Deepgram, RingCentral Cut Support Handle Times 60% With AI

Deepgram与RingCentral利用AI将客服处理时长缩短60% 部署智能体AI大幅削减了联络中心的常规工作流与成本，但也让留任员工因接连不断的复杂来电而不堪重负，并使企业直接面临聊天机器人虚假陈述的法律责任。 专家小组详细介绍了AI在客户支持与服务工作流中的实际企业落地应用，重点介绍了Deepgram的自主Tier 1支持智能体以及Franklin Templeton的统一智能中心等解决方案。与会嘉宾提出了结构化的“自动化、辅助、分析”（Automate, Assist, Analyze）框架，推动企业从被动救火转向主动异常检测及全量对话智能。尤为关键的是，讨论强调AI部署必须建立在稳固的数据治理之上，而非混乱的遗留基础设施，并引用了加拿大航空（Air Canada）聊天机器人裁决等具有警示意义的法律先例。各方达成共识，认为AI在快速信息综合和常规问题解决方面表现出色，但在处理涉及复杂情感的互动时，必须配合无缝且具备丰富上下文的人工交接。 企业必须在以客户为中心的AI自动化与公司盈利能力（EBIT）之间取得平衡，避免在缺乏清晰、可自负盈亏的ROI路线图的情况下为了AI而部署AI。 (0:04) Deepgram在人工团队旁引入了自主“智能应用工程师”（Agentic Applied Engineer），在处理大量Tier 1技术咨询的同时，为复杂企业客户保留专属的工程师支持。 (5:28) 通过AI对电话系统进行现代化升级，使平均处理时间缩短了60%以上，通话后整理时间从7分多钟降至2.5分钟，同时电话通信开销削减了50%。 (7:34) 联络中心常规通话的自动化带来了认知负荷困境，人工客服人员不得不连续处理大量高压复杂问题，缺乏自然的认知调剂间隙。 (7:34) RingCentral将AI客户旅程划分为三个阶段：前期交互自动化、通过实时副驾驶辅助在线人工客服，以及在通话后对100%的对话展开全量分析，而非传统的10%至15%抽样。 (12:03)

Deploying artificial intelligence in customer experience delivers measurable cost and capacity gains, but panelists stress that sustained ROI depends on strict error governance, proactive issue resolution, and deliberate human escalation paths. Operational Gains and Agent Workload Pressures (5:28) Modernizing telephony and tier-one support with AI-first systems produces steep efficiency improvements. Heather Gillbanks of CXPA highlighted an enterprise call center deployment where an AI telephony implementation reduced telephony infrastructure costs by 50%, cut post-call wrap-up times from over seven minutes to 2.5 minutes, and lowered average handle time by more than 60%. Similarly, Deepgram rolled out an autonomous Agentic Applied Engineer to absorb surge volumes of basic technical inquiries across thousands of developer accounts, freeing human engineers to handle complex enterprise implementations. However, stripping routine tasks introduces a notable operational strain. Gillbanks and Erin Fusaro observed that eliminating simple inquiries left human staff with an unrelenting queue of cognitively demanding issues, which overwhelmed average performers. To prevent burnout and fear of automation, organizations must accompany AI implementations with intentional upskilling programs that prepare frontline workers for higher-complexity interactions. Proactive Remediation and End-to-End Orchestration (12:03) Panelists emphasized shifting customer service from reactive ticketing to predictive, end-to-end resolution. Amir Hameed outlined RingCentral’s framework covering upfront automation (handling basic FAQs), in-flight assistance (providing context-rich handoffs and co-pilots), and comprehensive post-call analysis that evaluates 100% of interactions—contrasting with legacy quality management that sampled only 10% to 15%. This exhaustive coverage helps identify churn patterns and operational gold nuggets from top reps. Predictive analytics also enables silent issue remediation before users notice an error. Deepgram runs continuous telemetry sweeps that trigger autonomous sub-agents to resolve regional latency spikes or transcription degradation without human tickets. In enterprise asset management, Franklin Templeton uses unified advisor hubs to analyze customer history and market data proactively, surfacing product recommendations and routing high-value clients directly to senior portfolio managers. Handoff Boundaries, Messy Data, and Trust Governance (23:07) Panelists warned that AI cannot simply be applied on top of chaotic data lakes or obsolete legacy architecture. Chasing tools without clear business discovery risks repeating the big data era's mistake of dumping unstructured information into data lakes and producing unusable outputs. Successful modernization requires focusing first on bounded, high-impact areas with relatively clean data, such as modern telephony records, rather than immediately trying to overhaul decades-old core back-office systems. Clear guardrails at the boundary between bots and humans are equally vital for customer trust. Panelists agreed on an immediate override rule: when a user explicitly requests an agent, the system must hand off the call with complete conversational context. The penalty for unconstrained automation is steep, as demonstrated by the Air Canada court ruling that held the airline legally liable for bereavement policy misrepresentations made by its chatbot. Because customer experience research indicates it takes 11 consecutive positive touchpoints to recover goodwill from a single severe mistake, critical policy and pricing decisions require hardcoded limits and auditability. Balancing Retentive Metrics and Real-Time Sentiment (37:15) Long-term AI success requires tracking traditional contact center benchmarks—such as containment rate, first contact resolution, and repeat call frequency—against baseline Net Promoter Scores to confirm vendor performance claims. In consumer scenarios like gaming, automated credit distributions immediately following a technical glitch preserve active session time, mitigating costly customer acquisition expenses. In institutional financial settings, expanding advisory coverage through AI platforms deepens relationship sentiment and capital retention. Sentiment monitoring has evolved from retroactive survey tagging into dynamic real-time evaluation. Modern models track vocal distress during live interactions to trigger immediate human supervisor escalations and automated follow-ups. Panelists concluded that AI agents do not inherently replace human agents; AI excels at rapid information retrieval across disparate databases, while human representatives remain indispensable for authentic empathy and high-stakes problem resolution. While AI automation provides immediate call center savings and scalable coverage, the panel concluded that its effectiveness hinges on clean data inputs, strict factual guardrails to prevent legal liability, and seamless escalations to human agents when genuine empathy is required.

Ai4研讨会上，来自PadSplit、Deepgram、RingCentral及富兰克林邓普顿等机构的技术与客户体验负责人指出，人工智能正在将客户服务从被动的故障修复转变为全流程主动干预，但企业必须在自动化提效、人工兜底保障以及脏数据与法律合规风险之间维持务实平衡。 降本提效背后的认知负荷与规模化支持 (7:34) 在呼叫中心与工程支持等一线场景中，部署AI代理带来了显著的效率突破。Deepgram现场工程高级经理Kiley Williams介绍，其团队开发的“智能应用工程师”（Agentic Applied Engineer）作为自主协作伙伴处理一级支持，解决了向所有开发者开放支持后工单暴增的问题，让人工工程师专注于复杂企业架构。在传统呼叫中心改造中，CXPA的Heather Gillbanks指出，向AI优先的电话架构转型不仅削减了50%的基础设施支出，更将平均处理时长（AHT）降低超过60%，并将通话后整理时间从7分钟以上压缩至2.5分钟，大幅精简了离岸外包坐席。 然而，这种效率提升也带来了未曾预料的操作困境。Gillbanks与PadSplit的Erin Fusaro指出，当常规、琐碎的咨询被AI完全过滤后，留给一线人工客服的全是高度复杂、情绪紧绷的棘手问题，导致表现一般的坐席承受持续的认知重负。Fusaro强调，企业必须配合技能升级培训，帮助员工适应更高价值但更具挑战性的工作，以消除对自动化的恐惧。 全旅程智能化：自动化前置与数据全覆盖 (10:03) 针对端到端的客户交互，RingCentral全球技术解决方案高级副总裁Amir Hameed提出了“自动化、辅助、分析”（Automate, Assist, Analyze）的三阶段框架。前端通过AI接待员和数字代理过滤日常高频咨询；中端在向人工坐席转接时传递完整上下文，并提供实时辅助驾驶与主管看板；后端则彻底改变了传统质量管理仅能抽样10%至15%录音的局限，实现对语音及数字渠道100%对话的转录、评分和归类。管理者借此能够精准排查区域性流失和话术漏洞，并提炼顶尖员工的成功模式进行全员推广。 富兰克林邓普顿人工智能采用与解决方案主管Lily Li展示了针对财富管理条线落地的“情报中心”（Intelligence Hub）。该平台整合了此前分散在15个系统中的市场洞察、客户记录与交易报告，涵盖区域管理、顾问拜访准备与会后跟进。李指出，与未采用该系统的团队对比，该平台带来了更广的客户覆盖面、更深入的互动，并直接推动了可量化的销售额提升。 预测性运维与多模态主动干预 (16:24) 专家强调，卓越的客户体验在于客户意识到问题之前便予以解决。Williams解释，预测性分析的本质在于利用系统遥测数据实现从被动应对向主动修复转变。Deepgram通过自动化监控代理对海量实时语音数据进行全天候巡检，一旦发现局部延迟激增或特定行业转录质量下降，便会自动派生子代理排查根因，在终端用户提交工单前解决基础设施异常。 在投资与资产管理等高净值领域，富兰克林邓普顿结合了传统预测模型、生成式AI与智能体架构，根据市场波动和过往偏好向客户主动推送投资组合优化建议。针对不同层级的客群，AI能够根据记录系统实现分流：普通客户获得上下文连贯的自助流，而高净值客户则跳过常规排队直接转接至高级投资经理。这种主动服务缩短了解决路径，但也需严格把握沟通分寸，避免主动提醒反而引发客户不必要的焦虑。 人机边界设定与内部员工赋能 (23:07) 在自动化流程设计中，界定AI与人类代表的交接边界至关重要。研讨会提出了一项硬性规则：一旦客户明确说出“人工客服”（agent）等关键词，系统必须立即无缝转接人工，彻底规避繁琐的IVR菜单陷阱。虽然现代语音AI在语调、语速和口音上已有大幅提升，且具备情绪分析能力，但数字代理依然无法替代真实的人类共情。 当AI穷尽解决手段时，向人工坐席的切换必须携带完整的对话上下文与身份验证信息，避免客户重复陈述。同时，AI不仅面向外部客户，还能充当一线员工的内部辅助工具（copilot）。通过实时提供数据与策略建议，内部AI可将单个客户经理的服务覆盖能力从100名客户扩大到1,000名。为确保合规与安全，企业必须设置清晰的“人机协同”（human-on-the-loop）监管防线，要求AI提供可追溯的引用来源，并设置严格的业务熔断机制。 脏数据陷阱、商誉重建与法律合规红线 (27:48) 企业推进现代化时，切忌将AI盲目嫁接到混乱的数据和几十年历史的遗留系统之上。多位嘉宾警告，不要重蹈早期大数据时代将Hadoop当成随意倾倒杂乱数据的“乔治·福尔曼烤盘”（George Foreman grill）覆辙，输入垃圾数据只会放大幻觉并产生管理废料。务实的路径不是等待百分之百完美的数据，而是先选择电话系统等数据相对干净的高价值场景切入，而非强行重构比员工年龄还大的老旧ERP系统。 在风险防范方面，CX研究显示，一次严重的体验故障需要后续连续11次正面交互才能将客户信任恢复到基线水平。加拿大航空（Air Canada）的判例成为全场的警示教材：该航司官网聊天机器人向旅客提供了错误的丧亲优惠退票政策，加拿大民事法庭最终裁定航空公司必须对其公开AI聊天机器人作出的不实陈述承担全部法律责任。该案例表明，在定价、退改政策和法律合规等高风险场景中，企业必须对AI施加硬编码限制与合规熔断，杜绝幻觉可能引发的法律与声誉危机。 研讨会总结指出，评估AI代理与人类坐席的高下取决于业务场景：AI在海量数据综合、高响应速度及预测预判上具备无可比拟的优势，但在需要深度倾听、情感连接和化解高阻力冲突时，人类代表依然不可或缺。成功的架构必须将二者深度融合，依托留存率、真实客户满意度等长期商业指标持续校准系统。

