eGain Says Broken Knowledge Layers Drive Enterprise AI Failures

eGain称知识层缺陷导致企业AI项目失败 在碎片化且缺乏治理的内容管道上部署生成式人工智能会放大操作失误和合规风险，除非企业首先对源头数据治理进行彻底整改。 John Copeland概述了企业AI项目为何因内容管道脆弱性而受挫，并警告称，若将模型直接建立在陈旧、重复或冲突的数据之上，必然会产生不准确的回答和合规风险。为了取代混乱的“意大利面式”架构，他提出了一种由持续的“AI知识运维”进行治理的集中式AI知识中心，其递归生命周期涵盖源头获取、综合提炼和可审计性。该演讲深入阐述了eGain的知识成熟度模型，并展示了分析客户对话记录如何帮助企业将查询复杂性与合规严格程度相匹配，从而在削减处理成本的同时加速知识转型。 企业AI部署在客户体验领域经常失败，因为模型会放大不良、未经治理的数据，而不是对其进行修复。 (0:04) 知识管道存在四大核心故障：孤岛式知识库、缺乏源头级质量保障、将知识管理视为偶发性清理工作，以及防护栏设置过晚。 (1:43) 企业需要借鉴DevOps模式建立“AI知识运维”这一持续运营规范，以实现对企业知识的主动治理。 (1:43) 五大关键内容陷阱会破坏检索与生成效果：内容重复、陈旧过时、覆盖空白、分类体系不一致以及作者归属模糊。 (3:53) 行业预测显示，如果没有干净且适用于AI的数据基础，多达60%的企业AI项目最终将被废弃。 (3:53)

John Copeland, Vice President of Marketing at eGain, argues that enterprise customer experience AI initiatives fail primarily due to fragile, ungoverned data pipelines rather than generative model limitations. To prevent inaccurate responses and legal liabilities, organizations must transition from episodic content cleanups to continuous, demand-driven knowledge operations that validate information at the source. Structural Failures in the Content Pipeline (1:43) Copeland asserts that enterprise AI deployments repeatedly break down because organizations feed artificial intelligence models from uncurated, siloed sources like SharePoint, Confluence, and legacy search tools. This haphazard routing creates five foundational data traps: duplicated content that prompts retrieval systems to average conflicting facts, un-timestamped documents that cause models to serve outdated information, coverage gaps that trigger hallucinations, inconsistent metadata taxonomies that bypass security permissions, and conflicting document authorship. Applying guardrails only at the output stage cannot correct these root-level defects, exposing heavily regulated sectors like healthcare, banking, and government to direct legal liability. Underscoring the urgency of foundational data hygiene, Copeland references an industry forecast projecting that organizations will abandon 60% of their AI initiatives within the year if their deployments lack AI-ready data foundations. Rather than treating knowledge management as an occasional cleanup task, he advocates for "AI knowledge ops"—a continuous operational discipline modeled after DevOps to monitor, test, and govern internal information pipelines at every stage. Hub Architecture and Recursive Governance Lifecycles (10:32) To dismantle fragmented content pipelines, Copeland outlines an AI Knowledge Hub architecture where subject matter experts and knowledge managers curate disparate data into a single, verified source of truth. Once centralized, information is managed through a recursive five-stage lifecycle: cross-repository sourcing, curation and synthesis, continuous governance, personalized publishing, and analytics. Automated audit trails track ongoing regulatory compliance, while strict escalation parameters automatically shut off AI generation and route customers to human specialists during high-risk scenarios. Copeland maintains that unifying the knowledge layer directly enhances operational metrics, including first-contact resolution, average handle time, and Net Promoter Scores. He cites a case study from a global consumer packaged goods manufacturer presenting at an eGain conference, which reported that implementing single-source automation reduced its blended cost per inquiry from $10.00 down to 15 cents—a metric Copeland notes was audited and confirmed by the enterprise's corporate finance department. Transitioning to Demand-Driven Knowledge Operations (16:20) eGain categorizes enterprise readiness through a five-stage Knowledge Maturity Scale: Ad Hoc, Managed, Designed, Optimized, and Transformative. Copeland reports that approximately 80% of organizations remain trapped in the lowest two tiers—where content is either fragmented or inaccessible to AI parsing—while only about 4% reach the transformative stage required for fully compliant generative AI. He cautions that deploying artificial intelligence over an unorganized, ad hoc data foundation will inevitably fail. To bridge this maturity deficit, Copeland introduces the eGain AI Knowledge Method, which redirects organizational focus away from cataloging static data repositories toward analyzing actual customer demand. By evaluating real-world interactions across three dimensions—value, volume, and operational complexity—the system isolates priority inquiries termed "star questions." Copeland claims that using automated AI tools to curate, validate, and personalize answers for these high-impact questions allows organizations to complete knowledge transformations 90% faster and achieve time-to-value in one-third of the time required by manual engineering. Conversation Mining and Strictness Mapping (19:54) The operational implementation of demand-driven knowledge management begins by mining historical customer interaction transcripts across service channels. The platform extracts diverse colloquial phrasings—such as varied formulations for rescheduling a flight—and harmonizes them into single underlying customer intents. Administrators then ingest these transcripts to generate visual cluster graphs that map topic similarity against operational complexity and regulatory compliance strictness. Mapping queries along these two axes informs how an enterprise deploys AI: low-complexity, non-regulated inquiries can be safely automated with generative responses, whereas complex, highly regulated topics require rigid, deterministic instruction models and human oversight. Visualizing these topic clusters also highlights existing coverage gaps and surfaces stale legacy articles, enabling content administrators to strategically target remediation where customer demand intersects with business risk. Copeland's presentation was ended abruptly by conference time limits before he could complete the platform demonstration or address long-term operational costs. While the framework provides a structured blueprint for governing enterprise knowledge pipelines, successfully deploying it hinges on an organization's willingness to commit ongoing resources to continuous data operations and conversation analytics.

