Regal AI Automates Call Centers at 10% of Human Cost

Regal AI 实现呼叫中心自动化，成本仅为人工的 10% 语音 AI 解绑了运营效率与客户体验之间的权衡，提升了敏感任务中的信息透露率，且运行成本仅为每分钟 15 至 20 美分。 Lex Sivakumar 基于超过 5 亿次实际生产通话的数据，阐述了现代语音 AI 如何打破联络中心在运营成本与客户体验之间由来已久的权衡。他指出，在催收和潜在客户挖掘等场景中，消费者面对不带主观评判的 AI 代理时，往往比面对人工客服愿意透露更多敏感的财务及个人信息。通过道路救援服务商 Honk 等案例，Sivakumar 展示了自主 Agent 工作流如何通过并行执行 API 调用将解决时长缩短一半。该分享还详细介绍了 Regal 通过 Regal Copilot 实施的部署框架，重点突出了其每分钟 0.15 至 0.20 美元的全包计费模式，该模式具备无限并发能力，运行成本仅约为人工劳动力成本的 10%。 Regal 从传统 CCaaS 平台转型为语音 AI 代理，利用超过 5 亿次客户通话的数据集来明确成功部署的核心驱动因素。 (0:04) 与人工客服相比，消费者面对语音 AI 时往往会透露更多信息，发言篇幅也更长，因为他们不会感到被评判，也不用担心浪费别人的时间。 (1:40) 在催收和潜在客户挖掘过程中，不带主观评判的 AI 交互能促成更高的财务困难自主上报率，并收集到更丰富的背景信息。 (1:40) 语音 AI 通过提供不会降低服务质量的即时弹性，打破了联络中心在运营成本效率与品牌体验之间的两难权衡。 (6:13) 人工客服在经历高压对话后需要接听 2 到 3 通电话才能恢复基准工作效率，而语音 AI 则保持零情绪波动，能够即时维持基准标准化水平。 (6:13)

According to Lex Sivakumar, VP of Growth at Regal, enterprise Voice AI decouples contact center cost from service quality by executing autonomous multi-system workflows at roughly 10% of human labor costs, while eliciting greater customer disclosure in sensitive interactions. Consumer Candor in Non-Judgmental Voice Interactions (1:40) Addressing concerns that consumers will reject automated voice interfaces, Sivakumar argues that customers prioritize resolution speed over the medium. Drawing on data from over 500 million calls, Regal finds that consumers frequently share more personal information with AI than with human representatives. This dynamic is especially pronounced during sensitive interactions like delinquent payment collections, where human agents often unintentionally project judgment. Freed from social embarrassment, consumers voluntarily disclose financial hardships, target pay dates, and requests for installment plans. Similarly, across discovery calls in healthcare, education, and insurance, prospective leads provide longer, more detailed monologues to AI agents. Sivakumar attributes this to consumer psychology: prospects worry about wasting a live person's time, but feel comfortable taking extended time when interacting with an automated agent. Operational Decoupling and Multi-System Agentic Workflows (6:13) Traditional contact centers operate under a zero-sum trade-off between driving cost efficiency and delivering high-quality customer experience. Human staffing carries operational friction, including an average 30% turnover rate and performance variability. Specifically, Sivakumar points to post-call productivity recovery: a human agent typically requires two to three calls to return to emotional baseline after an abusive interaction, whereas Voice AI operates with consistent, standardized execution. Voice AI has also advanced from conversational Q&A into agentic systems capable of executing end-to-end workflows. In a case study of roadside assistance provider Honk, a human dispatcher historically spent three to four hours manually calling mechanics for quotes while placing drivers on hold. By contrast, an automated Voice AI workflow reduced total handling time by up to 50% by pinging mapping APIs and spawning concurrent outbound calls to five local mechanics simultaneously, securing bids and processing payment without placing the driver on hold. Deployment Benchmarks, Concurrency, and Automated Optimization (16:30) On financial and infrastructure metrics, Sivakumar reports that Regal's Voice AI costs $0.15 to $0.20 per minute—inclusive of model training and forward-deployed engineering—which represents roughly 10% of traditional human staffing expenses. The underlying architecture allows infinite simultaneous inbound and outbound call concurrency, preventing capacity bottlenecks during seasonal surges. To manage rollouts, Regal outlines a playbook starting with a single use case, sandbox testing across edge cases, and an initial production ramp handling 5% to 10% of call traffic. Ongoing maintenance relies on closed-loop optimization through tools like Regal Copilot, which ingests transcripts and sentiment data to isolate failure points and automatically generate updated agent prompts for administrative review. Usage-Based Pricing Versus Outcome-Based Margins (19:29) Regal advocates for transparent, per-minute usage billing over outcome-based pricing models. Sivakumar contends that as foundational LLM inference and token costs decline, vendors operating on fixed outcome-based pricing capture the entire cost-efficiency spread as margin. Under a per-minute pricing structure, cost scales directly with call volume and duration. Sivakumar asserts that this approach ensures that falling underlying compute and infrastructure expenses are passed directly back to enterprise clients as models become more efficient. While Voice AI eliminates capacity ceilings and staffing lag, Sivakumar notes that evaluation metrics cannot rely on rigid formulas: containment and transfer benchmarks must be calibrated to specific workflow goals, such as maximizing warm transfers to human closers in sales qualification while minimizing escalations in basic customer service.