## Background 该内容由行业峰会频道 Ai4 于 2026 年 8 月 22 日发布在 YouTube 平台上。视频记录了一场由 PadSplit 首席技术官 Erin Fusaro 主持的高管圆桌研讨，汇聚了来自 Deepgram、RingCentral、CXPA 和富兰克林邓普顿的行业领袖，重点探讨 AI 如何通过自动化交互、个性化支持、主动式预测服务及人机协同平衡来全面重塑客户体验（CX）与客户服务运营。 ## TL;DR 研讨会通过具体业务案例深入剖析了 AI 在一线工程支持、全生命周期呼叫中心质检及机构级智能中枢中的落地成效，展示了通话处理时长大幅缩减超 60% 与全量质检赋能培训的实际收益。嘉宾们详细拆解了融合预测性 AI、生成式 AI 与智能体 AI 实现主动式服务和高价值客户智能路由的架构设计，并强调人机平滑交接时必须完整保留上下文。针对落地陷阱，专家们警示了老旧系统数据质量问题与 AI 幻觉带来的法律信任风险（如加航诉讼案），呼吁企业从小步快跑切入，建立严格的评估闭环与安全熔断机制。最终指出，AI 的高效数据处理与人类独有的共情力并非零和博弈，两者的有机融合才是驱动客户长期留存与商业底线增长的核心动能。 ## Atmosphere & style 整场圆桌研讨节奏紧凑、干货密集，洋溢着务实求真与深度对话的专业氛围。嘉宾阵容涵盖了底层语音技术先锋、万亿级资产管理巨头以及经验丰富的企业转型老手，彼此间观点碰撞既有犀利幽默的真实吐槽（如面对老旧系统和海量销售推销），也有严谨硬核的架构推演。整体叙事从一线降本增效逐步升华至组织风控与长期客户信任构建，既展现了对前沿技术变革的热情，又始终保持着对商业 ROI、数据泥潭与员工认知负荷的理性审慎。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=4s) 企业在推进 AI 落地时应追求真实的投资回报率，通过将旧系统迁移上云节省成本，再以节省的资金滚动投资智能体语音 AI，形成自给型路线图，避免为了搞 AI 而搞 AI。 - [05:28](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=328s) AI 优先方案可大幅降低呼叫中心的平均通话处理时长（AHT）与后处理时长（ACW），但也带来了剩余复杂疑难问题集中、对留存坐席认知负荷要求急剧上升的新挑战。 - [10:08](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=608s) 大型机构可通过构建统一的情报枢纽打通分散系统与多模态数据，从而在财富管理和分销业务中实现真实的销售增长与客户覆盖面扩张。 - [10:08](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=608s) 生成式对话 AI 能够实现对 100% 通话的全量结构化质检与分析，不仅能精准定位区域性客诉诱因，还能提炼顶尖坐席的话术金矿以复制优秀实践。 - [16:30](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=990s) 现代客户支持正从被动响应转向主动预判，依靠全天候自主智能体体系实时监测微弱异常信号并自动修复，赶在客户感知前扑灭隐患。 - [18:37](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=1117s) 主动式客户体验需要打通融合预测性 AI（模式识别）、生成式 AI（交互生成）与智能体 AI（自主执行）等不同代际的技术体系。 - [20:41](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=1241s) 基于 360 度客户全景视图与预测性洞察，AI 能在后台自动识别客户层级，将高净值 VIP 客户无缝直通给资深投资经理，传递专属服务价值。 - [23:15](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=1395s) 人机协同必须设立清晰的边界与无条件转接机制，并在转接给人工坐席时必须完整保留上下文信息，坚决杜绝客户重复复述问题。 - [27:48](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=1668s) 面对老旧系统与混乱数据，企业切忌盲目接入 AI 放大混乱，也不应等待完美数据，而应锚定业务核心价值并从小步快跑的高杠杆局部切入改造。 - [34:23](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=2063s) 加航 Bot 虚假承诺诉讼案警示企业，无论信息是由真人还是 Chatbot 陈述均具有法律约束力，企业必须在核心规则与敏感业务节点严密杜绝 AI 幻觉。 - [38:27](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=2307s) 衡量 AI 对客户留存的贡献需要结合传统呼叫中心 KPI、重复提问率以及部署前后的基准满意度（NPS/CSAT）开展横向对照测试与闭环微调。 - [41:14](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=2474s) 大模型驱动的实时情绪分析告别了传统人工手动打标签模式，能够在通话中实时捕捉情绪波动并在触发流失导火索时自动发起主动外呼挽回。 - [42:25](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=2545s) AI Agent 与人类坐席并非零和替代关系，AI 胜在效率与海量信息预判整合，人类胜在深度同理心与复杂共情，两者融合才能实现极致客户体验。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=4s) 嘉宾介绍与各方在 AI 浪潮下的首要关注点 研讨会由 PadSplit（一家合住房屋双边市场平台）的首席技术官 Erin Fusaro 担任主持。开场时，四位来自不同领域的嘉宾依次分享了自己的背景以及目前在 AI 领域最关注的核心议题。 Deepgram 现场工程高级工程经理 Kiley Williams 率先发言。他介绍 Deepgram 是一家基础语音 AI 平台，旨在赋能开发者将语音功能无缝集成到自身软件中。对他而言，当前重中之重是打造“卓越的客户体验”（Exceptional Customer Experience），其极致标准在于让客户在与平台及客户支持工程师团队交互时，产生完全无缝、难以分辨是人还是机器的高水准体验，并且无论客户处于何种层级或面临什么问题，都能感受到高度个性化的服务。 代表客户体验专业人员协会（CXPA）出席的 Heather Gillbanks 则带来了不同的视角。她刚刚参与编写了《客户体验知识体系》（CX Book of Knowledge），并在现场打趣表示可以赠送一本。与在座的技术专家不同，Heather 坦言自己并非 AI 专家，而是拥有超过 30 年经验的“企业转型老手”。她坚信无论是数字化转型还是客户体验转型，底层逻辑完全相通。作为一名“客户痴狂者”（customer fixated），她同时强调企业必须兼顾息税前利润（EBIT），在客户体验与盈利能力之间找到平衡。在实操层面，她目前最大的痛点是应对海量供应商的轰炸：“我每周能收到 30 到 40 个推销联络，每个人都想占用我 15 分钟时间”，如何从这些供应商中筛选出真正有价值的合作伙伴，是当下从业者极其重要的课题。 RingCentral 全球技术解决方案高级副总裁 Amir Hameed 随后介绍了公司的业务布局，涵盖内部通信云平台、呼叫中心及客户体验，目前核心聚焦于智能体语音 AI（Agentic Voice AI）。Amir 带领着一支全球专家团队，日常工作即是协助不同阶段的客户推进转型——从深陷传统旧系统的泥潭，到描绘未来上云的蓝图。他特别强调，AI 落地必须讲求真实的投资回报率（ROI），RingCentral 倡导“自给型路线图”（self-funded roadmaps）：先通过将旧平台迁移到云端为客户省出成本，再用节省下来的资金逐步引入 AI 解决方案，形成良性循环，坚决避免“为了搞 AI 而搞 AI”（AI for AI's sake）。 富兰克林邓普顿（Franklin Templeton）AI 采纳与解决方案主管 Lily Li 最后做自我介绍。作为管理着 1.8 万亿美元资产的受监管大型金融巨头，她的工作专注于企业级 AI 战略落地与解决方案开发。在客户体验维度，公司致力于将客户端、市场端、产品端与投资专业知识等多模态智能打通，使客户体验具备前瞻性（proactive）、持续性（persistent）、一致性（consistent）与高度个性化，并在全企业范围内实现可扩展的“机构级”（institutional grade）落地。 - **PadSplit**：专注于合住房屋（co-housing）的双边交易市场。 - **Deepgram**：基础语音 AI 平台，致力于让语音交互在软件中达到“真假难辨”的极致个性化水准。 - **CXPA 与转化平衡**：转型方法论在不同技术浪潮中保持一致，核心是在客户体验与公司 EBIT（息税前利润）之间找寻平衡；从业者每周面临 30–40 个供应商的会议邀约。 - **RingCentral**：主打自给型路线图（Self-funded Roadmap），先迁移上云创造收益，再用收益滚动投资智能体语音 AI。 - **富兰克林邓普顿**：资产管理规模达 1.8 万亿美元，聚焦构建机构级、高可扩展的多模态客户端智能体系。 --- ### [05:28](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=328s) 实战案例：从一线工程工单到呼叫中心降本增效 主持人 Erin 邀请各位嘉宾分享目前已经上线或正在构建的 CX 实际项目，探讨 AI 究竟解决了什么痛点并带来了哪些实质性改变。 Kiley Williams 详细介绍了 Deepgram 现场工程团队旗下的“应用工程”（Applied Engineering）实践。Deepgram 的一项核心特色是：任何客户寻求技术支持时，与其直接对话的都是真实工程师。然而，当团队近期将支持服务向包括单兵独立开发者在内的所有用户开放后，海量用户带来了严峻的规模化瓶颈。为此，团队从 12 月开始构建并部署了“智能体应用工程师”（Agentic Applied Engineer）。这名 AI 智能体被全团队视为真正的“数字同事”，在与人类工程师并肩作战的同时负责处理一线（Tier 1）工单。智能体每一次与人交互都在持续进化变聪明，这不仅没有牺牲服务质量，反而将人类工程师从琐碎应答中彻底解放出来，去深度服务那些真正需要复杂、高触感支持的客户。 现场唯一的企业客户代表 Heather Gillbanks 接过话茬，幽默地表示：“我敢说主办方邀请我就是因为这个，我是少数被允许公开谈论实际落地成果的客户。”她的案例极具冲击力：她此前所在的企业拥有庞大的呼叫中心，团队采用“AI 优先”的解决方案全面替换了旧有电话通信系统。最初向 CEO 汇报时，立项理由仅仅是相比上一代供应商能直接节省 50% 的话费账单，但在深入运用 AI 功能后，业务指标出现了惊人的质变： - **平均通话处理时长（AHT）**：直接下降了超过 60%。 - **通话后处理时长（ACW）**：从过去的 7 分多钟断崖式压缩至约 2.5 分钟。 这一突破为公司在海外离岸业务流程外包（BPO）合同续约时大幅削减了外包坐席编制，直接砍掉了巨额运营成本。然而，Heather 也提出了发人深省的警示：当 AI 拿掉了所有“杂活差活”之后，那些最顶尖的 A 类坐席能够接住这种认知负荷，为复杂疑难客户提供极佳服务；但那些水平不够顶尖的员工，却无法承受全天面对高难度问题的脑力负荷。企业在拿走简单任务时，必须慎重审视剩下的人员是否具备消化复杂工作的认知能力。 对此，主持人 Erin 补充指出，在组织内部全面推行 AI 时，员工往往潜藏着“如果事情都被自动化了，我们该去干嘛”的焦虑与恐慌。因此，企业必须将 AI 演进与员工技能重塑（upskilling）绑定，让大家看到 AI 接管枯燥机械劳动后，自己能够发挥专业特长去承担更有价值的新工作。 --- ### [10:08](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=608s) 机构级智能中枢与端到端全旅程 AI 覆盖 富兰克林邓普顿的 Lily Li 分享了面向大型机构的落地案例。除了常规的 AI 联络中心外，他们于今年 1 月针对全美财富管理团队推出了名为“情报枢纽”（Intelligence Hub）的机构分销平台。在过去，财富顾问必须在多达 15 个分散的系统和信息源之间来回切换；如今，他们只需在统一平台开启一天的工作，从区域管理、投资顾问互动，到事前与事后沟通、可落地洞察、交易报告以及重点客户排序全部整合在一处。Lily 笑称，系统有一次发生小故障时收到了大量投诉抱怨，这反而证明团队已经彻底改变了原有的工作模式并高度依赖该平台。更重要的是，通过与未使用的对照组进行严格基准测试，该系统带来的不仅是效率工具层面的生产力，而是真实的业务价值：实现了客户覆盖面的增量扩张、更多高附加值的互动以及潜在的销售业绩拉升。 RingCentral 的 Amir Hameed 随后剖析了他们所践行的全生命周期客户旅程 AI 赋能模型。他指出，不应孤立地看待 AI，而应将其拆解贯穿在整个客户互动流中： ```mermaid flowchart LR accTitle: RingCentral 端到端客户旅程 AI 架构 accDescription: 展示从前端接待自动化、通话中协同到通话后深度分析的全流程架构。 subgraph PreCall["前端接触"] A["AI 接线员 / 数字人"] -->|"自动化解答营业时间与基础服务"| B["情境上下文完整交接"] end subgraph InCall["通话中协同"] B --> C["高价值客户转接人工"] D["实时 AI 辅助引擎"] -->|"实时提取全量数据 / 销售与服务赋能"| C end subgraph PostCall["通话后闭环"] C --> E["100% 全量智能分析专家"] E --> F["识别区域性流失与问题模式"] E --> G["提炼顶尖坐席话术金矿用于培训"] end ``` 在通话前，由 AI 接线员自动处理营业时间、地点与基础服务咨询；当需要人工介入时，数字助手会携带完整的上下文信息无缝转交给人类坐席。通话过程中，AI 实时为坐席调取关键背景资料，赋能坐席展开更有价值的销售或服务沟通，主管端也能同步获得监控杠杆。 最核心的质变发生在通话后分析阶段。在传统质检模式下，由于人力有限，企业拼尽全力也只能抽检 10% 到 15% 的通话录音；而借助现在的生成式对话 AI 专家系统，企业能够对 100% 的通话进行深度、结构化分析。这打破了传统的宏观数据盲区——例如某些地区客户流失率虽然处于“可接受范围”，但通过全量语义分析，企业能直接揪出东北区或西南区特有的重复性客诉诱因。 主持人 Erin 与 Amir 进一步达成共识：AI 质检的精髓不仅在于“挑错”和救火，更在于挖掘“极致优秀的标杆是什么样”（What does great look like）。通过对海量对话打分，企业能够精准捕捉顶尖坐席打动客户的“话术金矿”，并将其规律化、复制化，用于全员培训体系中。 --- ### [16:30](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=990s) 走向主动式服务：利用预测分析防患于未然 随着底层技术的高速演进，客户对企业的期望水准被大幅推高，甚至在某种程度上期望企业能够“读心”，提供极致的前瞻性个性化服务。针对主持人 Erin 提出的“如何利用预测分析在客户投诉或求助前解决问题”，Kiley Williams 给出了深入剖析。 Kiley 认为，“预测分析”听起来是个时髦词汇，其本质是对多维数据的上下文关联与深度挖掘。作为基础语音 AI 平台，Deepgram 每天处理着海量的实时语音数据流。基于这种天然的数据汇聚，技术支持团队的核心指导理念由“被动响应”（reactive）全面转向“主动预判”（proactive）。 Kiley 提到，最好的客户体验就像自己使用的某些工具发来的提示：“我们监测到底层存在微小异常，已经替您自动修复完毕。”为了达到这种境界，团队依靠全天候运行的自主智能体（Agent）体系。相比人类坐席每天只能盯盘 8 小时，AI 智能体永不疲倦地实时梳理系统运行状态，一旦在某个特定地域、特定客户类型或特定垂直行业发现微弱异常信号，主智能体便会自动派生出子智能体（sub-agents）去深入排查诊断，赶在客户感知到异常或举手提单之前，就将隐患彻底扑灭在萌芽状态。 ### [18:37](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=1117s) 主动式客户体验与多代 AI 技术的融合 谈及异常检测与主动服务，嘉宾指出，卓越体验的核心在于走在客户前面：“终极的客户体验就是你根本不需要客户做任何事；一切都已经替你办妥、修复好了，你甚至根本察觉不到曾有过问题存在。”Lily Li 对此表示高度赞同，特别是在财富管理和金融服务领域，投资与市场专家长期依赖对市场动态、产品特征以及客户历史行为与偏好的洞察。借助技术手段，机构能够在客户意识到需求前，主动出击为其提供投资组合优化建议。 Lily 进一步强调，企业在布局主动服务时切忌画地为牢、局限于单一技术形态。从 20 世纪 90 年代延续至 2020 年左右的预测性 AI（Predictive AI）、2022 年随 ChatGPT 爆发的生成式 AI（Generative AI），再到近期兴起的智能体 AI（Agentic AI），这些不同代际的系统应当打通融合。企业应当借助数据资产中的全维度血缘关系与情报，综合覆盖业务流程中的上行机遇与下行风险，全面构筑主动式客户支持体系。 - **Predictive AI**：深耕数十年，擅长基于历史偏好和市场指标进行模式识别与需求预判。 - **Generative AI**：提供自然的理解与内容生成能力，革新了人机交互与客户触达界面。 - **Agentic AI**：多智能体协作驱动自主执行与复杂工作流编排，实现从“分析”到“行动”的跨越。 --- ### [20:41](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=1241s) 客户分群、预测性洞察与高净值分级服务 Amir Hameed 将话题拓展至个性化与客户全景视图（360-degree view）。在金融等高价值行业中，客户通常具有明确的分级体系。若前线坐席能够在统一的记录系统（System of Record）中无缝调阅客户资产、已有服务及预测模型给出的下一步推荐，不仅能打消客户“在通话中重复解释已有投资组合”的烦躁感，还能让他们切实感受到“自己的时间受到了尊重”。 更关键的是基于客户画像的服务路由分级。Amir 指出，AI 完全可以在幕后完成客户识别，并根据客户层级做出精细化调度。对于顶级高净值客户（High Rollers），系统甚至不该让他们接入普通坐席或陷入排队等待，而是应当直接将其无缝转接给资深投资组合经理。通过 AI 在幕后的精准画像与专属 VIP 路由，机构向客户传递出清晰的信号：“您用长期的业务赢得了这份专属地位，我们尊重并极力维系这份信任。” ```mermaid flowchart TD accTitle: 客户分级智能路由流程 accDescription: 展示客户来电后如何经由背景识别与分级判定分流至不同服务链路 subgraph Identification ["身份与数据识别"] A["客户发起会话"] --> B["记录系统整合识别\n(资产/服务/历史偏好)"] B --> C["预测分析引擎\n(推荐下一步行动/画像评级)"] end subgraph Routing ["动态分流与干涉"] C --> D{"客户层级判定"} D -- "普通/大众客户" --> E["智能助理解答 / 赋能坐席"] D -- "高净值 VIP 客户" --> F["跳过排队，直通资深投资经理"] end ``` 针对产品与客群规划，Erin 与 Kiley 补充道，产品设计必须深刻理解不同用户分群（Personas）的核心目标。但在追求“主动解决”的同时，必须警惕交互文案与渠道是否符合客户预期，绝不能在试图“主动消灭痛点”的过程中弄巧成拙，反而制造出新的客户矛盾与困惑。 --- ### [23:15](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=1395s) 人机边界：智能体与人工的无缝交接 关于 AI 应当在何处止步、人工坐席何时介入，讨论激起了热烈的经验分享。Erin 谈到了自己设立的一项硬性规则：只要客户说出“人工坐席”（Agent）这个词，系统必须无条件立即转接到真人。“无论对方是我们最小的客户还是大客户，只要想找真人，就必须能找到，绝不能让他们被困在迷宫般的 IVR（交互式语音应答）系统里团团转。” Amir 则从技术演进的角度提出了互补看法。过去人们之所以痛恨 Chatbot，是因为它们极其生硬，而如今数字助理在语气、抑扬顿挫、口音以及语速控制上已逼真到令人惊叹的程度。但无论技术多成熟，“人机协同”（Human-in-the-loop）的底线不可动摇。当 AI 耗尽了信息检索与解答的价值空间，或者遇到需要真正同理心（Empathy）的情感化场景时，必须当机立断切入人工。最关键的痛点在于**交接过程中的上下文完整性**——绝不能让用户换到人工后把问题从头复述一遍。AI 沉淀的所有意图与背景信息，都必须原封不动地交接给真人坐席，并写回核心业务系统。 ```mermaid flowchart LR accTitle: 智能体与人工平滑交接流 accDescription: 描述从用户交互到同理心触发转接及上下文回流的全过程 subgraph Digital ["数字智能体处理"] A["用户交互"] --> B["意图与情感分析"] B --> C{"是否触碰边界？\n(主动呼叫人工 / 需深度共情)"} end subgraph Handover ["平滑交接机制"] C -- "是" --> D["完整会话上下文与画像打包"] D --> E["人工坐席介入处理"] E --> F["更新记录系统 (System of Record)"] end C -- "否" --> G["AI 自主闭环解决"] ``` Lily Li 进一步拓展了“智能体”的定义维度：智能体不仅是面向终端客户的客服机器人，更是赋能一线销售与运营人员的放大器。过去一名员工只能服务 100 位客户，在多智能体支持下如今能覆盖 1000 位。在“Human-on-the-loop”的管理模式下，人类不需要微观干预每一个决策，而是充当“熔断机制”（Circuit Breakers）和安全护栏的制定者。AI 系统必须以高度的可解释性、可审计性、数据透明度以及清晰的溯源链接，去真正“赢得”内部团队和合规风控的信任。 --- ### [27:48](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=1668s) 破除遗留系统与数据泥潭的现实策略 面对现场观众和普遍企业所面临的“老旧系统与凌乱分散的数据”，专家们剖析了传统架构打补丁时的典型陷阱。Kiley Williams 警告道，切勿被光鲜亮丽的新概念冲昏头脑，必须反复追问核心诉求：“我们究竟想要增强什么业务环节？为什么要把既有 IVR 换掉？”唯有立足领域核心诉求，从辅助内部员工提效做起，才能看清真正的数据依赖。 Amir 直言，数据是 AI 的燃料，如果数据底座本身是一团乱麻，接入 AI 只会以更快的速度“放大混乱”，最终不仅拿不到业务成果，反倒会因为耗费巨资毫无产出导致负责人被解雇。Erin 借用了早期大数据时代的比喻：当年大家把 Hadoop 当成包治百病的“George Foreman 智能电烤炉”，一股脑把所有脏数据扔进数据湖等待奇迹，结果得到的只有垃圾（Garbage in, garbage out）。如今 AI 也是同样的道理，拥有好数据才能提炼好洞察，否则只会以极高的效率产出“AI 废料”（Slop）。 然而，Lily Li 给出了务实的破局建议：**“不要等待完美的时机或完美的数据才开始转型。”** 这种等待往往会演变成无限期拖延。正确的做法是抛开 AI 本身的光环，回归业务的核心产出价值，锁定那些能够快速验证价值、且数据相对干净的高杠杆局部（例如 Erin 团队选择率先改造的电话通讯系统），小步快跑，逐步对核心系统动手术。 - **老旧系统的现实幽默**：Erin 透露，公司里某些运行中的 ERP 系统，甚至比操作它们的年轻员工年龄还要大。 - **历史技术包袱**：Heather 回忆早年在 Travelers 保险公司工作时，有些老旧的 Copybook 数据比自己的年纪（45岁）还要古老。 --- ### [34:23](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=2063s) 信任崩塌的代价与幻觉引发的法律风波 在探讨“AI 翻车事件与信任重建”时，Heather 引用了客户体验领域的经典研究结论： > “传统的 CX 研究告诉我们，一旦出现一次糟糕的体验，平均需要 **11 次连续的积极互动** 才能抵消那一次负面影响，仅仅是为了回到原点（回到零分水平）。” Amir 随之分享了加拿大航空（Air Canada）轰动一时的著名诉讼案：一名乘客因亲属离世需要订票出行，在加航网站与客服 Bot 对话时，该机器人承诺乘客可以在 90 天内申请事后丧亲折扣。然而等乘客参加完葬礼提交凭证时，人工客服却依据公司真实条例文明拒绝，声明丧亲折扣必须在购票前申请，事后一概不予退差。该纠纷最终闹上法庭，加拿大民事裁决法庭（CRT）裁定航空公司败诉并承担赔偿责任，法官明确指出：**无论信息是由真人坐席、数字助理还是网页 Chatbot 所陈述，只要对客户当作事实表达出来，企业就必须予以承认并承担法律责任。** 这一教训为业界敲响了警钟：企业必须在核心业务规则与退款等高风险节点上建立最严密的防护边界，采取强力手段彻底杜绝幻觉（Hallucinations）。AI 在琐碎支持上犯错或许有补救余地，但在涉及政策、法务和敏感流程的关键领域，一次生成错误便可能是灾难性的。随后，主持人引申到投资回报层面：虽然业界频繁宣扬 AI 降低成本与响应时间，但究竟哪些核心指标能够真正反映 AI 与客户长期留存率（Retention）之间的因果关联？Kiley 开始从客户在产品内停留时长等参与度指标切入展开剖析。 ### [37:15](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=2235s) 快速响应与主动服务：以游戏行业为例衡量留存价值 在评估客户体验与平台黏性时，用户在平台上的停留时间与企业为其解决问题的速度存在直接的正相关。以游戏行业为例，当游戏中出现系统故障（glitch）或遭受攻击导致异常时，如果支持团队或自动化系统能第一时间介入，并立即将游戏点数或积分退还给玩家，玩家就会毫不犹豫地重返游戏继续畅玩。相反，如果流程拖沓迟滞，玩家就会在挫败中流失到竞品平台。 更进一步的理想状态是变被动响应为“主动预判与修复”。如果系统能够主动监测到异常的发生，并在玩家甚至还没来得及投诉之前，就通过应用内推送（in-app message）或电话通知对方：“很抱歉我们刚才遇到了一点小状况，我们已经为您的账户补偿了 1000 积分”，玩家的满意度与忠诚度反而会大幅提升。虽然各行各业的业务模式各不相同，但核心逻辑始终如一：确保客户的业务或娱乐流程不被打断，并在出现插曲时迅速且透明地修复，这种无缝体验与持续互动正是企业最需要追踪和衡量的关键指标。 ### [38:27](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=2307s) 传统联络中心 KPI 的重塑与前后对比评估 如果将评估体系放回传统的呼叫中心与联络中心语境，一些经典指标仍然具有极高的参考价值。像自助解决率（containment rate）、平均通话/处理时长（average handle time, AHT）以及首次呼叫解决率（first contact resolution, FCR）依然十分关键。然而在引入 AI 之后，还必须重点关注“重复提问率”（repeatability），即客户在 AI 介入后是否仍然需要反复咨询同一个问题。 为了准确衡量 AI 项目的真实成效，绝不能仅仅依赖技术供应商的宣传，而必须在技术上线前后开展严密的对照测试。在部署 AI 之前先行采集净推荐值（NPS）及相关客户满意度调研，并在上线之后再次进行同等维度的横向对比。如果对比发现承诺的客户留存率并没有提升，或者客户的不满并未减少，这并不意味着要把 AI 工具全盘推翻，而是提示团队需要回过头去微调模型与策略，找出 AI 在特定场景或特定客户群体中出现“脱节”的具体症结。毕竟在这个竞争极其残酷的市场中，获取一个新客户的成本（CAC）远高于留存现有客户，一旦把客户彻底惹恼，拥有海量替代选择的用户会立即用脚投票。 ```mermaid flowchart TD accTitle: AI 客户服务体验优化与评估闭环 subgraph Traditional_KPIs ["传统基础监控"] A["呼叫中心指标<br/>(AHT / FCR / 解决率)"] B["重复进线率监测<br/>(Repeatability)"] end subgraph Benchmarking ["对比与反馈循环"] C["部署前 baseline 调研<br/>(Pre-AI NPS / CSAT)"] D["部署后效果追踪<br/>(Post-AI NPS / 留存率)"] E{"是否达到留存与体验预期?"} F["微调模型配置与特定交互流程"] G["巩固客户信任与延长生命周期价值"] end A --> E B --> E C --> D --> E E -- 否 / 表现未达预期 --> F --> D E -- 是 / 满意度提升 --> G ``` ### [39:41](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=2381s) 机构金融视角：大额资金流、销售覆盖面与商业底线 从大型金融机构的视角来看，留存与转化的衡量尺度被放大到了更高维度。例如在富兰克林邓普顿（Franklin Templeton），客户往往是家族办公室、投资顾问团队或大型机构全国性账户，因此最终最清晰的商业指标就是资本资金流（capital flows）。但在漫长的成交周期中，客户留存很大程度上取决于一线市场主管与销售团队与客户展开深度、高价值对话的广度与频率。 通过为一线团队配备 AI 工具与平台赋能，企业不仅能够敏锐捕捉大客户情绪变化与个性化诉求，还能显著提升一线人员对机构客户的覆盖率（coverage）。现场数据显示，这种高附加值的互动频次在部署 AI 后持续上升，而覆盖面的扩大与情感纽带的强化，最终直接映射并转化为业务销售数字的提升，在底线营收（bottom line）上带来了立竿见影的投资回报。 - **获客成本压力**：获客成本居高不下，使得管理层的核心战略重心必然向“如何长期锁住既有客户”倾斜。 - **高净值客户的情绪与需求**：在机构级服务中，及时洞察顾问与家办的情绪起伏，是维系长期信任与合作关系的命脉。 - **业务覆盖度与营收转化**：AI 释放了一线销售的生产力，使其能覆盖更多高净值客户，将服务覆盖率直接转化为资金流入。 ### [41:14](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=2474s) 实时情绪感知与主动外呼：告别周末手动打标签 情感与意向分析（sentiment analysis）在近几年迎来了颠覆性突破。在过去几十年中，客户体验分析人员常常不得不在周末对着问卷调查里的大量主观文本（verbatim），一行行手工打标签来做归类分析；而如今，大模型不仅能够高精度地推断出用户在此前通话中的真实情绪，更具备在通话进行中实时捕捉情绪波动的能力，从而精准判定当前客户是否处于爆发边缘、是否需要立即转接至人工高级坐席。 结合企业沉淀数十年的历史业务数据，AI 能够清晰识别导致用户流失的敏感触发词与痛点。AI 在后台静默运行全文本实时检索，一旦发现特定导火索——例如产品包装破损、品质异常或是屡次未解决的售后卡点，系统就能自动触发工单，指派专人进行主动外呼挽回。一句主动真诚的“我们注意到您之前致电反馈的问题未能妥善解决，我们特意致电来协助您”，能瞬间化解危机。正如与会者所强调的，在技术赋能的同时，坚守“真实性与透明度”（authenticity and transparency）才是真正赢得人心的关键。 ### [42:25](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=2545s) 观众问答：AI Agent 是否优于人类坐席？ 在圆桌会议尾声的最后两分钟互动环节，现场观众抛出了一个核心问题：“与人类坐席相比，AI Agent 到底是更好还是更差？” 针对这一尖锐提问，嘉宾指出“更好还是更差”纯粹取决于应用场景和期望达成的业务目标： - **效率与信息预判维度**：如果诉求是实现全天候超快速的问题解决，或是希望系统能够提前收集海量跨渠道上下文、预先推演用户需求，那么具备全局数据整合能力的 AI Agent 无疑更胜一筹。 - **共情与高难度升级维度**：如果问题极为棘手，且需要极高的情感温度、同理心与复杂人际协调，那么迅速将对话升级并转交给真人坐席依然是不可替代的选择。 技术与人工并非零和博弈，根据场景将 AI 的高效分析与人类独有的共情力有机融合，才是客户体验升级的终极答案。随着提问解答完毕，嘉宾向全场听众致谢，圆桌讨论在热烈的掌声中顺利闭幕。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=4s) 我每周能收到 30 到 40 个推销联络，每个人都想占用我 15 分钟时间 — Heather Gillbanks > [05:28](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=328s) 我敢说主办方邀请我就是因为这个，我是少数被允许公开谈论实际落地成果的客户。 — Heather Gillbanks > [05:28](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=328s) 如果事情都被自动化了，我们该去干嘛 — Erin Fusaro > [18:37](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=1117s) 终极的客户体验就是你根本不需要客户做任何事；一切都已经替你办妥、修复好了，你甚至根本察觉不到曾有过问题存在。 — Lily Li > [23:15](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=1395s) 无论对方是我们最小的客户还是大客户，只要想找真人，就必须能找到，绝不能让他们被困在迷宫般的 IVR（交互式语音应答）系统里团团转。 — Erin Fusaro > [27:48](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=1668s) 我们究竟想要增强什么业务环节？为什么要把既有 IVR 换掉？ — Kiley Williams > [27:48](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=1668s) 不要等待完美的时机或完美的数据才开始转型。 — Lily Li > [34:23](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=2063s) 传统的 CX 研究告诉我们，一旦出现一次糟糕的体验，平均需要 11 次连续的积极互动 才能抵消那一次负面影响，仅仅是为了回到原点（回到零分水平）。 — Heather Gillbanks > [34:23](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=2063s) 无论信息是由真人坐席、数字助理还是网页 Chatbot 所陈述，只要对客户当作事实表达出来，企业就必须予以承认并承担法律责任。 — Amir Hameed > [37:15](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=2235s) 很抱歉我们刚才遇到了一点小状况，我们已经为您的账户补偿了 1000 积分 > [41:14](https://www.youtube.com/watch?v=SLMHDyO5a2U&t=2474s) 我们注意到您之前致电反馈的问题未能妥善解决，我们特意致电来协助您 ## Mentioned Erin Fusaro, PadSplit, Kiley Williams, Deepgram, Amir Hameed, RingCentral, Heather Gillbanks, CXPA (Customer Experience Professionals Association), 《客户体验知识体系》(CX Book of Knowledge), Lily Li, Franklin Templeton (富兰克林邓普顿), Intelligence Hub, Travelers 保险公司, Air Canada (加拿大航空), 加拿大民事裁决法庭 (CRT), Hadoop, George Foreman ## Video description Erin Fusaro - Chief Technology Officer, PadSplit Kiley Williams - Senior Engineering Manager, Field Engineering, Deepgram Amir Hameed - Senior Vice President - Global Technical Solutions, RingCentral Heather Gillbanks - Executive Editor, CXPA Book of Knowledge Lily Li - Head of AI Adoption and Solutions, Franklin Templeton This session will explore how AI is revolutionizing customer experience (CX) and customer service by automating interactions, personalizing support, and improving response times. Panelists will discuss how AI-powered tools, such as chatbots and virtual assistants, are transforming customer service operations in contact centers, enabling faster resolutions and reducing operational costs. Learn how organizations are integrating AI into their CX strategies to improve customer satisfaction, streamline workflows, provide 24/7 support, and retain customers. The discussion will also examine the challenges of balancing AI automation with human interaction and strategies for creating a seamless, customer-centric experience. [Source](https://www.youtube.com/watch?v=SLMHDyO5a2U)

[0:04] Heather Gillbanks: 我每周能收到 30 到 40 个推销联络，每个人都想占用我 15 分钟时间 [5:28] Heather Gillbanks: 我敢说主办方邀请我就是因为这个，我是少数被允许公开谈论实际落地成果的客户。 [5:28] Erin Fusaro: 如果事情都被自动化了，我们该去干嘛 [18:37] Lily Li: 终极的客户体验就是你根本不需要客户做任何事；一切都已经替你办妥、修复好了，你甚至根本察觉不到曾有过问题存在。 [23:15] Erin Fusaro: 无论对方是我们最小的客户还是大客户，只要想找真人，就必须能找到，绝不能让他们被困在迷宫般的 IVR（交互式语音应答）系统里团团转。 [27:48] Kiley Williams: 我们究竟想要增强什么业务环节？为什么要把既有 IVR 换掉？ [27:48] Lily Li: 不要等待完美的时机或完美的数据才开始转型。 [34:23] Heather Gillbanks: 传统的 CX 研究告诉我们，一旦出现一次糟糕的体验，平均需要 11 次连续的积极互动 才能抵消那一次负面影响，仅仅是为了回到原点（回到零分水平）。 [34:23] Amir Hameed: 无论信息是由真人坐席、数字助理还是网页 Chatbot 所陈述，只要对客户当作事实表达出来，企业就必须予以承认并承担法律责任。 [37:15] 很抱歉我们刚才遇到了一点小状况，我们已经为您的账户补偿了 1000 积分 [41:14] 我们注意到您之前致电反馈的问题未能妥善解决，我们特意致电来协助您

[0:04] Heather Gillbanks: I also understand EBIT's a thing. So, you really have to find a balance between the customer and the profitability of the company. [0:04] Heather Gillbanks: 15 minutes of my time. [0:04] Amir Hameed: AI for AI's sake [5:28] Kiley Williams: Agentic Applied Engineer [12:03] Amir Hameed: gold nuggets [23:07] futuristic IVR jail [23:07] earn the right [27:48] George Foreman grill [27:48] garbage in, garbage out [37:15] Hey, we had an issue, but we credited your account a thousand points, [40:08] We noticed an issue and want to help.