eGain营销副总裁John Copeland指出，企业客户体验（CX）人工智能项目的成败取决于知识层的数据治理，单纯在交付末端设置安全护栏无法修复源头数据缺陷。他提出必须建立类似DevOps的持续“AI知识运维”机制，通过统一知识中枢与客户真实对话驱动的方法，解决内容无序并防范高监管行业的法律合规风险。 内容管道结构性缺陷与五大基础陷阱 (1:43) Copeland指出，企业AI部署在客户体验领域频频受挫，根源在于内容供应链的脆弱性。主要表现为四个运营层面的失误：缺乏治理的信息孤岛（如多部门无监督维护的SharePoint冲突文档）、数据源头缺少质量控制（QA）层、将知识管理误当作一次性清理项目而非持续运营机制，以及将AI安全护栏置于交付末端而非数据源头。Copeland特别强调，在银行、医疗与保险等高度监管行业，源头数据错误会直接带来法律责任与监管合规风险。 此外，Copeland将企业内容混乱现象概括为“意大利面图”（Spaghetti Chart）：分散在网站、PDF、Confluence及CRM中的非结构化文档直接杂乱接入自建检索系统，导致输出相互冲突。具体表现为五大陷阱：重复内容使RAG系统计算冲突文档均值；陈旧文档缺少时间戳致使大语言模型调取过期信息当作事实；覆盖空白促使AI生成幻觉回答；分类法标签不一致破坏基于角色的访问权限并引发数据泄露；文档归属权模糊更使系统无法在政策冲突时判定真实依据。Copeland援引市场预测警示，由于缺乏AI就绪数据，企业在当年内将放弃60%的AI项目。 集中式知识中枢与闭环递归治理 (10:32) 为打破无序的网状架构，Copeland建议构建集中式知识中枢（Knowledge Hub），由业务专家与知识管理者协同提炼企业内容，形成唯一的真实数据源（single source of truth）。统一的内容在通过基于角色的权限控制与个性化规则配置后，再分发至数字机器人、官方网站及人工坐席。Copeland引用其伦敦大会上一家跨国消费品（CPG）制造商的案例说明业务成效：该企业经公司财务部门审计验证，通过工作流自动化与首问准确解答，将单次客户咨询的综合成本从10美元削减至15美分，并同步改善了首次联络解决率（FCR）、平均处理时长（AHT）及净推荐值（NPS）。 知识管理并非一次性的线性工程，而是一个闭环递归的生命周期：涵盖源头对接、提炼综合、持续治理、个性化发布以及运营指标分析。eGain平台内置自动化审计追踪，以适应不断变化的监管程序；在个性化发布环节，治理规则不仅界定访问权限，还设立了严格的升级边界——当系统识别到高风险问询时，会自动切断AI应答并将客户直接转接给人工专家，同时借助闭环分析指标指导后续知识库更新。 知识成熟度模型与需求驱动方法 (16:20) Copeland将企业知识准备程度划分为五个阶段：临时阶段（Ad Hoc，数据分散且缺乏分类）、管理阶段（Managed，具备集中存储但缺乏统一编写规范，致使多达80%的知识无法被AI解析）、设计阶段（Designed，有统一定义但系统脱节且缺乏上下文交互）、优化阶段（Optimized，融入情境化个性化应答）以及变革阶段（Transformative，全面结构化且满足合规要求）。据Copeland透露，约80%的企业仍停留在低成熟度水平，仅有约4%达到变革阶段。他告诫处于初始阶段的企业切勿盲目接入AI，必须“先理顺知识库，再叠加AI应用”。 为弥合这一差距，Copeland提出“eGain AI知识方法”，主张将重心从被动管理现有堆积数据，转向精准识别客户的真实知识需求。该方法通过分析客户实际交互记录，从价值、咨询量和复杂度三个维度对问题进行量化评分；在三项指标上均表现突出的话题被归类为“明星问题”（star questions）。平台将此类重点问题与企业知识库精准匹配、提炼内容并配置规则，据其介绍，借助AI驱动的自动化审核，企业知识转型效率比传统手工方式提升了90%。 真实交互解析与多维AI治理分配 (19:54) Copeland阐述了如何从海量客服对话记录中提取真实意图。在客户服务场景中，多类不同的口语化表达（例如关于改签航班的数十种提问方式）最终指向完全相同的底层意图。通过直接接入通话与对话记录数据库，系统能够系统化提取原始问题，进行语义相似度归类，并降低对传统手工知识工程的依赖，将实现价值的时间（time-to-value）缩短至传统方式的三分之一，同时提升合规性与答复准确度。 在实操流程中，管理员只需上传交互记录文本并添加基础元数据，后台系统即可生成问题与主题相似度的可视化图谱。系统将业务操作复杂度与合规严格度作为两个核心维度进行交叉映射，根据聚类在坐标系中的分布，为不同主题分配匹配的AI模型与治理策略。管理员还可通过生成的层级化主题视图审查现有知识覆盖面，精准排查高频疑问、陈旧信息和内容缺失。 Copeland强调了以AI知识运维和规范知识层为底层支撑的重要性，但由于演讲时间限制，现场演示在展示管理员界面层级化主题审查处戛然而止；演讲并未进一步展开说明在复杂监管与高动态业务环境下，具体如何自动化维护庞大相似度图谱的长期准确性。

## Background 本内容由 Ai4 于 2026 年 8 月 17 日发布在 YouTube 平台上，演讲人为 eGain 市场营销副总裁约翰·科普兰（John Copeland）。视频简介指出，当前企业在客户体验（CX）中仓促部署 AI 暴露了底层知识碎片化与失管的致命弊端，本演讲旨在探讨如何超越传统知识管理，构建可靠的企业级 AI 指令层，从而为自助服务、座席辅助及智能体自动化提供坚实支撑。 ## TL;DR 演讲系统性地剖析了企业 AI 项目频频受挫的根源——“内容管道的脆弱性”与散乱的“意大利面式”知识架构，指出重复、陈旧、盲区、分类不一及权属不明等五大硬伤彻底摧毁了 AI 的可信度。面对 80% 的企业知识对 AI 不可见的行业现状，科普兰提出必须建立涵盖全生命周期的“AI 知识运维”与统一中枢，并分享了一家快消巨头将单次呼叫成本从 10 美元降至 15 美分的标杆案例。最后，eGain 演示了其需求驱动的“AI 知识方法论”，展示了如何通过自动化聚类分析真实对话并筛选高频核心问题，将知识梳理速度提升 90%，实现兼顾严苛合规与全渠道一致性的智能化升级。 ## Atmosphere & style 整场演讲氛围轻松幽默且兼具严谨的实操干货，科普兰在傍晚时分登台，以自嘲“我们是最领先者”和现场悬赏找官网 Bug 的风趣互动迅速拉近与观众的距离。作为深谙业务痛点的前数据分析主管，他在共情知识管理者繁重苦差的同时，以敏锐而犀利的行业洞察直击 AI 乱象本质。伴随着生动的架构图谱拆解与振奋人心的真实财务降本数据，全场节奏由剖析痛点的共鸣逐渐过渡到展示前沿解决方案的笃定与信服，带给观众清晰、专业且极富启发性的视听体验。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=4s) AI 在客户体验中频繁出错的核心根源是底层“内容管道的脆弱性”，给大模型喂入破碎混乱的知识只会放大错误输出。 - [00:04](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=4s) 企业在知识治理上常犯的四大硬伤包括：知识底座缺乏治理、源头缺少质量把控、将知识整理当成突击大扫除，以及将安全护栏置于输出末端而非源头。 - [03:53](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=233s) 典型大型企业普遍陷入“意大利面架构”困境，底层知识在官网、PDF、SharePoint 与 CRM 之间割裂，导致前后矛盾的错误答案向全渠道蔓延扩散。 - [05:14](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=314s) 瓦解 AI 可信度的五大内容硬伤分别是：内容重复、内容陈旧、覆盖盲区、分类法不一致以及作者权属不明与政策冲突。 - [05:14](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=314s) Gartner 预测企业在整年内将放弃高达 60% 缺乏“AI 就绪数据（AI-ready data）”支持的人工智能项目。 - [10:42](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=642s) 企业必须构建统一的知识指令层，将散落的数据经过领域专家与知识管理员提炼策展，归入单一事实源（Single Source of Truth）的知识中枢后再行分发。 - [11:56](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=716s) 实现全渠道知识一致性能够大幅提升首呼解决率（FCR）并显著降低平均处理时长（AHT），从而直接拉动客户净推荐值（NPS）。 - [13:09](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=789s) 知识管理是包含“多源接入、提炼分类、治理审计、个性化发布、业务反馈分析”五大环节的持续闭环生命周期，且支持在高风险时主动熔断 AI 转交人工。 - [13:09](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=789s) 全球顶尖快速消费品制造企业通过标准化知识体系，将加权后的单次客户呼叫处理成本从 10 美元断崖式降至 15 美分（$0.15）。 - [16:20](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=980s) 在企业知识成熟度模型中，约 80% 的企业仍处于孤岛化的初始阶段或低成熟度阶段，仅有约 4% 达到足以全面驱动 Agentic AI 的变革阶段。 - [16:20](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=980s) 企业内部目前约有 80% 的知识因为格式不兼容而对 AI 不可见，盲目在初级知识库上叠加 AI 是本末倒置，必须“先理顺知识库，再部署 AI”。 - [18:54](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=1134s) eGain AI 知识方法论主张从“拥有什么”转为“用户正在寻找什么”的需求驱动模式，基于真实对话从价值、体量和复杂度三个维度聚类筛选明星问题。 - [18:54](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=1134s) 通过 AI 自动化知识提炼与聚类图谱工作流，企业知识梳理构建速度可提升 90%，业务价值落地周期（Time to Value）缩短至原来的 1/3。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=4s) eGain 登顶行业榜首与内容管道的脆弱性 演讲在一片轻松的氛围中拉开帷幕。时近傍晚，eGain 市场营销副总裁约翰·科普兰（John Copeland）向坚守在现场的观众致谢，随即切入当天的核心话题：AI 时代的“垃圾进，可信出”（Garbage In, Trusted Out），以及知识管理对于企业级人工智能成败的决定性作用。科普兰幽默地调侃道，台下观众很可能从未听说过 eGain 这家公司，甚至会纳闷“eGain 到底是谁”。他随即展示了最新出炉的行业权威背书：在几周前刚发布的 Gartner 首届客户服务知识管理系统魔力象限（Magic Quadrant for Customer Service Knowledge Management Systems）中，eGain 不仅被评为领导者，更在执行能力（Ability to Execute）与愿景完整性（Completeness of Vision）两个维度双双位居榜首。科普兰笑着借用了一个自创词称自己是“最领先者”（leaderist）。作为一家深耕知识管理长达 27 年的行业老兵，eGain 见证了最新 AI 技术如何为这一传统领域注入颠覆性的能量。 在客户体验与服务场景中，为什么知识管理对 AI 来说至关重要？科普兰一针见血地指出，AI 之所以频繁翻车，根源在于“内容管道的脆弱性”（content pipeline fragility）。当底层喂给 AI 的知识本身残破混乱时，再强大的大模型也只会生成荒腔走板的答案。科普兰梳理了造成这一脆弱性的四大核心硬伤： - **缺乏治理的知识底座**：企业内部往往运行着数个彼此割裂的 SharePoint，同一主题的内容由不同员工各自撰写，多版本并存却无人统筹监管，导致底层基础异常薄弱。 - **源头端缺失质量保证（QA）**：企业在把答案推送给 AI 机器人或一线客服之前，没有在内容源头上进行严格的质检与审核，错误信息未经拦截便直接推向终端。 - **将知识管理视为一次性大扫除**：许多企业错误地把整理知识库当成突击清扫项目，但知识管理实际上是一门持续运营的严谨学科。eGain 将其定义为“AI 知识运维”（AI Knowledge Ops），类比于软件工程领域的 DevOps，强调必须建立常态化、持续在线的治理流程。 - **安全护栏置于流程末端**：许多组织寄希望于在 AI 最终输出时才加装护栏，但科普兰强调，“末端的安全护栏根本无法修复源头缺乏治理所造成的错误”。特别是在银行、保险、医疗保健及政府机构等强监管行业，提供错误答案不仅破坏客户体验，更会带来极其严重的法律合规责任与罚款风险，因此护栏必须从最前端筑起。 --- ### [03:53](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=233s) 混乱的“意大利面图”与内容断层 科普兰展示了一张在客户企业起步阶段屡见不鲜的架构现状图，并亲切地称之为“意大利面图”（spaghetti chart）。画面中纷繁交错的连线如同乱麻一般，形象地勾勒出典型大型企业的知识断层现状：底层散落着官网、PDF 文档、多个独立部署的 SharePoint、Confluence 页面以及 CRM 系统，各自接入不同的检索通道、搜索后台或自研的知识库系统。 图上密密麻麻标出的问号与红叉，记录了各种糟糕体验的诞生瞬间——员工或系统在错误的地点调取了过时的源文件，导致得出的答案完全错误。更致命的是，这些错误答案会一路顺着渠道扩散蔓延，无论是客户直接浏览网页、使用手机 App，还是与呼叫中心客服、官网数字员工乃至外勤人员沟通，最终都会接收到前后矛盾的错误回答。问题不在于企业没有正确答案，而在于正确的知识从未被正确地连通与分发。 ```mermaid flowchart TD accTitle: 碎片化知识体系的意大利面架构混乱流程 subgraph SOURCING["底层碎片化源头"] S1["企业官网 / PDF 文档"] S2["割裂的 SharePoint / Confluence"] S3["各类 CRM 系统"] end subgraph RETRIEVAL["混乱的检索层"] R1["自研知识管理工具"] R2["分散的搜索后台"] R3["互不连通的索引管道"] end subgraph CHANNELS["终端受害者"] C1["呼叫中心坐席"] C2["移动端与数字客服"] C3["外勤人员与官网用户"] end S1 -. 错误引用.-> R2 S2 -. 抓取过时信息.-> R1 S3 -. 孤岛阻隔.-> R3 R1 ==>|"错误传播 (X/?)"| C1 R2 ==>|"混乱回答 (X/?)"| C2 R3 ==>|"合规暴雷 (X/?)"| C3 ``` --- ### [05:14](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=314s) 瓦解 AI 可信度的五大内容硬伤 科普兰深入剖析了导致 AI 彻底失信的五大内容层顽疾。首当其冲的是**内容重复（Duplication）**。同一业务流程往往在不同系统存在多个版本，当部署 RAG（检索增强生成）系统时，大模型在缺乏明确指令与护栏的情况下只会“在多个版本之间求平均”，自作聪明地东拼西凑，导致同一个渠道甚至不同渠道针对同一问题给出的解答五花八门。 第二个顽疾是**内容陈旧（Staleness）**。科普兰调侃现场观众上一次系统性清理老旧 SharePoint 是在什么时候，并打趣地自嘲道，当天就有人在 eGain 自己的官网上翻出了一条陈年旧内容；他现场悬赏 5 美元给能在周四会议结束前找到该错误的观众，并附带提示“与人事相关”。他坦言，模型本身无法辨别内容的新旧，如果文档未标注有效时间，AI 便会堂而皇之地检索两年前的作废规定并呈递给客户。为了让 Claude 产出准确答案，他自己平时都必须附加大量详尽的 Prompt 来防范过时内容。 第三是**覆盖盲区（Coverage Gaps）**。当业务中存在客户或员工频繁咨询、但系统里根本未曾沉淀相关知识的领域时，AI 会本能地抓取字面最相关的边缘信息强行拼凑，造成空洞的套话或直接“一本正经地胡说八道”（幻觉）。大模型本身没坏，它只是忠实地执行计算，错误在于组织从未发现知识版图上的这些黑洞。 第四是**分类法不一致（Taxonomy Inconsistencies）**。