面对双十一，智能客服 Agent 既要答得准确，也要服务可靠。如何打通业务数据、规范工具调用，让 Agent 经得起真实业务考验？影腾智联的售后智能客服实践提供了一个参考。
影腾智联聚焦智能家居与物联网消费电子，面向家庭用户和小微商户提供智能摄像头、智能门铃、免插卡 4G 随身 WiFi 等产品及服务。这些产品的售后问题涉及设备、网络、账号与套餐等多个环节，对客服的专业诊断与协同处理能力提出了要求。
为此，影腾智联基于火山引擎 AgentKit 构建售后智能客服“小影”，将其嵌入影腾智联 App，串联知识问答、设备诊断与人工协同，逐步承接真实售后咨询。
本文聚焦售后智能客服“小影”从问答验证走向生产应用的实践，介绍影腾智联如何借助 AgentKit 打通业务工具、保障服务可靠运行，并通过观测与评测持续优化服务质量。
01企业痛点：海量咨询下的效率与服务质量挑战
影腾智联拥有数百人规模的售前与售后客服团队，日均处理工单数量达数万单。为缓解大量重复咨询与排障带来的服务压力，影腾智联希望引入 Agent，降低售后人力成本、提升服务体验的一致性。
- 重复咨询挤占人力，高峰体验不均。 海量重复性咨询（使用教程、故障排查、流量套餐、设备共享等）占据客服大量时间。高峰时段排队、等待时间长，不同客服的应答口径与处理质量难以统一，影响用户体验和店铺评级。
- 设备复杂，自助排障能力不足。 影腾智联设备种类多（太阳能/插电款摄像头、4G/WiFi 款、单目/双目等），故障类型涉及网络、供电、存储、画面等多个维度。同一句"摄像头看不到画面"，可能对应供电、网络制式、信号质量、设备在线状态、套餐、权限或存储等不同原因。App 内自助排障能力有限，大量问题最终转人工。
- 稳定性是电商经营红线。 电商平台对店铺稳定性有严格要求，一旦产品或功能出现大面积不可用，将导致店铺评级降低，严重可能被关停整改，直接影响营收。售后 Agent 不仅要回答问题，更要稳定、可控、可审计。
02解决方案：基于 AgentKit 构建生产级售后服务 Agent
为提升售后服务效率与体验一致性，影腾智联建设售后智能客服“小影”，面向智能摄像头、随身 WiFi 等设备用户，解答设备使用、套餐及配置等高频问题，并结合设备状态与告警数据诊断故障、提供处置建议。
为支撑“小影”稳定承接生产流量，影腾智联选择高代码开发路线，采用火山方舟豆包大模型，基于 VeADK 编排业务流程，并依托 AgentKit 的弹性运行、工具接入、身份权限及观测评测能力，实现诊断过程可追溯、业务数据访问受控、Skill 变更可评测回归，保障服务可靠性与迭代质量。
影腾智联售后 Agent 基于火山引擎 AgentKit 的三层能力构建：
打通知识与工具，
串联 AI 客服的咨询和诊断
“小影”覆盖使用教程、功能咨询、流量套餐和账号配置等高频问题，根据设备型号匹配教程与处理建议。针对设备故障，结合可获取的状态和告警数据，调用设备、套餐及权限等工具，分析可能原因并给出排查步骤。
影腾智联将问题分类、排障规范和转人工策略封装为 Skill，进行版本化管理；存量业务接口经 AgentKit Gateway 转换为 MCP 工具，统一管理鉴权、路由、限流与审计，确保业务数据在授权范围内访问。
对数据不足或无法确认的问题，Agent 携带设备信息和已完成的排查步骤转交人工，减少重复询问，形成咨询、诊断与人工协同的服务闭环。
托管运行，保障服务稳定可靠
售后 Agent 上线后，需要持续处理模型生成、工具调用和多轮交互。对影腾智联而言，运行环境的可靠性直接关系到用户能否顺利完成咨询与故障排查。
“小影”托管于 AgentKit Serverless Runtime，由平台提供运行资源管理与按需扩缩容能力，支持服务随实际负载调整资源，减少团队自行配置和维护运行环境的工作。
业务工具调用则通过 AgentKit Gateway 统一管理，以鉴权和限流等机制控制接口访问。Runtime 与 Gateway 配合，为 Agent 运行和后端业务访问提供支撑，让影腾智联团队能够将更多精力投入排障流程与服务质量优化。
观测与评测，优化响应和效果
智能客服的问题可能出现在模型生成、知识匹配、工具调用或多轮交互等不同环节。仅查看最终回答，难以判断一次诊断为什么出错、用户等待时间消耗在哪里。
借助 AgentKit 的观测能力，影腾智联可以追踪模型与工具调用链路，定位耗时和异常环节，使问题排查从查看最终答复深入到具体执行过程，为调整知识、工具与业务流程提供依据。
影腾智联同时自建会话管理后台，支持会话追溯、摘要、分类、质检与标签管理，并以 3 名金牌客服的聊天记录为参考，提炼服务知识与处理规范，建设人工与 AI 对比分析体系。
真实会话持续沉淀为评测集，用于验证模型、知识和 Skill 的变更。通过“发现问题—定位原因—调整流程—评测回归”的迭代方式，团队能够检查修改是否解决原有问题，以及是否影响其他场景，降低依赖零散对话测试带来的遗漏风险。
引入观测评测一体化之前，影腾智联对售后 Agent 的问题定位与质量把控主要依赖零散对话测试和人工抽检：一旦出现回复异常，难以判断问题出在模型生成、知识匹配、工具调用还是多轮交互环节；质检仅靠每日抽检 10-20 条会话，远低于每日3000条的工单量。
依托 AgentKit 观测评测一体化能力（全链路观测 + 评测与 Skill 中心打通 + 会话管理后台）后，团队实现了"发现问题—定位原因—调整流程—评测回归"的迭代闭环。前后效果对比如下：
03落地成效：接入生产流量，Agent 规模化运行
基于 AgentKit 构建的“小影”已在影腾智联 App 上线，为用户提供设备使用指导与故障排查，并在转人工时同步交接上下文与排查记录，减少重复询问，让客服更专注于复杂问题处理，对于影腾来说，智能客服Agent的建设已经在如下三个方面初见成效：
1. 降低售后人力成本。 智能客服自动承接高频重复咨询与基础排障，减少人工客服接待量，让客服团队聚焦复杂故障与高价值服务，逐步优化售后人力投入，实现降本提效。
2. 提速问题处理。 Agent 可快速调取人工客服难以直接获取的设备状态、套餐与权限等后台数据，结合设备型号精准定位问题、给出排查建议，加快问题解决速度，减少用户等待与重复描述。
3. 服务稳定可靠、质量可衡量。 依托 Serverless 弹性运行与观测评测闭环，Agent 可稳定承接高峰期咨询，会话全流程可追溯、可质检，服务质量从"凭感觉"变为"可观测、可迭代"。
目前，“小影”已逐步进入常态化生产运行，项目累计沉淀会话约 4 万条，日均导入约 3000 条，Token 消耗超过几十亿。真实会话持续转化为知识补充与评测素材，为后续服务优化提供依据，进一步提升客服响应效率与问题诊断准确性。
点击这里，获取 AI 客服示例代码，开启你的客服 Agent 搭建实践。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
In the face of the double 11, the smart customer service Agent needs to be both accurate and reliable. How do we get through business data, standardize tool call, and make Agent stand up to the real business test? The after-sale smart customer service practice of IQW provides a reference.
