendo consultas procedentes de la IA y otros canales.","description":null},"id":{"alt":"Pertanyaan pelanggan dialihkan melalui chatbot AI, telepon, dan email ke agen, dengan dasbor analitik","caption":"Ilustrasi menunjukkan agen manusia menyelesaikan pertanyaan dari AI dan berbagai kanal layanan.","description":null},"pt":{"alt":"Dúvida de cliente encaminhada por chatbot, telefone e e-mail até uma atendente, com painel de análises","caption":"A ilustração mostra uma atendente resolvendo solicitações vindas da IA e de outros canais.","description":null},"zh-hant":{"alt":"客戶問題經由AI聊天機器人、電話和電子郵件轉交真人客服，下方顯示分析儀表板","caption":"圖中呈現真人客服接手AI與多種服務管道的問題並完成處理。","description":null},"de":{"alt":"Kundenanfrage wird über KI-Chatbot, Telefon und E-Mail an eine Mitarbeiterin geleitet, darunter Analysen","caption":"Die Grafik zeigt, wie eine Mitarbeiterin Anfragen aus KI- und weiteren Servicekanälen löst.","description":null}}}],"published_at":"2026-08-14T01:01:58+09:00","updated_at":"2026-08-14T01:01:58+09:00","license":"cc_by","translation_status":"reviewed","available_locales":["ko","en","ja","es"],"data_locales":["ko","en","ja","es","id","pt","zh-hant","de"],"url":"https://injoys.com/en/articles/why-ai-chatbots-do-not-reduce-call-center-calls"}