不同团队在跨系统撰写文档时所用的元数据标签各不相同。一个部门标记为“敏感信息”，另一个部门写成“受限信息”，第三个部门则写成“机密信息”。在人类看来这些词语含义相近，但对 AI 系统而言却代表完全不同的数据定义，极易导致基于角色的访问权限控制（RBAC）失效，进而把涉密或敏感信息泄露给无权访问的受众群体。 第五则是**作者权属不明（Ambiguous Authorship）与政策冲突**。当系统里存在多份关于同一命题的文档、且作者与权限不明时，AI 根本无从判断该信任哪一份。科普兰以大家在购买健康保险时的痛苦经历为例：打电话问人工客服，对方明确回答“保单予以报销”；回过头查阅官网，网页却赫然写着“不予承保”；再去问 AI Agent，又给出第三种说法。这种相互打架、左右矛盾的内容让客户无所适从，彻底摧毁了品牌的信任度。 - 混乱的数据生态直接引发项目流产危机：Gartner 预测，在整整一年内，企业将放弃多达 60% 缺乏“AI 就绪数据”（AI-ready data）支撑的人工智能项目。 - 作为前数据科学团队负责人，科普兰感同身受地指出，构建任何模型的前提永远是让数据准备就绪；如果输入端是一堆垃圾，整个 AI 项目必然面临被高层砍掉的厄运。 --- ### [10:42](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=642s) 构建企业级知识指令层与单点真实源 面对盘根错节的意大利面架构，企业必须建立系统化的知识指令层（Knowledge Instruction Layer），从而将无序的数据流转化为清晰可靠的单点真实源。正如前一场圆桌论坛所探讨的那样，首要前提是全面打通包含客服与体验数据在内的所有企业级数据源，完成全面的“源头接入”（Sourcing）。 在这一体系中，企业不能仅仅把原始文件丢给向量数据库，而是必须将内容统一聚合、抽取、提炼与策展。该过程深度融合了领域专家（业务作者，SME）的专业知识与知识管理员（Knowledge Managers）的分类架构能力，最终将沉淀后的高质量知识归入统一的“知识中枢”（Knowledge Hub）。只有在知识中枢建立起无可争议的唯一事实来源（Single Source of Truth）之后，系统才能进一步叠加精细的访问控制与个性化规则（Personalization），决定哪类受众在何种权限下查看特定信息，最终高保真地分发推送到官网、App、坐席工作台及各类数字终端中。 ### [11:56](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=716s) 全渠道一致性带来的业务收益与传统知识管理的困境 当企业能够确保无论客户是通过电话呼入、在官网上自主查询，还是与人工坐席或数字化智能体（Digital Agent）互动，都能获得完全一致且准确的答案时，整个客户体验体系将迎来质的飞跃。首呼解决率（First Contact Resolution, FCR）会大幅攀升，而平均处理时长（Average Handle Time, AHT）则会显著下降。站在客户的角度，如果第一次提问就能迅速得到正确答案，满意度必然提升，企业客户的净推荐值（NPS）也会随之水涨船高。 然而，在现今这一波自动化与智能化技术成熟之前，知识管理一直是一项极其沉重且繁复的纯人工苦差事。演讲者由衷感慨：“我的心与所有知识管理人员同在（My heart goes out to knowledge managers）。”面对庞大企业内浩瀚如海的信息负荷，传统流程完全依靠人力梳理和维护，但这种笨拙的现状正迎来彻底改变。 ### [13:09](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=789s) 知识生命周期管理与跨国企业的降本奇迹 企业必须将知识管理视为一个持续循环的生命周期过程，而非单向线性的任务。从最初连接并吸纳所有业务系统的内容源头开始，团队需要推进知识的整理与提炼，包括构建统一的内容结构、分类法（Taxonomy）与元数据体系；随后进入内容治理与合规阶段，淘汰过时内容并替换上最新资产，同时提供完整的审计追踪（Audit Trails），实时监控内容是否符合法规与标准操作流程。随后是精准发布阶段，系统需界定谁有权限查看哪些内容，并在面临高风险议题时果断决策：“关掉 AI，不让 AI 回答该问题”，及时将其升级转派给人工作席。最后，在反馈与分析环节，闭环监控通话时长、FCR 和运营成本等核心 KPI。 ```mermaid flowchart LR accTitle: 知识生命周期闭环流程 subgraph S1["摄取与整理"] A["多源接入<br/>(Sourcing)"] --> B["提炼分类<br/>(Curation & Taxonomy)"] end subgraph S2["治理与分发"] B --> C["治理与合规审计<br/>(Governance & Audit)"] C --> D["个性化发布与分流<br/>(Publish & Personalization)"] end subgraph S3["闭环优化"] D --> E["业务反馈与分析<br/>(Feedback & Analytics)"] E -.->|持续迭代| B end ``` 演讲者结合自己多年的数据分析背景，分享了一家全球顶尖快速消费品（CPG）制造企业的惊人成效： - **单次通话成本断崖式下降**：该企业跨渠道加权后的单次客户呼叫处理成本从过去的 **10 美元暴跌至 15 美分（$0.15）**。 - **效率提升根源**：得益于知识体系的标准化，绝大多数常规问题实现了自动化处理，复杂问题也在首次呼叫中被迅速正确解决。 - **数据真实可靠**：该企业在伦敦大会上分享这一成果时特别强调，所有数据均经过了财务部门与 CFO 的严格审计与核实。 ### [16:20](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=980s) 知识成熟度模型：为什么80%的企业未准备好迎接AI 并非所有企业都有资格立即落地生成式 AI。在知识成熟度评估模型中，企业通常会经历五个阶段： | 成熟度阶段 | 阶段特征与核心瓶颈 | |:--- |:--- | | **初始阶段 (Ad-hoc)** | 内容孤岛严重，缺乏正式的采集、整理流程与统一分类法，**完全不具备上线 AI 的条件**。 | | **管理阶段 (Managed)** | 建立了单一事实来源（Single Source of Truth），但缺乏标准化的编写规范，知识格式杂乱。 | | **设计阶段 (Designed)** | 拥有统一的单一事实源、结构化内容和元数据，但尚未与后端企业业务系统深度打通，缺乏上下文个性化。 | | **优化阶段 (Optimized)** | 深度整合了提问者的上下文背景，能够根据业务场景进行动态适配。 | | **变革阶段 (Transformative)** | 知识层完全打通并具备高度自动化治理能力，可全面驱动生成式 AI 与 Agentic 自动化应用。 | 演讲者发出明确警告：如果企业仍处于孤岛化的初始阶段，切忌盲目在其上叠加 AI，正确的路径必须是“先理顺知识库，再部署 AI”。eGain 的核心洞见指出，企业内部目前有 **80% 的知识对 AI 而言是不可见的**，因为它们从一开始就不是按照 AI 可读取、可理解的格式构建的。目前受访企业中约有 **80% 处于低成熟度状态**，仅有大约 **4%** 达到了最高阶的变革阶段。 ### [18:54](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=1134s) eGain AI知识方法论与系统实操演示 为了帮助企业跨越成熟度鸿沟，eGain 推出了“AI 知识方法论”（eGain AI Knowledge Method）。该方法论的核心在于**需求驱动**——不再盲目整理企业“拥有什么知识”，而是聚焦于用户“正在寻找什么知识”。 系统直接从真实的客户对话记录切入，利用自动化工具提取高频问题，并从价值（Value）、体量（Volume）和复杂度（Complexity）三个维度进行综合打分，筛选出至关重要的“明星问题”（Star Questions，例如将各种不同表述统一归类为“我该如何改签现有航班？”）。系统将这些核心问题与后端知识进行匹配与合成，随后结合合规严格度与个性化规则完成发布。借助这套 AI 自动化流程： - 知识梳理与构建速度提升了 **90%**； - 方案价值落地周期（Time to Value）缩短至原来的 **1/3**； - 知识的一致性、准确性、合规度与相关性均得到质的提升。 