The Internet-enabled smart home and goods network of the WikiLeaks network provides products and services such as smart cameras, smart doorbells and 4G off-plugs for household users and microbusinesses. After-sale issues related to equipment, networks, account numbers and packages, and required the professional diagnostic and co-processing capacity of the service.
To this end, IFJ is based on Volcengine AgentKit's construction of a post-sale smart customer service, “Small Shadow”, embedded in IUPU App, a series of knowledge questions and answers, equipment diagnostics and manual collaborations, gradually taking over after-sales counselling.
The short-shot of this post-focused smart customer service, from question-and-answer to production applications, describes how IQU, through AgentKit, can access operational tools, guarantee reliable operation of services and continuously optimize the quality of services through observation and evaluation.
Businessache 01: Efficiency and quality of service challenges with mass counselling
IFJ has hundreds of pre-sale and post-sale customer service teams, with a daily average of tens of thousands of work orders. In order to alleviate the pressure on services resulting from a large number of repeated counselling and exclusions, IQF hopes to introduce Agent to reduce the human cost of selling and improve the consistency of service experience.
- Repetition of consultations over manpower, with uneven peak experience. Mass repeated counselling (use of tutorials, failure screening, flow packages, equipment sharing, etc.) takes up a significant amount of time for passenger service. The peak hours are queued, long waiting times, and the response calibration and processing quality of different customer service are difficult to reconcile, affecting user experience and store ratings.
- Complex equipment and insufficient capacity for self-help barriers. There are many types of IQU equipment (solar/interpolated cameras, 4G/WiFi, mono/biscope, etc.) and the type of failure involves multiple dimensions such as network, power, storage, image, etc. In the same sentence, "the camera cannot see the picture", there may be different reasons such as power supply, network style, signal quality, equipment online, food packages, permission or storage. App has limited capacity for internal self-help barriers and a large number of problems end up in labour.
- Stability is a red line run by an electrician. Electrician platforms have strict requirements for shop stability, and when products or functions become unusable in large areas, they result in a lower rating of the shop, which can be seriously disrupted and modified, directly affecting the collection. After the sale, Agent not only answers questions, but also is stable, manageable and auditable.
02 Solution: Construction of after-sales service based on AgentKit
To improve the efficiency and consistency of post-sale services, the IQU built the smart customer service “Small Shadow”, which is designed for users of equipment such as smart cameras, walk-in WiFi, to answer high-frequency (HF) problems such as the use of equipment, food packages and configurations, and to provide recommendations for disposal in conjunction with equipment status and alarm data.