Regal增长副总裁Lex Sivakumar基于超过5亿次通话数据指出，语音AI已脱离简单的问答检索，正通过消除情绪损耗、降低约90%人力成本以及并发执行复杂业务流来重塑企业联络中心；同时，消费者在催收与线索挖掘等特定场景下，表现出比面对人工坐席更深的信息披露意愿。 解耦成本与体验：消除情绪波动与流动性摩擦 (6:13) 传统呼叫中心运营长期受困于在效率/成本控制与品牌体验之间的二选一权衡，优化一方通常会牺牲另一方。Sivakumar指出，语音AI打破了这种关联，使运营成本与客户体验成为可同时优化的独立变量。传统呼叫中心面临约30%的年均坐席流失率、季节性业务激增以及繁琐的人工质检摩擦，而语音AI能够提供即时弹性与标准化交付能力。 在通话后的生产力恢复指标方面，人工客服在经历一次负面或对抗性通话后，通常需要经历至少2至3通后续电话才能恢复正常的情绪基线与业务表现；语音AI则不受情绪波动影响，能够在全量通话中维持即时的基准稳定性。在经济效益上，Sivakumar说明语音AI的综合运行成本为每分钟0.15至0.20美元（包含模型训练、技术支持与驻场工程部署），仅相当于传统人工坐席成本的约10%。 消除人际评判：敏感场景下的高信任披露与接受度 (1:40) 针对企业管理者担忧客户抗拒与AI交谈的疑虑，Sivakumar引用实际交互数据表示，消费者核心关注的是解决问题的速度与成效，而非交互介质本身。分析显示，在账单催收与销售线索挖掘等特定敏感场景下，消费者对语音AI的坦诚程度甚至高于人工坐席。 在逾期催收场景中，人工坐席往往会无意中流露出道德评判色彩，而面对中立客观的语音AI，消费者的心理防备显著降低，更愿意主动披露真实的财务困难、预定还款日期或申请分期方案，使企业得以捕获以往难以采集的高价值数据。在医疗、家政服务、保险及教育等线索初筛场景中，由于潜在客户不存在“浪费他人时间”的心理负担，因而更倾向于在与AI的交互中展开较长独白并详尽说明自身背景。 代理型工作流与并发协同：全美道路救援案例 (8:57) 语音AI已从基础的FAQ知识库问答，演进为能够独立调用底层API、更新CRM系统并仅将边界问题升级至人工的自主代理型工作流。Sivakumar以全美道路救援平台Honk为例进行了对比：传统人工调度需耗时3至4小时，客服人员需反复让受困司机保持在线等待，同时手动逐一致电3至5家当地汽修厂进行询价与比价；而引入语音AI后，系统在识别司机地标的同时，能一边调用地图接口，一边并发调起多个外呼AI坐席同时向5家维修厂询价，无需司机挂起即可在通话内完成比价确认与付款扣款，将整体业务耗时压缩了高达50%。 在技术吞吐能力方面，语音AI在基础设施层面具备无限并发能力，且在面临超出处理边界的复杂诉求时，可通过“三方连线温和转接”（three-way warm transfer）在挂断前与人工坐席平滑交接上下文。Sivakumar特别指出，转接率的高低不能孤立看待，销售线索资格审查流程通常以高转接率将合格潜客交由人工成单为目标，而常规售后支持则以高留存阻断率作为评估基准。 敏捷工程部署、自优化闭环与按时长计费模型 (12:41) 为缩短企业方案的上线落地周期，Regal推出了基于自然语言对话界面的Regal Copilot，支持业务团队在单个半天内完成提示词调试、沙盒环境测试及代理部署，替代了此前工程团队数天的开发周期。在交付路线上，工程团队采用分阶段放量机制：先锚定单一高价值场景并在沙盒中完成极端边缘用例的压力测试，初期仅接入5%至10%的实际通话流量，待阻断率与质量达标后再逐步扩充至全量。 针对上线后的系统运维，平台构建了自动化反馈闭环，通过持续摄取通话录音、转写文本与情绪分析数据，自动识别代理的应答短板与知识库漏洞，生成靶向改进建议，管理员点击确认即可直接更新代理策略。在商业化定价方面，Sivakumar主张采用透明的按分钟用量计费（每分钟0.15至0.20美元），而非部分竞品采用的按结果付费；他解释道，随着底层大语言模型推理与Token成本逐步下调，按用量计费能够将算力降价红利直接让渡给客户，而按结果计费则会导致服务商独占底层成本下降带来的利润差额。 Sivakumar的论点表明，语音AI在联络中心的价值主张正从技术替代转向复杂业务编排与边际成本解耦。然而，简报中消费者高信任披露的实证主要集中于欠款催收与线索初筛等特定场景，对于高情绪对抗或诉求极度复杂的争议场景，AI能否维持高满意度仍有待更大范围的数据检验；同时，随着大模型效率提升与单次交互时长趋于缩短，按分钟用量计费的商业模式如何平衡厂商的研发投入与营收扩张，也是行业演进中值得关注的问题。