在最后的平台实操演示中，演讲者展示了系统的工作流：管理员批量导入客户交互文本数据并配置元数据后，后台任务自动启动分析，前端迅速生成一张直观的“问题相似度与主题聚类图谱”。运营人员可以在图谱上直观叠加“业务复杂度”与“合规严格度”两重维度——根据问题所属的象限区域，系统会自动判定该类问题应当交由何种 AI 模型处理，或者直接触发人工审批与升级流，从而在确保绝对合规的前提下实现全流程智能化。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=4s) 末端的安全护栏根本无法修复源头缺乏治理所造成的错误 — 约翰·科普兰 > [03:53](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=233s) 问题不在于企业没有正确答案，而在于正确的知识从未被正确地连通与分发。 — 约翰·科普兰 > [05:14](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=314s) 模型本身无法辨别内容的新旧，如果文档未标注有效时间，AI 便会堂而皇之地检索两年前的作废规定并呈递给客户。 — 约翰·科普兰 > [05:14](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=314s) 构建任何模型的前提永远是让数据准备就绪；如果输入端是一堆垃圾，整个 AI 项目必然面临被高层砍掉的厄运。 — 约翰·科普兰 > [11:56](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=716s) 我的心与所有知识管理人员同在（My heart goes out to knowledge managers）。 — 约翰·科普兰 > [13:09](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=789s) 关掉 AI，不让 AI 回答该问题 — 约翰·科普兰 > [16:20](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=980s) 先理顺知识库，再部署 AI — 约翰·科普兰 > [16:20](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=980s) 企业内部目前有 80% 的知识对 AI 而言是不可见的，因为它们从一开始就不是按照 AI 可读取、可理解的格式构建的。 — 约翰·科普兰 > [18:54](https://www.youtube.com/watch?v=wWsLaa6zfvY&t=1134s) 不再盲目整理企业“拥有什么知识”，而是聚焦于用户“正在寻找什么知识”。 — 约翰·科普兰 ## Mentioned 约翰·科普兰（John Copeland）, eGain, Gartner, Gartner 客户服务知识管理系统魔力象限（Magic Quadrant for Customer Service Knowledge Management Systems）, SharePoint, Confluence, Claude, DevOps, RAG（检索增强生成）, eGain AI 知识方法论（eGain AI Knowledge Method） ## Video description John Copeland - VP of Marketing, eGain The rush to deploy AI in customer experience has surfaced an uncomfortable truth: most organizations are pointing powerful models at organizational knowledge that is fragmented, inconsistent, and ungoverned. But direct a brilliant model to broken knowledge and it only leads to disaster. Across self-service, agent assist, and automation, teams are learning that AI does not fix bad knowledge, it amplifies it. Conflicting policies, stale procedures, duplicate content, and undocumented expertise turn into wrong answers to customers, broken automations, compliance exposure, and a steady erosion of trust in both the AI and the brand behind it. This session examines how leading CX organizations are building a trusted instruction layer for enterprise AI: the foundation that lets these systems deliver instant, trusted answers across every channel. Drawing on lessons from hyper-scale AI deployments, we will look at how to move beyond traditional knowledge management toward AI-ready infrastructure that reliably powers self-service, agent assist, and agentic automation, so customers get answers they can rely on and AI earns the trust it needs to scale. [Source](https://www.youtube.com/watch?v=wWsLaa6zfvY)

[0:04] 约翰·科普兰: 末端的安全护栏根本无法修复源头缺乏治理所造成的错误 [3:53] 约翰·科普兰: 问题不在于企业没有正确答案，而在于正确的知识从未被正确地连通与分发。 [5:14] 约翰·科普兰: 模型本身无法辨别内容的新旧，如果文档未标注有效时间，AI 便会堂而皇之地检索两年前的作废规定并呈递给客户。 [5:14] 约翰·科普兰: 构建任何模型的前提永远是让数据准备就绪；如果输入端是一堆垃圾，整个 AI 项目必然面临被高层砍掉的厄运。 [11:56] 约翰·科普兰: 我的心与所有知识管理人员同在（My heart goes out to knowledge managers）。 [13:09] 约翰·科普兰: 关掉 AI，不让 AI 回答该问题 [16:20] 约翰·科普兰: 先理顺知识库，再部署 AI [16:20] 约翰·科普兰: 企业内部目前有 80% 的知识对 AI 而言是不可见的，因为它们从一开始就不是按照 AI 可读取、可理解的格式构建的。 [18:54] 约翰·科普兰: 不再盲目整理企业“拥有什么知识”，而是聚焦于用户“正在寻找什么知识”。

[0:04] John Copeland: garbage in, trusted out John Copeland: AI does not fix bad knowledge, it amplifies it. [0:04] John Copeland: leaderist [1:43] John Copeland: AI knowledge ops [3:53] John Copeland: spaghetti chart [3:53] John Copeland: customer zero [16:20] John Copeland: get the knowledge base correct first, then put the AI on top of it. [16:20] John Copeland: star questions [19:54] John Copeland: 1/3 the time [19:54] John Copeland: how do I reschedule my existing flight?