To support the stabilization of the production flow of the small image, IFAP selected a high code development route, using the Volcanic Ark Doubao LLM to organize business processes based on VeADK and based on AgentKit's flexible operation, tool access, identity permission and observational assessment capabilities to achieve traceability, operational data access control, Skill change assessable regression, and to safeguard service reliability and iterative quality.
Agent is built on three layers of Volcengine AgentKit:
This post is part of our special coverage Global Voices 2011.
Advice and diagnostics for a serial AI client service
`Small image' covers high-frequency issues such as the use of tutorials, functional counselling, traffic packages and account configurations, matching the curriculum and processing recommendations according to the equipment model. In the case of equipment malfunctions, a combination of accessible status and alarm data, tools such as equipment, food packages and permission are called, possible causes are analysed and steps are given for screening.
IFJ has sealed the issue classification, debarment regulation and conversion strategy to Skill for version management; the inventory business interface has been converted from the AgentKit Gateway to the MCP tool, which centrally manages assurance, route, rate limit and audit to ensure that business data are accessed within delegated authority.
In cases where data are inadequate or unconfirmable, Agent carries equipment information and completed clearance steps to be handed over manually, reduces repeat queries and creates a service closed loop for consultation, diagnosis and manual collaboration.
Host operation, secure service.
After selling Agent online, continuous processing of model generation, tool call and multi-wheel interaction is required. For IAF, the reliability of the operating environment is directly related to the ability of users to successfully complete counselling and troubleshooting.
“Small Image” is hosted by AgentKit Servers Runtime, which provides operational resource management and needs-based build-up capacity from the platform to support services to adjust resources to actual loads and reduce team self-configuration and maintenance of operating environments.
Operation tool call is managed centrally through AgentKit Gateway to control access through mechanisms such as assurance and rate limit. Runtime, in collaboration with Gateway, supports Agent 's running and back-end business access, allowing the IQW team to devote more energy to the process and service quality of the barrier.
Observation and evaluation, optimizing response and effectiveness
Problems with intelligent passenger service may arise at different stages of model generation, knowledge matching, tool call or multi-wheel interaction. It is difficult to judge why a diagnosis went wrong and where the user waits.
Through AgentKit ' s observation capability, IQU can track the model and tool call links, locate time-consuming and unusual links, allow question-checking from the final response to the specific implementation process, and provide a basis for adjusting knowledge, tools and business processes.
IFJ also built its own back-office for session management, supporting session traceability, summary, classification, quality checking and tag management, using as a reference the chat logs of three gold-soldier guest clothes, extracting service knowledge and processing norms, and building manual and AI comparative analysis systems.
Real sessions continue to sink into assessment and are used to validate model, knowledge and Skill changes. Through the iterative approach of Discovery — Reasons for Positioning — Repositioning Processes — Assessing Return, teams are able to check whether the modification addresses the original problem and affects other scenes, reducing the risk of leakage from reliance on sporadic dialogue testing.
Prior to the introduction of the integration of observational assessments, FIT relied mainly on sporadic dialogue tests and manual sampling for the problem positioning and quality control of Agent after sale: in case of a response anomaly, it was difficult to judge whether the problem arose from model generation, knowledge matching, tool call or multiple rounds of interaction; and the quality check was based only on daily sampling of 10-20 sessions, which was well below 3,000 work orders per day.
Following the AgentKit observational integration capability (whooping full-link observation + assessment and communication with the Skill Center + session management backstage), the team achieved an iterative closure of the Discovery Problem - Positioning Reason - Adjustment Process - Assessing Return. The effects are compared as follows:
03 Landing effects: access to production flows, Agent scale operation
The "Small Shadow" based on AgentKit is on-line with the IQU App to provide users with guidance on the use of equipment and troubleshooting, and to synchronize the context and Quest logs when handing over labour, to reduce the number of repeated queries and to focus more on complex issues. The construction of the smart guest service Agent has begun to yield results in three respects:
Reduce the cost of after-sale manpower. Smart passenger service automatically covers high frequency repeated counselling and basic debarment, reduces the volume of manual passenger service reception, allows the passenger service team to focus on complex malfunctions and high-value services, gradually optimizes post-sale manpower inputs and achieves reduced impact.
Speeding up. Agent is able to quickly access back-office data such as the state of equipment that is difficult to obtain directly from the manual passenger service, the food package and permission, to combine precision positioning problems with equipment models, to give recommendations for queuing, to speed up problem resolution and to reduce user waiting and repetition.
3. The services are reliable and of measurable quality. Based on the Serverless elasticity and observational closed loops, Agent is able to carry out a steady peak consultation, full conversational process is traceable, qualitative, and quality of service is changed from "favourable" to "observable, iterative".
At present, the “small shadow” has gradually entered a normal production operation, with the cumulative deposition of the project with about 40,000 sessions, with an average daily import of about 300 million and a consumption of over billions. The continuous transformation of genuine sessions into knowledge supplementation and assessment of asset provides a basis for the optimization of follow-up services and further improves client response efficiency and problem diagnosis accuracy.
Click here, get an AI client service example code, open your customer service Agent build practice.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.