[0:04] Hello. Hello. Uh, I'm Aaron Fizaro. I will be  your friendly moderator for this discussion. Uh, [0:12] I am the chief technology officer of  Padsplit. Padsplit, we are a marketplace, [0:17] two-sided marketplace for, uh, co-housing.  Let's start off with some quick intros. Um, [0:24] I'll start at the end at uh, Kylie. So, what uh  what do you do? Who are you? What do you do? And [0:33] uh like what's your number one thing on  your radar with regard to AI right now? [0:39] Yeah. Hi everyone. I'm Kylie Williams. I am the  senior engineering manager of field engineering [0:43] at Deepgram Graham. We're a foundational  voice AI platform that enables software to [0:49] uh enables developers to uh provide voice  in their software. Um, one of the the the [0:56] the number one thing for me is providing a an  exceptional customer experience. Exceptional [1:02] being that uh it's an indistinguishable feel  indistinguishable feeling for people when [1:07] they're interacting with our platform with our  our uh team of customer support uh engineers and [1:15] they feel like they are getting that personalized  service um no matter what they're going through, [1:19] no matter what level of uh of customer they are.  So that to me is what AI is really allowing us [1:25] to to do. Awesome. Next, I'm Heather Gilbanks.  I'm here representing the Customer Experience [1:31] Professionals Association. We recently published  a book called the CX Book of Knowledge. So, [1:35] if you're into CX, check it out. I actually  have a copy here to give away to somebody. Um, [1:40] I do not claim to be an AI expert other unlike  most of the folks on this panel. I'm here because [1:45] I know how to do transformation. I've been  doing transformation for over 30 years. To me, [1:49] it doesn't matter whether you're talking about  digital transformation, customer, whatever it [1:53] might be, it's all the same. Um, couple things to  know about me. I am customer fixated, which is how [1:58] I ended up in the customer experience space, but  I also understand EBIT's a thing. So, you really [2:03] have to find a balance between the customer and  the profitability of the company. Um, and I'm also [2:09] not here to sell you anything unless you're hiring  a transformation or customer experience lead. Um, [2:13] but top of my priority list right now as a person  who's primarily a practitioner is trying to figure [2:18] out how to sift through all the vendors. We I was  talking to somebody about this at lunch recently. [2:23] I get 30 to 40 outreaches a week from people who  just want 15 minutes of my time. If you do the [2:28] math on that, that's, you know, that's a pretty  significant chunk of of time. So figuring out how [2:32] to sift through those folks and decide which ones  to meet with is pretty important for for those of [2:37] us in the practitioner space. Awesome. Amir.  Hi, I'm Amir. Turn the mic on test. Okay. Uh [2:45] Amir Hamid and I'm the senior vice president of  global technical solutions at Ring Central and we [2:51] basically provide cloud-based communications for  both internal communications and contact centers [2:57] customer experience as well. And really we've been  kind of focused on aentic voice AI. So I have the [3:04] pleasure of leading a team of global experts and  basically we talk to customers day in day out. [3:10] And you know speaking of transformation that's  really what we're doing. We're seeing customers [3:14] in various stages of transformation. Some of them  are stuck in legacy systems. And then we try to [3:20] kind of understand where they are today. We paint  a picture of where we can take them and then we we [3:27] do the hard work of kind of getting through. And I  think the key is, you know, now that we're talking [3:32] about aentic AI and all this, it's got to truly  deliver value, right? So we try to do things like [3:38] self-funded road mapaps where if we're talking to  the customers, we build out an ROI to say, hey, [3:43] if you were to first migrate to the cloud, you can  see these benefits off your legacy platforms and [3:49] then again we bring in more and more of the the  AI solutions and as those benefits are achieved, [3:56] they kind of help fund moving forward. So  it's really not a uh a a automatic turn it on, [4:02] right? you you've got to really look at the  customer's environment, see what their pressing [4:06] needs are first, and then determine which  AI makes the most sense to deliver that ROI, [4:13] or else it's just AI for AI's sake, and that's not  going to help anybody. Agreed. Pragmatic use of AI [4:21] is one of my uh big tenants right now. Lily. Hi  everyone. I'm Lily Lee. I'm head of AI adoption [4:31] and solutions at Franklin Templatin. If you guys  don't know the company, it's a asset manager who [4:36] manage about $1.8 trillion. So big deal regulated  industry. U my focus I guess as part of my job is [4:43] really enterprise AI strategy enablement and  looking for transformation solutions that we [4:48] built across using AI as a catalyst enabler in  particularly for this panel context right for the [4:56] customer experience standpoint that we're really  looking for a connecting client experience and [5:02] engagement with the leverage of all the mode  intelligence that we can really unlock from [5:08] client market product and investment expertise  that we have and really make it more proactive, [5:16] persistent, consistent as well as personalized and  for a companies like us that we want to make sure [5:22] definitely we create things are institutional  great and then institutional great at scale. [5:28] Awesome. Well, thank you all for being willing  to entertain me today. Um, let's start with a [5:35] real example. Uh, and I'll call on Kylie for  this one. Can you tell us about an AI project [5:41] your team has shipped uh or is building right  now for within the realm of CX and what problem [5:49] is it solving and what's it moved the needle on?  Certainly. So within the field engineering team [5:57] uh we have a team a sub team called applied  engineering and one of the the core offerings of [6:03] deepgram is that anyone you talk to as a customer  is going to be an engineer when when you're coming [6:09] for help anything that you you need assistance on.  And recently we've opened up our support platform [6:14] or support capabilities to everyone. So even if  you're a single independent developer you can [6:19] come get help. Now that introduces a a a problem  with scale. You know there's we have thousands [6:26] and thousands of customers. So how do we scale? So  we've introduced the agentic applied engineer and [6:32] we believe that uh being able to serve that tier  one of ticketing is can be done with agents but it [6:40] can be done with agents at a in a place or excuse  me in a way that doesn't sacrifice the ability to [6:45] provide that exceptional experience that I that  I spoke about. And so that's something that we [6:50] started building out in uh roughly December. Um  it's if you do reach out to our team, there is [6:56] an agent, there is an Agentic applied engineer  that we consider a colleague that is working [7:00] alongside our real engineers. And every single  time that they interact or someone interacts with [7:06] that Agentic applied engineer, it gets smarter,  it gets more intelligent, it works with us. And [7:11] that's one those are one of the things that uh we  find we're able to to do more for our customers. [7:16] Now what that does is that frees up the human  engineers to provide even more exceptional service [7:21] to our customers that really really need that  hands-on touch. So that's one thing that we're [7:27] really proud of that uh is running day in and day  out for us. Yeah, absolutely. Does anybody else [7:34] have a story they want to share? I'm pretty sure  that's why they invited me. [laughter] I'm one [7:40] of the few customers who's allowed to talk about  actually having done this. Um my use case is way [7:45] simpler than yours. Um we are um the company  I was most recently with we had a large call [7:50] center situation and uh we basically replaced our  telefan systems with an AI first solution and I [8:00] sold it to the CEO on the basis that it was going  to save us 50% on our telefan bill as compared to [8:05] the previous vendor. um which was great and that  was true but then we started digging into some of [8:11] the AI capabilities and we've been able to reduce  our average um handle time by over uh 60% and the [8:20] post call processing time has gone from uh just  over 7 minutes to about 2 and a half minutes [8:25] which if you're not familiar with call centers  that's a huge piece of the cost. So for yeah [8:31] for my offshore BO we were able to cut the um we  were in the midst of a renewal so we were able [8:36] to really reduce that headcount quite a bit which  had additional savings and like you were talking [8:41] about being able to free people up. So what we're  finding is that having taken some of that scut [8:46] work off of people's plates our better agents are  able to take that cognitive load and they're able [8:52] to give better service to the the customers who  call in with a little bit more complicated things. [8:58] What's an interesting flip side to that that we  were starting to see is also that some of the [9:02] agents, maybe the ones who are not the A players,  they're not able to handle that cognitive load. [9:07] So, I would just give you that little caution that  you need to think about that as you take the the [9:11] simple work away, can the humans that you have  left, can they handle that that remaining work? [9:17] Yeah, I've gotten feedback on um you know, as  I implement AI across my whole organization, [9:25] um there's the overwhelming sort of unspoken fear  of, you know, if we if we automate this stuff, [9:34] what are the employees that we have going  to do? And [snorts] um so pairing your AI [9:42] um journey with upskilling the folks you have and  giving them the opportunity to take on things that [9:49] maybe uh they can use their expertise in that  uh the AI is automating. Um it gets a lot of of [9:57] good adoption. Folks are are excited about the  opportunity to do something other than what AI [10:03] is automating. So cool. Lily, you had something.  