[0:04] Want to thank you for hanging out so long  at the getting close to the end of the day. [0:09] Yes, we're going to talk about garbage in and  trusted out and the importance of knowledge [0:14] and knowledge management for AI success. I  am John Copeland, VP of marketing at eGain. [0:20] You probably have never heard of eGain. If  you have, that's great. But if you haven't, [0:24] it would be normal to ask who the heck is  eGain. This is eGain. Right? We are a leader [0:31] in the inaugural Gartner Magic Quadrant for  customer service knowledge management systems. [0:38] I like to say the leaderist, even though that's  not a thing. Uh highest on ability to execute [0:46] furthest on completeness of vision.  This just came out uh a few weeks ago, [0:50] so it's somewhat hot off the presses. We're very  proud. The company's actually been in business [0:56] in knowledge management for 27 years, right? So,  kind of expert at the category. And leveraging AI [1:04] is sort of the latest greatest technology in that  application, and it's made a world of difference. [1:11] So, we're going to talk a little bit about  that. And you might say, "Well, you know, [1:14] why does knowledge management matter in customer  service, particularly in the context of AI?" [1:21] Because AI fails due to content pipeline  fragility, right? There there are a number [1:28] of factors that that go on that actually  cause the knowledge that's being read by AI, [1:35] leveraged by AI, to make a bad CX, incorrect CX  uh kind of experience. The first one is that, [1:43] you know, ungoverned knowledge basically  creates a very weak foundation. [1:48] Uh how many of you have multiple SharePoint's in  operation at your organization? Multiple versions [1:55] of the same kind of content authored by different  people. Nobody's kind of overseeing it at all. [2:02] That's that sort of ungoverned process that  creates a lot of issues. Secondly, but related, [2:07] there's no QA layer at the source. And the source  matters, right? Before you go sending an answer [2:15] off to an AI bot or to a customer service agent,  you want to know it's the right one. But there's [2:20] been no often no QA applied before the answers  are are pushed to the experiences that are being [2:27] delivered. Third, a lot of organizations treat  knowledge management as like a one-time cleanup [2:34] exercise. It's actually a discipline. It's an  operation. At eGain we call it AI knowledge ops. [2:42] You could think about it as a parallel to  DevOps. There is a discipline to this, right? And [2:48] if you're going to be really effective at it, it  has to be an ongoing kind of always-on process. [2:55] Finally, uh a lot of companies try to establish  guardrails, AI guardrails, at the very end of the [3:02] process. Guardrails at the end will not fix errors  at the source, lack of guardrails at the source. [3:10] We're going to talk a little bit about the kinds  of errors that we see in the implementations that [3:15] we've been doing across our pretty broad customer  base, very large enterprise. Um we're actually [3:22] very highly rated in uh regulated markets. So,  if you're a bank or you're an insurance agent [3:29] or you're in healthcare or you're a government  agency, you're often talking to us because errors [3:35] come with costs. Not just a bad customer  experience, but you may be legally liable [3:41] for providing an incorrect answer in a regulated  market. So, you want to make sure that the [3:46] guardrails are applied at the knowledge source,  which is kind of the beginning of the process. [3:53] That's lovely. Let me show you what we see when we  start an implementation in a lot of customers. We [3:59] lovingly refer to this as our spaghetti chart.  It looks like spaghetti, right? All those [4:06] wires going everywhere. What you're ultimately  seeing is multiple knowledge content sources, [4:12] right? You have your website, your PDF, your  SharePoint, your Confluence, or other sources, [4:19] your CRMs. They're often connected into different  kinds of retrieval systems, right? A search admin, [4:26] a homegrown knowledge management system. Um  those question marks and X's that you see [4:33] are actually the representation of  somebody getting a wrong answer, [4:38] right? Because they were probably looking  in the wrong spot, getting it from the wrong [4:43] uh source material, and as a result the  answer itself is wrong. And then it propagates [4:49] all the way through to the final channel in  which the answer is served. So, on the website, [4:55] over a mobile device, over the phone with the  call agents, uh digital agents on your website, [5:01] for example, and field staff. They  get the wrong answer to the question, [5:06] not because it the right answer wasn't there,  but because it's not appropriately connected. [5:13] So, this leads to the kind of  content foundation problems we see. [5:17] One of the biggest is duplication, right? This  used to be an issue with uh my team in some of the [5:23] places I used to work. It still is a little bit  of an issue at eGain uh to be honest. You know, we [5:27] want to be customer zero, but our practice could  use some improvements, some muscles of our own. [5:33] So, you get duplication, right? You get multiple  versions of the same article in different places, [5:39] right? Uh in separate systems. So, the  the rag implementation is basically, [5:45] in the absence of any guardrails or guidance,  it's going to kind of average over them. [5:50] Well, I'm going to take a little bit from this  and a little bit from this. They all seem to be [5:53] relevant. Nobody's telling me otherwise, right?  There's no instruction for AI in terms of how [5:59] it reasons. And so, as a result, you will get  inconsistent answers to the very same question [6:05] in the same channel and in different channels. The  second one is staleness. When was the last time [6:10] you went through and cleaned up your SharePoint  repository and got rid of things that could have [6:15] been gotten rid of a year ago or even longer. I  found a piece of eGain content actually today. [6:22] I I didn't find it. Somebody else found it. It's  on our website. Give you $5 if you can find out [6:27] where it is. Um and I will be here through the end  of uh Thursday. Um I will give you a hint. It has [6:34] something to do with personnel. Um but there  is a an error there and when stuff is stale, [6:41] right? AI is going to retrieve older stuff  not knowing that it's stale. There may not [6:46] be any date on it that says this was from 2 years  ago and there's another one that's more recent. [6:51] Right? And so, you will get AI constantly  presenting an answer that is ultimately [6:57] wrong. I get that often on Claude. I actually have  to give it a bunch of prompts around what it has [7:03] to do to make sure stuff isn't stale um and stuff  is accurate. It's It's not just making things up. [7:10] Coverage gaps are another one. So, there may be a  whole body a line of questioning that people are [7:16] interested in or demands from the employees.  They're interested in knowing certain things [7:21] and there's nothing there on it, right? We do see  that sometimes. And in that particular case, what [7:27] AI will tend to do is it will look for things that  are as relevant to the question as it possibly can [7:33] find even though it's not accurate and that's  when you tend to get either generic answers [7:39] or what we would call hallucinated answers.  