## Background 本内容由 Ai4 于 2026 年 8 月 21 日发布在 YouTube 平台上。视频记录了 Regal AI 增长副总裁 Lex Sivakumar 的主题演讲，他基于 Regal 累计驱动超过 5 亿次真实通话的业务数据，深度拆解了高绩效语音 AI 落地的基准指标、客户成功案例，以及区分成功上线与停滞试点的关键因素。 ## TL;DR 演讲全面回顾了 Regal 从传统 CCaaS 向语音 AI 智能体平台的转型历程，并揭示了诸多反直觉的落地洞察。Sivakumar 指出，AI 的无道德审判属性显著提升了用户在催收与潜客跟进场景中的倾述意愿，打破了呼叫中心在运营效率与客户体验之间的传统权衡。通过 Honk 道路救援等标杆案例，他展示了 AI 并发外呼与调度将履约耗时缩短 50% 的实战成果，并介绍了 Regal Copilot 如何实现智能体“自构建与自迭代”。在问答环节，他还详细阐述了沙盒渐进式落地路径、每分钟 0.15–0.20 美元的全包定价模型，以及按时长计费如何将底层算力降本红利切实让渡给客户。 ## Atmosphere & style 整场演讲呈现出干货密集、务实沉稳且极具实战导向的专业氛围。主讲人 Lex Sivakumar 语言精炼、条理清晰，既有对行业宏观趋势的洞见，又有来自一线 5 亿次通话数据的扎实支撑。从开场调动全场互动，到深度拆解技术与心理学维度的反直觉现象，再到 Q&A 阶段对敏感成本与竞品差异化问题的坦率直答，现场情绪始终保持在专注而高效的行业研讨状态。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=4s) Regal 创始团队具备管理 3 万人呼叫中心的丰富经验，推动公司从传统 CCaaS 转型为语音 AI 智能体平台，并在生产环境中累计驱动了超过 5 亿次通话。 - [01:36](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=96s) 用户的核心诉求是快速解决问题；在催收还款等敏感场景下，AI 的无道德评判属性消除了沟通羞耻感，反而使用户更愿意坦诚沟通困境与还款计划。 - [01:36](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=96s) 在潜客跟进场景中，消费者因不必担心“浪费坐席时间”而倾向于面对 AI 进行大段深入陈述，帮助企业挖掘更详实的个性化需求与筛选资质。 - [06:13](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=373s) 语音 AI 将联络中心的运营成本与客户体验解耦为可同时优化的独立变量，彻底规避了人工坐席 30% 的高流失率与负面通话后的情绪恢复周期。 - [08:57](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=537s) 语音 AI 深入端到端复杂业务协同，例如 Honk 道路救援通过主智能体保持通话并并发呼叫 5 家维修商比价派单，将整体脱困耗时压缩 50%。 - [12:06](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=726s) Regal Copilot 开启了“以 AI 构建与管理 AI”的模式，将专业智能体部署缩短至一个下午，并支持基于通话录音与情绪分析的闭环自愈调优。 - [14:11](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=851s) 企业落地语音 AI 应采取克制的渐进式策略：历经业务脚本盘点、沙盒极限压测、5% 至 10% 小流量灰度放量，达标后再逐步规模化推广。 - [16:56](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=1016s) 全包式语音 AI 部署成本约为每分钟 0.15 至 0.20 美元（含算力与驻场工程交付），整体支出仅相当于传统人工坐席劳动力成本的 10% 左右。 - [17:51](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=1071s) AI 具备无上限的并发承载能力，遇到长尾未覆盖场景时通过“暖转接”（Warm Transfer）无缝交接给人工坐席，保障服务平滑过渡。 - [18:47](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=1127s) 转接率指标需依业务场景定制：在线索初筛场景中追求 100% 精准识别并转接销售，而在售后客服场景中则追求尽量压低转接率以提升拦截率。 - [19:33](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=1173s) Regal 坚持按时长计费而非按成效收费，核心逻辑在于将底层大模型 token 成本不断下降的技术红利直接让渡给企业客户。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=4s) Regal 的创立背景与 5 亿次通话里程碑 演讲开场，Regal 的增长副总裁 Lex Sivakumar 介绍了自己的职责范围，他目前全面掌管公司的市场营销、业务拓展（BD）以及收入运营（RevOps）团队。在向现场观众互动提问时，举手表示听说过 Regal 的人并不多，Sivakumar 借此详细介绍了公司的创立渊源：Regal 是一家语音 AI 智能体平台，约 6 年前由 Alex Levin 和 Rebecca Greene 联合创立。在创办 Regal 之前，这两位创始人曾在全美最大的本地家政服务平台之一 Angie 管理着一个规模达 30,000 人的庞大呼叫中心。 起初，Regal 是一家直接与 Five9、Genesys 等老牌巨头竞争的呼叫中心即服务（CCaaS）平台；在大约三到四年前，团队敏锐捕捉到技术变革趋势，果断切入语音 AI 智能体领域。如今，Regal 已经达成了为客户驱动超过 5 亿次通话的重大业务里程碑。现场调查显示，台下已有约 30% 至 35% 的企业投资了语音 AI。Sivakumar 表示，无论是已经入局的先行者，还是剩余 65% 仍在观望的企业，Regal 从这 5 亿次真实通话中沉淀的生产级经验都能提供极具价值的避坑指南与实施路径。 - **创立团队背景**：创始人 Alex Levin 与 Rebecca Greene 曾统领 Angie 旗下 30,000 人的大型呼叫中心，深谙呼叫中心一线运营痛点与业务指标。 - **业务转型历程**：从最初对标 Five9、Genesys 的传统 CCaaS 架构，转型为深耕 3-4 年的语音 AI 智能体平台。 - **业务体量**：生产环境累计驱动超过 5 亿次通话，沉淀了海量的人机互动对比数据。 --- ### [01:36](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=96s) 客户接受度反直觉洞察：为什么用户更愿意向 AI 敞开心扉 在与企业沟通语音 AI 时，Sivakumar 最常听到的顾虑就是“客户接受度”——企业往往担心真实客户反感与机器对话，甚至一接通就强硬要求转接人工客服。然而，Regal 分析海量通话数据后发现，用户的核心诉求极为务实：尽管人类在沟通中对信任的门槛极高，但用户真正在乎的仅仅是以最快的方式达成诉求，至于沟通载体是面对面交谈、呼叫中心坐席还是语音 AI，用户其实并不介意。更令人意外的是，在多类敏感或复杂的业务场景中，消费者面对语音 AI 时反而表现出了比面对人类坐席更高的开放度。 最典型的验证来自催收还款（Payment Collections）场景。催收通话高度敏感，当人工坐席询问还款计划时，潜台词往往暗示对方违约失信，用户极易感受到审判与道德压力（Judgment）。而换成语音 AI 拨打后，这种人际间的道德评判感彻底消失，用户反而愿意坦诚沟通自身困境，主动说出诸如“这个月手头确实紧”、“我下周五发工资就还”或者“能否帮我申请分期付款”。这促成了双赢：消费者摆脱了心理压力与羞耻感，而企业则获取到了过去人工沟通中极难挖掘到的真实财务状况信息。 类似的情况也出现在潜客线索跟进（Lead Generation Follow-ups）场景中，广泛覆盖医疗、本地生活、保险和教育行业。