Yeah, I also have well obviously we we are also [10:09] doing AI contact center which is probably um  something similar so I'll probably skip that [10:13] example but I do have another kind of example  on more of an institutional scale side that we [10:20] do have um launched and distribution AI platform  called intelligence hub earlier in January this [10:27] year. So ever since launch our entire US wealth  team are using them on daily basis. So in the old [10:34] days that have to navigate a lot of workflows and  and information sources from about 15 systems. Now [10:41] they can actually start their day and manage  their territories and advisor engagements and [10:47] pre and post engagement activities thinking about  uh you know actionable insights and trade reports [10:54] and focus list non-focus list prioritize their  day all in one place. And one day we actually [11:00] had a little bit hiccup and we got crazy  of complaints. People saying why this is [11:04] not working. So that's one good sign that you  feel about people are actually really changing [11:10] some of the way of working fundamentally and  then getting getting the benefits out of the [11:14] platform. By now we're actually already seeing  measurable values out of it. Here what I'm saying [11:20] is values not productivity because we have very  clear benchmarking against the folks or who have [11:27] not used that before that we've seen incremental  client coverage we have seen um more value added [11:33] engagements and also um potentially sales lift  associated with the platform. So this type of [11:40] transformation is require people really to think  about how to work end to end. Obviously AI is a [11:47] really fundamental of it. However, there's still  also some deterministic you need to consider to [11:51] put into automation. So, it is a full to a  point or all of you guys are talking about [11:56] full transformation, but AI it's really a core and  center. Awesome. Did you have anything you wanted [12:03] to add? You you look like you have Yeah. Well, why  should I miss out? No. [laughter] I think I think [12:08] from uh so I don't agree lots of value across  the board. I think the way we're looking at [12:12] it is we're kind of breaking it up into chunks,  right? So we we try to follow the full customer [12:18] journey and apply AI accordingly. So upfront we  want to automate as much as possible. So I think [12:24] you gave a great example. You know it's not about  replacing humans per se. Those those same humans [12:29] shouldn't have to you know get on a call or or  answer a text and and talk about you know uh [12:35] location hours services that they provide. We can  easily automate that with with an AI receptionist [12:41] as an example. So we would do that right up front.  We'd leverage the AI. Then when you're actually in [12:46] a conversation where you need an agent, then  you'd have a handoff. So you've got the the [12:52] uh the digital agent handing off with context  with all the information over to a human agent. [12:58] And that and that human agent we want to apply  AI to assist. So now we want to assist them [13:04] because again at the other end there may be a  high-profile customer. You want to have all of [13:08] the information available to that agent so they're  empowered to have a very meaningful conversation [13:15] whether they're trying to to sell whether trying  to service it doesn't matter we want to give them [13:20] all of that assistance from an AI perspective and  of course the supervisors in the contact center [13:26] would also have all of that leverage and then  then post call we want to analyze right and I [13:32] think this is the real value especially I'm coming  from a voice perspective a gentic voice AI and you [13:38] look at contact centers and you know you look at  historical or legacy quality management and the [13:43] tools that we had it was very difficult to analyze  even say 10 or 15% of the conversations you try [13:50] your best and say I think we've got a problem  over here so I'm going to analyze as much as I [13:55] can there but now with the tools that we have AI  conversational experts something that we've got [14:00] you know 100% of all conversations can be deeply  analyzed you know with pre-built uh formats with [14:08] with uh certain things that you're keying in on  and then you can actually come into problems and [14:14] all of those patterns are are available to  you, right? So you're not missing anything. [14:18] So for example, in the quality management  realm, you know, maybe you've got a certain [14:22] um churn rate that's acceptable to you. So you're  looking across the nation and you're going to say, [14:27] "Okay, well this churn rate is okay. Oh, but this  area it's a little bit high." So you're just going [14:32] to focus in on that area. you're actually missing  out because those other areas while the churn rate [14:37] may be high, if you're actually leveraging AI to  analyze those conversations, maybe there are key [14:43] repetitive problems that are happening in the  northeast or in the southwest that you can now [14:49] analyze independent of the rest and not accept a  certain level of uh of churn. Like why should we [14:55] accept that? We want to be able to do that. And  again, so again, leveraging AI from the beginning [15:00] to the end of the journey is is really what we're  focused in on. And we're and we're doing that for [15:05] customers of all shapes and sizes, right? From  from the very small to the very large enterprises. [15:11] Yeah. I think that reminds me of something that  I've learned [snorts] I think over the years that [15:17] AI is making a lot easier which is [snorts] um I  think in the past we tend to look at what's not [15:25] doing well and then figure out how to fix it and  something that now you can look at uh very easily [15:33] where are things going really really well and are  those things or those patterns are they able to be [15:39] replicated in places that maybe aren't doing well  or in places where they don't exist that might [15:44] even allow another place to do even better. So  using it not just to find and solve problems but [15:51] looking at what is it that we are kicking butt at  and can we replicate that. Um yeah so like AI has [15:59] helped a lot with that 100% and that's what we  do as well. We actually score the conversation. [16:03] So to your point, what does great look like? And  then how do we replicate it and leverage that for [16:08] training? Because you may have certain agents that  have just absolutely excelled. They're delighting [16:12] the customers all the time. Now you know exactly  every single, you know, conversation interaction [16:18] that they're having. What was it? What was the  gold nuggets that they were doing that others [16:22] aren't and then replicating it? Yeah. All right.  Well, let's uh drilling into that a little bit. [16:30] customers like tech has gotten uh gotten really  advanced and customers basically expect you to [16:38] to some degree uh read their minds. They expect  a certain amount of personalization. Um how are [16:45] you using predictive analytics to get ahead of  uh problems before the customer even reaches [16:52] out to complain or ask for assistance? Uh anyone  want to take it first? Yeah, I'll take that. So [16:59] predictive analytics, it's a a fun word. It gets  tossed around quite a bit and it's it's data. [17:05] It's working with data. It's looking what you  have. It's looking at what you know and trying [17:09] to glean as much as you can around the context of  what's going on. And so as I mentioned, Deepgram [17:14] is a foundational AI voice AI platform. So, we're  feeding a lot of customers and we're seeing a lot [17:20] of voice data come through and we're seeing a  lot of data that is potentially having issues [17:26] um especially the issues he was just talking about  and and what we can do is say hey let's try to [17:31] figure out something before a customer raises  their hand and so uh one thing that I tell the [17:36] team is hey we are we do have a support arm of our  team and instead of being reactive we should be [17:42] proactive I love when I get notices from different  systems or tools that I use and say, "Hey, there [17:48] was something that's a little bit abnormal. We  fixed it for you." And so, how far can we go and [17:53] thinking ahead of