the model's not wrong. It's doing what it's [7:45] designed to do. It doesn't have access to the  information it needs to provide the right answer [7:51] because it's a coverage gap. You don't know  that it's an area that you're not covering. [7:56] Number four, taxonomy inconsistencies. So,  when you're in a situation where different [8:01] teams are authoring on the same topics, on the  same content, storing it in different places, [8:08] they may be using different kinds of metadata. For  example, you might call in one team they'll say, [8:14] you know, sensitive information. Another might  say restricted information. Another might say [8:19] confidential information. You know, they To us,  they all mean the same thing. To AI, they don't [8:25] necessarily. And when you get that different  use of metadata, those different taxonomies, [8:30] you run the risk of actually giving access to  information to people and groups of people, [8:36] role-based access and filters that they  shouldn't see and potentially exposing [8:42] sensitive content to the wrong audiences. The  last one, ambiguous authorship. If you don't [8:49] know if there are two pieces of content on the  same topic and you don't know who authored which, [8:56] you don't necessarily know which one you  should trust more, which one you would lean on, [9:01] especially if they are conflicting, right? The  conflicting answers. Which actually isn't on here, [9:07] but that's another one that we see. Two or  three or more pieces of content that say [9:12] different things about the same topic. One says  go left, one says go right. In a regulated market [9:18] like health insurance, for example, if you've  ever had the frustration of saying, you know, [9:22] does my policy cover this? And somebody says,  yes. And then you look on the website and it says, [9:28] no. And then the AI agent says something  else. That is a problem of inconsistency. [9:37] Now, [9:39] predicts that throughout this year [9:42] organizations will drop 60% of AI projects that  aren't supported by AI-ready data. Data that [9:51] basically succumbs to one or more of these  issues. So if you haven't fixed those up [9:57] and you are in the midst of deploying or working  on an AI project, if the data is not ready there [10:02] is a better than chance opportunity or likelihood  that your project will get dropped. And in this [10:08] day and age, I mean from what we've seen from  customers and companies that we work with that's [10:13] not a good place to be in, right? I empathize  because as a former leader of data science team, [10:20] we used to have to have our data ready before we  could do, you know, um different kinds of models. [10:26] So it's about the data as a starting point. And  in AI, if your data is not ready you might get [10:32] dropped. So how do you build that infrastructure  layer? Um and the knowledge instruction layer. [10:42] The spaghetti chart after you go through sort of a  systematic process would look something more like [10:46] this, right? So you've connected I actually  heard in the prior panel discussion the idea [10:51] of the notion of you should connect to all of  those enterprise data sources. If they contain [10:56] information that's relevant in this case to  customer service, customer experience, they need [11:02] to be connected into that knowledge source. So uh  once you do that and you you can see we actually [11:08] call that sourcing, right? So you source to the  content, you bring it together, you synthesize it, [11:13] you curate it, and we're going to go through kind  of what that process looks like in a second. Um [11:19] and you involve knowledge authors, so those are  typically subject matter experts in a particular [11:24] domain, plus knowledge managers who are working  on how does this information, how should it be [11:29] organized? And then you store it in a knowledge  hub, right? The the one source, the single source [11:35] of truth for the answer to a specific question,  looking across the set of questions. Once you have [11:44] that, you can start to work on personalization.  Who gets to see which information? Right? So, [11:50] there are rules set up for that, and then you can  publish to all of the channels you need to publish [11:56] to. So, whether somebody's calling in on the  phone, whether they're looking up on the website, [12:01] whether they're using you're talking to a  phone agent or working with a digital agent, [12:05] they will get the same answer to the question.  What happens when that happens? Uh the first time [12:11] or first contact first call resolution rates go  way up. You get the answer right the first time. [12:17] Average handle time, this is what we've seen in  our customers, goes down because you can work [12:22] through customer questions quickly. And, you know,  if you're the customer, if you get your answer [12:28] right on the first try and you get it quickly,  you tend to be happier, right? So, our customers [12:35] NPS scores, for example, tend to go up  uh when they start to implement this. [12:42] Now, traditionally, prior to a lot of  technology and a lot of automation, [12:46] knowledge management has been a really heavy lift.  All of what we were talking about has been done [12:52] manually. My heart goes out to knowledge  managers just for the load of information I mean, [12:57] a typical enterprise can I mean, I was going  to say, can you imagine? If this is your job, [13:01] you can imagine, right? You live through it. Um  it's been all very manual, but that changes now. [13:09] Right? You should think about managing  knowledge as a life cycle, right? And [13:17] even though this is drawn as a linear diagram,  a linear process, it is actually recursive. [13:24] It continues, right? You hopefully source into  all the systems at the very beginning and you [13:29] don't need to continue to source, right? Anywhere  content would show up, you have plugged into. [13:35] But then you start working on the curation and  the synthesis like I talked about, right? What's [13:39] the structure? What's the taxonomy? What's the  metadata that we want to use? And then you start [13:44] thinking about managing all of that content and  governing all that content. What's old? Should we [13:50] get Well, I'll get rid of the old stuff if you've  got new stuff that I should be using in its place. [13:54] What are the knowledge workflows you need to be  building and managing over time? And how do you [14:01] prepare for and get ready to respond to compliance  and audits, right? Our Our product, as an example, [14:08] provides audit trails. Um we make sure that the  knowledge that's there is compliant with whatever [14:15] regulatory uh requirements might be there or  at least we signal to the knowledge managers [14:21] that content is either in or out of compliance  with, you know, particular procedures. [14:28] Then the personalization and publish. Where is  the data going? Who needs to see it? What are [14:33] they allowed to see? What are they not allowed to  see? When do you say, "You know what? I'm going to [14:37] turn off AI. I'm not going to let AI answer that  question." It is serious enough that it needs to [14:43] be escalated to a human, right? As an example.  