当潜在客户在网站填表未完成或需要进一步咨询时，传统认知认为人工回访更显诚意，但数据对比揭示了一个反直觉现象：面对语音 AI，客户的回答篇幅显著更长，甚至会出现大段的主动陈述（Extended Customer Monologues）。背后的心理机制非常微妙——人类会本能地愧疚于“浪费另一个人类的时间”，担心自己提出“愚蠢的问题”；但在语音 AI 面前，这种顾虑荡然无存。无论是本地房屋翻新项目的个性化细节，还是购买保险背后的隐秘缘由，客户都会毫无保留地倾述，为企业后续提供极度个性化的服务与客户资质筛选（Disqualification）提供了宝贵的第一手信息。 - **核心驱动心理**：用户对沟通形式并不教条，唯一追求的是“快速解决问题”；AI 的无评判属性消除了沟通耻感与顾虑。 - **催收数据表现**：消除违约者的被审判感，大幅提升分期协商意愿与真实还款日期披露率。 - **线索转化表现**：消除客户担心“浪费坐席时间”的心理包袱，诱发长篇幅输入与深度业务诉求挖掘。 --- ### [06:13](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=373s) 打破联络中心的传统权衡：效率与体验的兼得 Sivakumar 指出，在传统的呼叫中心管理逻辑中，管理者始终在两个互斥的目标之间走钢丝：一是实现预期的业务产出与效率指标（如营收增长、以可预测的低成本解决进线），二是提供代表品牌形象且令人满意的客户体验。在过去的单维管理模式下，追求其中一项往往必须以牺牲另一项为代价，二者就像是互相制约的“因变量”。但语音 AI 的引入彻底打破了这一二元对立的权衡，将原本绑定的变量解耦为可以同时优化的“独立变量”。 管理人工呼叫中心不仅要应对剧烈的季节性业务波动，还必须承受行业平均高达 30% 的员工流失率，耗费大量资源在招聘、绩效管理与人工质检（QA）上。语音 AI 不仅能做到业务波峰波谷的瞬时弹性伸缩，还能实现自动化呼叫监控与全量质检。更为关键的是情感劳动维度的差异：人工坐席在经历一场充斥负面情绪或冲突的艰难通话后，往往难以立刻调整心态，通常需要经历至少 2 到 3 通电话的缓冲期才能恢复到基线生产力水平（Return to baseline productivity）；而语音 AI 智能体则完全不存在情绪耗竭或服务质量波动，在每一通电话中都能保持绝对标准化与一致性的高水平交付。 ```mermaid flowchart TD accTitle: Traditional Contact Center vs Voice AI Paradigm accDescription: Comparison between traditional trade-offs in contact centers and the voice AI independent variable paradigm. subgraph TRAD ["传统呼叫中心：相互制约的权衡"] direction TB T_EFF["运营效率 / 成本控制"] <-->|"不可调和的拉扯"| T_CX["客户体验 / 品牌形象"] T_LOSS["痛点: 30% 年均流失率 / 负面情绪需 2-3 通电话恢复"] end subgraph VAI ["语音 AI 平台：解耦为独立变量"] direction TB V_CORE["语音 AI 智能体"] V_CORE --> V_SCAL["即时弹性伸缩与自动化 QA"] V_CORE --> V_EMO["零情绪损耗，秒级恢复基线表现"] V_SCAL --> V_GOAL["兼顾降本增效与标准化高体验"] V_EMO --> V_GOAL end ``` - **行业基准流失率**：人工呼叫中心的平均人员流失率高达约 30%，带来极其沉重的培训与管理负荷。 - **情绪恢复周期**：人类坐席遭遇高压或恶劣通话后，需经历“至少两到三通电话”才能重回基线生产力，而 AI 智能体实现了“瞬时重置”。 - **变量关系转变**：将成本与体验从互相拉扯的约束关系，转化为可同时提升的双赢指标。 --- ### [08:57](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=537s) 超越简单应答：Honk 道路救援如何将耗时压缩 50% 语音 AI 的能力范畴早已脱离了前几年单纯基于知识库/FAQ 的简单问答阶段，正在深入到复杂业务流程的自主决策、系统集成与端到端工作流接管。