the customer and thinking ahead  of our our people and starting to say, "All right, [17:57] we notice an anomaly. Let's figure out what this  is. It might be geographic. It might be related to [18:02] this type of customer or this vertical. Why are  we seeing this? What's going on?" And before a [18:08] customer can even notice anything might be going  on before anything even reaches them. And so uh [18:12] when it comes to predicting and understanding and  trying to figure out what are the trends that that [18:17] are telling us more than than we've seen in the  past, that's what we tend to look at. We tend to [18:21] look at our and again those those agents that are  uh running alongside us, it allows them to always [18:27] be on. And so instead of a human having to be up 8  hours and saying, "Hey, I noticed something while [18:32] I was online," that agent is constantly combing  and looking and spawning sub agents to say, "Hey, [18:37] I noticed an anomaly. Go run off and look over  here." and if you come back and you see something, [18:42] let me know. And then we can start to figure  out we're seeing anomalies and patterns that [18:46] can inform us in ways that we can then inform the  customer and that makes them happy. It again goes [18:51] all the way back to giving them that exceptional  customer experience. And an exceptional customer [18:55] experience is if you really have to do nothing.  Everything is just done for you. It's fixed for [19:00] you. You don't even know there's a problem. And  so that's something that's that's critical for our [19:04] operations. I I want to echo on that proactiveness  because I think you're absolutely right that the [19:10] best client experience supposed to be you're  doing all the work that client before realize [19:15] they actually need them and I think you mentioned  something about like the fixing the issue analyt [19:23] in a lot of times in investment space or financial  space that you like to have your engagement [19:28] leaders your market leaders and your investment  expertise to actually identify your needs of [19:34] your portfolio. based on the market activities,  based on the offering of product intelligence [19:38] that we have and all the client behavior and their  preference in the past and actually proactively [19:43] give you recommendations of how to really  optimize your solutions and those are things [19:48] that we're actively thinking and putting into our  platform and engagements and also like the way we [19:54] think about proactiveness and and the technology  that offer you don't have to really limit to one [20:00] technology, right? So we say predictive AI  that's probably like the late late 90s all [20:05] the way to like 20 probably 20 and generated AI  that you heard about 2022 introduction of chat [20:11] GBT and last year Gentic AI they all can all  work together and to really extract them all [20:19] out of the data intelligence that you have and  all the linear you have between the data across [20:25] different dimensions of the journeys. I think  that's something for you to all considering like [20:30] very comprehensively to make sure that the  needs and the issues are both on both side, [20:36] right? The the downside and upside all cover for  your customers. Yeah. And I think you started off [20:41] by talking about personalization, right? So  that's a deep understanding of the customer. [20:46] So obviously we talked about the problems that  they have. You want to anticipate it, but I'm [20:50] I'm looking at it from another angle as well.  You you talked about financial services. Well, [20:54] you've got tiers of customers, right? You've  got tiers of customers and if your agents are [20:59] empowered to have access to the various systems  of record, you know everything you need to know, [21:05] what services they have, what are they likely and  then you can actually use the predictive analytics [21:10] from a from a services offering perspective to  say, hey, you know, you're already here. Um, [21:16] this is something that you might want to consider  next. And then the person feels like, hey, you're [21:20] not wasting my time. I don't have to tell you  what services or what products or what investment [21:25] portfolio. it's already all there and then you  can also offer that level of differentiation [21:31] right because obviously you've got various people  who can respond to various uh customers and again [21:37] if you're high rollers they need or and expect  a certain level of discerned service and I think [21:43] leveraging the AI knowing everything about them  but making the agents aware of it to say hey [21:48] you know what maybe I don't even want an agent  speaking to that person you know if this person [21:53] comes to us it's going to go straight to a senior  portfolio manager, somebody uh more senior that [21:59] can have that level of conversation. Show them  the respect that they're kind of expecting to say, [22:04] "Yes, you you uh you've earned my business. You  clearly want to keep my business because you're [22:09] not putting me in a queue. You're not treating  me like everybody else. You're giving me that [22:13] status that quite frankly I've earned based on my  business." Yeah. It's almost uh so putting on my [22:20] product hat, but when you think about personas  and audience and segmentation of your audience, [22:27] it's understanding the goals of your customers and  then really focusing on the ways in which you can [22:36] uh sort of serve them, answer their  questions before they need it, [22:40] highlight the things that they might be concerned  about, which is very different depending on sort [22:44] of the the customer segment that uh is in the  pipeline. Yeah. And I think to add on to that, [22:50] making sure that your messaging matches what  that segment or group expects is really really [22:56] important. You don't want to solve a problem  proactively and cause a problem in the process, [23:01] right? You want to make sure you're What do you  mean? I love doing that. That would never happen. [23:07] Great. All right. Um let's see. So, as AI takes  over more of the customer journey, we're talking [23:15] about customer journeys a little bit here. Um,  how do you decide where the bot stops and where [23:22] a human needs to step in? So, um, like what's a  friction point that you think virtual assistants [23:29] will help make disappear in the next few  years? Uh, anybody want to take that one first? [23:34] I mean, we we set a very strong rule where I  was that if the person said the word agent, [23:41] they would go immediately to a human. I don't care  if they're our smallest customer. If you want a [23:45] human, you should get a human. And you shouldn't  be forced to go through a labyrinthine IVR to get [23:50] to them. You know, that that just frustrates  people. So, that was one thing that we said, [23:55] no matter how great the AI is, if  they just want to talk to a person, [23:57] let them do it. My father would appreciate that.  A lot of people do. Yeah. Yeah, I mean I guess [24:04] the counter to that is certainly five years ago  you could you knew you were talking to a chatbot, [24:09] you didn't like it and it's like what's happening?  Um I'll be honest the the advances are so strong [24:15] now that I don't mind it. I don't mind it  because um if I can get the information that [24:20] I need irrespective whether it's a human agent  or a digital agent, I'm just fine with it. I'm [24:25] absolutely fine with it. What I'm not fine with  coming uh coming back to your answer is if I've [24:30] had an interaction with the digital agent where  they've already determined who I am, what I have, [24:34] what my issue is and for whatever reasons they  need the handoff because as a best practice you [24:39] always want to have a human in the loop and there  are certain uh times where the AI has exhausted [24:45] its value and you've got to you've got to code  that right into your agents like now you know we [24:49] need like for you know you know um agents can do  a little bit of sentiment analysis but they can't [24:56] truly empathize the way a human can and when  true empathy is required you want to shift to [25:02] a human uh as quick as possible but at the same  time that handoff has to be seamless with all [25:08] of the context given with it so when you when you  do that I think irrespective whether it's a human [25:14] agent or a a digital agent or back and forth it  doesn't matter as long as that journey for the [25:20] individual has all of the context available and  then all of that information that was gleaned is [25:25] then fed back into whatever system is of record or  whatnot as you resolve the issue. Um, like I said, [25:31] I don't I don't mind digital agents anymore. I  think they've gotten and and you're going to see [25:35] it out in the floor. You can come to our booth  or any booth. You I mean the the uh the tone, [25:40] the inflection, the accents, the the speed in  which people talk depending on which region of [25:46] of the US or wherever you might be. It's it's just  amazing now. And uh so personally