And then finally, and you know, my career uh [14:50] for a long time has been in analytics, so I love  this last step. Feedback and analytics, right? So, [14:56] how are the KPIs for the business impacted?  Are we seeing faster uh call times, right? Uh [15:04] average handle time going down. Are we seeing  increases in first call resolution? Are we seeing [15:09] impact on customer success? I will tell you we  have one customer um in the consumer packaged [15:16] goods industry. It's a a global manufacturer.  Their old uh cost to handle a customer call [15:24] uh uh balanced across all the channels was about  $10 per call. Right? After they applied this uh [15:33] way of working and these tools that uh that we're  talking about, they reduced the average customer [15:39] call cost from $10 down to 15 cents. Because a  lot of the effort uh has been automated and a lot [15:48] of the um the first calls are resolved correctly  the first time. Right? So, the calls are shorter. [15:56] Uh they can automate a bunch of them. So,  they're driving those weighted costs down. [16:00] And when the customer presented that at  our conference in London a few months ago, [16:05] um he made a point of saying, "These numbers are  verified by finance." Right? So, uh it's it's [16:12] really nice when the CFO then kind of is on your  side for this operation that you're trying to do. [16:20] Now, as you begin to apply these uh principles,  for example, uh you will end up kind of at some [16:27] point in our knowledge maturity scale. I apologize  cuz the colors, I don't think, come through as [16:32] nice as I'd like them to. Um but there are Every  column here has some blue, which is good, and some [16:41] pink, which is a problem. So, usually when you're  starting out the first time, you're kind of in the [16:46] ad hoc uh knowledge maturity point. You have um  siloed content, right? There's no formal process [16:55] for capturing it and curating it, and there's no  consistent taxonomy. You're living in, you know, [17:00] a really hard time. Uh you are not ready for AI.  If that is your situation, don't throw AI on top [17:07] of it. In fact, we have a number of customers  who would say the the way to do this right [17:14] is get the knowledge base correct first, then  put the AI on top of it. Don't do it in reverse [17:20] order cuz it it's it's a really big hassle. When  you advance to sort of the more managed stage, [17:27] what that tends to align with is you've got a  single source of truth now. Right? You've got [17:32] one place where answers to key questions live,  but you don't have any kind of formal authoring [17:40] uh guide. So, when new knowledge comes to bear,  it can look like a bunch of different things, [17:45] right? Um one of the things we say at eGain  is 80% of the knowledge in your systems [17:51] is invisible to AI because it's not crafted in a  way that AI can read it and understand what it is. [17:58] So, one of the things you want to  do is you, you know, mature more, [18:02] is use that managed approach. I'm sorry, use the  um uh the formal approach. When you're designed, [18:08] and I'm I'm sorry, I'm speeding up a little bit  cuz I see the time ticking down. When you're [18:12] designed, you've got the single source of truth,  you've got a consistent structure and metadata, [18:17] but you still aren't connected to enterprise  systems uh the way you would want to push it out. [18:22] And you don't have contextual personalization. For  optimized, uh you actually bring in the context, [18:29] so now you know more about the context of the  question coming in. And when it's transformative, [18:34] that's when you're really ready to use  knowledge to drive gen AI applications. [18:41] Now, we tend to see the companies that we  talk to about 80% of them are at low maturity. [18:47] Right? Only about 4% are off  on that far right-hand side. [18:54] But there's a method for moving up. We call it  the eGain AI knowledge method. It's helping you [19:00] understand the knowledge you need, not necessarily  the knowledge you have. Where does that [19:06] understanding come from? It comes from looking at  what people are looking for. Where is the demand [19:12] of knowledge? So, we actually start by looking at  conversations. Uh the conversations customers have [19:18] uh with different people. What questions  are they asking? We score those questions [19:23] uh on value, volume, and complexity. The ones that  are very high, we call star questions. They're [19:29] very important questions to get right. We match  the questions with the knowledge in your systems [19:35] and bring it together, the synthesis, the  curation, um and then add the personalization [19:40] on the back end before we then go to publish.  This is our approach, the eGain knowledge method. [19:46] Because we're applying automation and AI, we can  now run through that process 90% faster than when [19:54] knowledge management was fully manual. And you get  to value from that implementation in 1/3 the time. [20:02] And you get improved consistency, accuracy,  compliance, and relevance. And it starts [20:07] with a question. These questions on the left  really sort of resolve to how do I reschedule [20:14] my existing flight? Right? Examples, a bunch of  a bunch of different ways that are all really the [20:21] same question. Well, the thing that people want to  know determine the knowledge you need. Right? So, [20:27] we definitely need to be able to figure out how  to answer changing my flight. So, we connect to [20:32] the conversation stores. I'm going to show you a  very short little video here in a second of how [20:37] that works on our platform. We extract questions  from the conversation. We group similar questions, [20:43] and then we do volume, complexity analysis. And  we need to look at the AI that we apply, right? [20:49] Is it complex? Is the query complex or simple?  Is it very strict from a compliance standpoint [20:55] or less strict? So, very quickly, this is let's  see us. This is how it works in our system. [21:03] So, you connect to a store of questions. [21:08] You're going to see it here. They're  going to drop in a set of transcripts. [21:16] Right, you provide a little  metadata. You start the job. [21:22] The job is going to be running across the top. [21:26] Right, it's going to continue to run, but in  the demo, we're showing you what you kind of get [21:29] for the output. That lovely graph right there is  basically a graph looking at questions similarity, [21:35] topic similarity. And what we can do is  overlay, working with you, we overlay [21:40] the level of complexity and the level  of strictness, right? Does it require [21:45] compliance? And then the AI will be a function of  where a cluster of questions lands in this graph. [21:53] You can also see the questions on the left. They  get assembled into topics and subtopics that [22:00] you can look through. And you can find out which  questions have All right. Old data. Sorry. Yeah, [22:06] no, we've come down to the end. Uh, don't mean to  cut you off like that, but it's like everything's [22:10] set on timers here. Thank you, everybody. Thank  you so much for coming out. Thank you, sir.