I'm not averse [25:52] to it. But I would agree if a customer really  wants to speak to an agent, they should absolutely [25:58] have that and not be in some sort of futuristic  IVR jail. I I like to add another angle of this [26:06] because we when we think about agent, there's more  than just the agent directly facing your customers [26:12] and customers. You can also have a lot of agents  supporting your sales staff, your service staff [26:18] or serving your clients and amplify that serving  and scaling that serving in a sense that in the [26:24] past they probably only be able to manage 100. Now  they probably can't cover a thousand and that also [26:30] grow your business and booked and in that case  I think the new model coming into the way there [26:37] could be some sort of human on the loop where  you don't have human to do every micro actions [26:42] or decisions however human needs to really set the  gu guideline policies and circus breakers or rules [26:49] to determine to your questions where to stop the  bot the agent the multi- aent systems at certain [26:56] guardrails And based on how your how your um  business operates and how the risk rules you have [27:01] to comply. Meanwhile, also your agent has to earn  the right like they have to earn the rights to [27:08] um be utilized by those um individuals or teams  by their you know explanability, auditability, [27:16] the transparency, the link sources, all  of those needs to be like fully earned [27:21] the trust of your your sales force your  your um servicing force so that they can [27:26] actually ultimately provide the best client  engagement and experience to your end clients [27:32] And that um I would say that's the guideline. The  oversight accountabilities will set the rules for [27:39] the breaker or an operator to go beyond agents  or humans. Yeah. So, uh I know we've got a lot [27:48] of different size companies here represented uh  both on stage and also out in the audience. So, [27:56] um I know most of you are probably not starting  from a blank slate. Who knew? Um, so you're [28:03] bolting AI onto systems that have been around for  a decade and in some cases probably much older [28:08] than that. Um, so for the folks in the room that  are wrestling with messy scatty scattered data [28:17] um or just in general having to deal with legacy  systems. Uh, what's your advice? I'll take one [28:25] here. So, one of the one one of the things that  I really really preach to the team when they go [28:32] out and consult with our customers is to really  understand the data and the domain that you're [28:38] working within because there's all this cool and  and glamorous and shiny stuff that that can that [28:44] can be done. But if you really focus and say what  am I trying to do? What am I trying to augment? [28:51] What am I trying to enhance? And what they're  talking about the agents and the human agents [28:56] and not having to have everything be agentic  to the customer to the end user but being able [29:02] to support your in-house agents being able to make  their lives a bit easier. Trying to understand and [29:07] we we've seen this many times over. We we've seen  the trans different transformations and different [29:11] really cool technologies and everyone wants to be  an early adopter. But if you really take a step [29:17] back and and take a deep breath and look and say,  "All right, let me adopt this over here and figure [29:22] out that I want to go from from IVR to something  a bit newer. Why do I want to go that way?" So [29:28] asking why and continuously asking why and trying  to understand what the re reason is behind this [29:35] transformation that you're pushing, that's really  going to to start to to clear up the needs and the [29:41] data and the needs and the the complexity that you  might have. Yeah, I I would agree. So discovery [29:46] is absolutely key, right? We need to have that  discovery uh you know understand uh the why of [29:51] why we want to migrate. But coming back to your  question, you you talked about you know legacy [29:56] systems and and the data itself. Well, the reality  is you know as a best practice you have to ensure [30:02] that your data is clean. So I think there's a lot  of planning that happens upfront to ensure that [30:08] you know again data is the fuel for the AI. So if  your data is messy, you're just going to amplify [30:13] that mess that the AI is not going to give you  those business outcomes that you were promised [30:18] if the data itself is suspect. So I think it's  extremely important again as you mentioned that [30:24] discovery. Why do you want to do what you want to  do? What is your expectations? How would you uh [30:29] you know what's the end state? Why do you want to  get to that end state? And then let's take a real [30:34] reality check of where we are right now. Are you  truly ready for AI? Because if you're not ready [30:39] for AI, bolting on AI uh solutions here and there,  it's just going to make it messy and someone's [30:45] going to get fired because they spent a heck of  a lot of money and it's not driving any business [30:49] results. And it it was something as simple as  taking a step back kind of kind of walking slowly [30:54] to to run later, but first make sure you clean up  your data, make sure it's not messy, so that all [31:00] of the AI that feeds off the data is truly going  to provide benefits for you. Yeah, I think um it's [31:08] a step that a lot of companies skip is looking at  the quality of their data and uh having come from [31:17] the olden days when we were all excited about  Hadoop and I remember folks putting they they [31:24] were sold on Hadoop being the George Foreman grill  and they put all of their data in these giant data [31:29] lakes and then they waited for magic to happen and  it turned out garbage in garbage out. Um, and the [31:36] thing that I think AI does really well is if you  have great data, it will bring out great insights. [31:42] If you have not so great data, it's going to  very quickly give you a lot of slop. Um, so [31:48] like your data governance, your data cleanliness,  understanding what is in there and what it means [31:53] is is deeply important to getting the value out of  it. Um, so like the prep work is super important. [32:00] I I completely echo all you guys mentioned about  the data but then what I'm saying next will sound [32:06] like a a life advice don't wait for the perfect  moment just don't wait for that perfect data [32:12] or perfect moment to come and then start your  transformation and I think that's like one thing [32:18] I will add on to the importance of data and the  other thing I will mention is also forget about AI [32:25] forget about AI or technologies think about what  core function you want to deliver out of your own [32:32] industry businesses and see with the new processes  that you want to deliver with the value you want [32:38] to drive what legacy system or what data you need  to clean what legacy system you need to really [32:45] change or update or evaluate where you need to  plug in AI to drive the most valuable components [32:51] out of it so I think that's probably another way  for you to both align with your leaders and get [32:57] your clear values out of your might also help  you to create the functionality that's going to [33:03] be really lasting and long-term strategy for  your farm. Yeah, AI is a large topic and so [33:12] uh don't allow yourself to get so stuck in  everything isn't perfect, but like scope it [33:17] out and figure out where is the biggest impact and  then you know get get a piece of it out. It's just [33:23] tech like so don't let the mystique of it terrify  you. Um, it is not in fact a George Foreman grill [33:30] or a magic box. Yeah, that's actually why we  started with the telefan system because we had [33:34] relatively clean data and we knew that we could  dip our toe into the water and and get some pretty [33:39] quick success out of it. But yeah, to go back and  and clean all the data in all the systems. I have [33:44] ERPs that are older than the people using them.  I'm not making that up. That's the truth. So yeah, [33:50] I I used to work for the Travelers and they had  uh data in um old copy books that went back to [33:58] before I was born and I'm old. I'm 45. So um it  was a lot of data. Sorry, Heather. It's okay. [34:08] I'd like to see 45 again is all I'm going  to say. [snorts] Tech has done a number [34:14] on me. Um so all right. So, here's one that I  think you may or may not have answers to. Um, [34:23] but trust takes a lot of time to build with your  customers. And, um, one really bad rogue chatbot [34:32] interaction can take things sideways. Um, so tell  us about a time AI got it wrong um, and frustrated [34:41] a customer and what happened and how did you  recover? I don't have an AI example, but I do [34:46] have a statistic for folks. Traditional customer  experience research tells us that when you mess [34:51] up, it takes 11 positive interactions to get over  that one negative interaction just to get back to [34:57] zero. So, you really, really, really want to make  sure that you don't let this happen more than it [35:02] absolutely has to. And that's true whether it's  a human or a bot. It doesn't matter, right? But [35:06] just keep that in mind that it takes a long time  to rebuild that trust incrementally. Yeah. And I [35:12] would just add um you know within our uh AI agents  we build parameters in to kind of absolutely [35:18] mitigate hallucinations and certain areas that  are absolutely where you don't want to go. Um [35:23] you know I I I live in Toronto, Canada and there  was a pretty famous story of Air Canada right [35:28] somebody had passed away and so uh the individual  went online and the bot told them that hey don't [35:34] worry you know after within 90 days if you get  back to us we'll discount your ticket because it's [35:39] bereavement. So the individual went, attended the  funeral, came back and spoke to a human agent this [35:44] time. And the human agent said, "No, absolutely  not. This is our policy. There will absolutely be [35:49] no refunds after the fact. You should have told us  beforehand, and that's how we would have done it." [35:56] And so this actually went to the courts in Canada.  It was it was a pretty famous and and the courts [36:00] came back and said sorry you are resp whether it's  your human agent or your digital agent or your [36:05] chatbot if something is stated as a fact you have  to honor it. So so again coming back to you know [36:13] you want to mitigate eliminate the hallucinations  especially in areas where it could get extremely [36:20] serious. So I'm not going to say AI is perfect.  It's not perfect, but there are certainly certain [36:25] uh areas where it can be catastrophic and that  you absolutely can and and and just absolutely [36:31] have to shut down. Yeah, for sure. Um so we  gloss we talked about it a little bit earlier [36:40] and everyone talks about how AI can cut costs  and response times. Um but does it actually keep [36:45] customers around? So customer retention is a key  um sort of metric. What metrics are you tracking [36:53] um or seeing other clients track uh that connect  AI investment to long-term retention? Um I can [37:02] take it. Uh so one of the things that that we've  uh that I've seen just over my my my career in [37:09] conversational AI is how much time customers  are spending whether it's in app in game uh [37:15] on your platform and that is directly correlated  to how quickly you can resolve issues for them. [37:22] For example, in the gaming industry, if you have  customers that write in or dial in and you're [37:28] able to a glitch happens or a hack happens and  you're able to get them credits back immediately, [37:33] they're going to go back to playing the game.  And so, you're going to have happier players, [37:37] happier customers that are doing more because  you're able to help them more quickly. Otherwise, [37:41] they'll get frustrated and they'll leave.  And so if you're able to get back to them, [37:44] if you're able to identify, again, going back  to what I said about proactively seeing what [37:48] happened, they get an inapp message or they  get a phone call, hey, we had an issue, but [37:52] we credit your account, thousand points, whatever  it is, then you're going to be happier, right? And [37:58] um and so just sort of thinking about what it is  that's going to keep people using your product, [38:04] using your platform. Um that's those those are  the the areas that you do want to measure. Um, [38:09] it's going to be different for every industry and  it's going to be different for a lot of different [38:12] companies, but seeing that a customer is able  to continue doing what they were doing and not [38:18] have interruption, maybe they get a notice and  a notification, that's going to that that that's [38:23] what matters and being able to track that  that level of uh of of engagement is what [38:27] matters. Yeah. And if we bring it to a traditional  contact center in terms of KPIs, you know, you've [38:32] got containment rate, your average handle time,  certainly first contact resolution. I mean those [38:37] things are still very valid but also repeatability  right is the customer coming back asking the same [38:42] thing if you've deployed AI I think another uh  really important thing is doing a a um a survey [38:50] in advance the NPS score things like that prior  to the AI and then post uh deploying the AI and [38:57] then actually comparing the two and saying hey we  were made all these promises that the AI would do [39:03] uh you know specifically on retention well is that  actually happening and are our customers actually [39:08] happier or not? And if they're not, it doesn't  mean the AI gets thrown out. It means you've got [39:13] to tweak it. Maybe there's something in there that  is not clicking for your particular customers in a [39:19] particular space. So, I think uh those traditional  KPIs absolutely still apply, but certainly [39:25] followed up by surveys and NPS scores because  uh at the end of the day, if you can't keep [39:30] your customers, I think you mentioned it earlier,  right? uh how how how much more difficult it is [39:35] to acquire a customer versus keeping a customer  happy. And if you've made that customer I rate [39:41] uh it's a very competitive world. They h everybody  has options. Yeah. Like I have to think about [39:47] things like customer acquisition cost all the time  and it's very expensive to acquire customers. So [39:53] it makes a ton of sense to um focus on like how  do we keep them? How do we keep them around long [40:00] term? Um yeah. Yeah, I I would say in our industry  it's probably even larger scale. So one it would [40:08] be very very clear at the end of the measurement  of capital flows that's like very uh measurable [40:14] but in the meantime um the the retention is  really based on how how much coverage our um [40:21] people on the field our market leaders ourselves  people can actually interact and spend meaningful [40:26] conversations with our clients because a lot of  our clients are group of advisors they could be [40:30] family office institutional like national accounts  they're all large large clients and um you know [40:36] really paying attention to how their sentiment is  going and how their needs are being felt that is [40:42] really going to be very clear to see um the trust  being built the relationships so as I mentioned [40:48] that the platform that we offer to our field  people are already showing those valuable um value [40:55] added engagements increasing you're seeing the  coverage increasing and those are direct linked [41:02] formula that we're leading towards the post AI  high um improvement in terms of the actual sales [41:08] numbers. So, we're really seeing that it's coming  down to the bottom line. Yeah, the sentiment's a [41:14] really interesting one because for decades, many  of us have spent our weekends tagging verbatim [41:18] that was, you know, written in by somebody on a  survey, right? And now the AIS are able to give [41:23] you a much better approximation of what was that  person really feeling on that call or even better, [41:29] what is that person feeling on that call right  now? and do they need to be escalated to a person [41:34] at this moment and then you add to that you do  have those decades of data you know what those [41:39] triggers are what those things are that make  people churn the AI can do full text searches [41:44] for you it's a super simple thing but if it's  doing that in the background automatically and [41:49] then spawning a ticket to have somebody call right  like you were talking about the proactive service [41:54] right hey we noticed that you had a problem with  the quality or you know the box wasn't good or [41:59] whatever it is in in your particular industry  that thing that triggers people to get upset. [42:04] If somebody calls them and says, "We saw that you  called in and it wasn't resolved. We want to help [42:08] you." That goes a long long way with folks. Yeah.  I think authenticity and transparency goes a long [42:15] long way. And the more uh that we're able to use  AI and not lose that um is key. All right. Well, [42:25] we got two minutes left. Uh and so if anyone has  a quick question uh I'll take a question otherwise [42:33] I'm going to do a little soft shoe or you know  so the question was are AI agents better or worse [42:43] compared to the human agents? I can take that  really quickly. Um better or worse is subjective. [42:51] If if you want something to get habitate to happen  and get resolved quickly or you want something to [42:56] just try to think ahead of you, then AI agents  are going to likely have a lot more information [43:01] they can collect altogether. Um, but if you want  to escalate quickly and have that empathy that [43:05] Amir was talking about, you're going to probably  want to have that conversation with a human. So, [43:09] I think we can talk a little bit more after  this, but yeah, that's my quick answer. Great [43:13] questions. Thank you all again. Thank you all for  coming and listening to us. It's been a pleasure.