Sivakumar 以全美知名道路救援服务商 Honk for Help（模式类似 AAA）为例，深度拆解了语音 AI 如何重塑高并发、重协同的履约工作流。 在接入 Regal 之前，司机在高速公路突发车辆抛锚时，人工调度模式极为繁琐低效：车主进线报备故障与位置后，坐席需在地图系统人工搜索附近 3 至 5 家汽修厂或拖车行；随后坐席将车主置于静音等待（On Hold）状态，挨个给修理厂拨打电话比选报价；收集完毕后切回车主通话汇报比价方案，车主确认后坐席再次将车主置于等待状态，重新呼叫选定的修理厂确认派单；最终救援车辆赶往现场，车主在线下自行结清款项。整套人工流转耗时通常高达 3 到 4 个小时，令身处危险公路边等待的抛锚车主苦不堪言。 ```mermaid flowchart LR accTitle: Honk for Help Voice AI Orchestration accDescription: Parallelized roadside assistance workflow enabled by Regal AI voice agents. subgraph INBOUND ["进线与定位"] A["车主呼叫进线"] --> B["AI 实时调用地图 API<br/>精准锁定周边维修点"] end subgraph CONCURRENT ["无挂起并发调度"] B --> C["车主保持在线沟通"] B -.->|"触发子智能体并发外呼"| D["并发呼叫 5 家维修商<br/>获取实时报价"] D -.->|"汇总比价数据"| C end subgraph FULFILL ["履约与结算"] C --> E["车主在线秒级决策"] E --> F["系统自动派单 &<br/>在途电话扣款"] F --> G["端到端耗时缩短 50%"] end ``` 通过引入 Regal 语音 AI 智能体，原先必须依赖人工线性推进的流程被高度并行化压缩：车主进线说明附近地标，AI 瞬间调用地图 API 匹配周边救援商；在完全不中断与车主通话、不进行通话保持的情况下，主智能体直接调度系统并发向 5 家服务商发起外呼比价并即时汇总数据；车主在电话中完成确认后，智能体不仅完成派单，还能在救援人员前往现场的途中直接在通话中完成线上支付。这一重构彻底消除了多轮电话转接与等待，将车主从报案到脱困的端到端耗时直接压缩了多达 50%。 - **传统模式痛点**：多重电话保持（On Hold）与人工线性比价，导致道路救援全流程耗时长达 3 至 4 小时。 - **并发能力释放**：利用 AI 可以发起无限进出线呼叫的特性，在保持与车主通话的同时并发呼叫 5 家维修商完成竞价。 - **业务价值落地**：将端到端履约周期大幅缩减高达 50%，人类坐席得以从低价值询价事务中抽离，专注于极具挑战性的关键个案。 --- ### [12:06](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=726s) AI 构建 AI：Regal Copilot 与下一代智能体管理 面对语音 AI 深入复杂业务流所释放的巨大价值，Regal 思考的核心命题转向了如何最大程度压缩客户的“价值实现时间”（Time to Value）。为此，团队在大约半年前推出了代号为“以 AI 构建 AI”（AI building AI）的核心产品——Regal Copilot。 以往搭建一套上述复杂的专业级语音智能体，需要 Regal 内部专业的部署工程师团队耗费数天时间手工配置与调试。而在 Regal Copilot 发布后，企业客户只需登录官网，便能通过类似 Gemini 或 ChatGPT 的纯自然语言对话界面，在短短一个下午的时间内自主完成高可用语音 AI 智能体的搭建与上线。更深远的意义在于全生命周期的元治理（Meta-governance）：管理人员可以在该对话界面中持续下达指令监控与优化智能体，例如直接询问“该智能体近期表现如何？”、“平均处理时长（AHT）是多少？”、“调出它出错的边缘案例（Corner Case）”或“我需要更新哪些 Prompt 和知识库？”。 Sivakumar 指出，行业正在步入一个“智能体管理智能体”的元管理时代（Meta world）。在可预见的未来，企业的组织架构图（Org Charts）下方将出现层级分明的 AI 梯队，由上层的高阶调度智能体管理下层执行各项细分业务的子智能体。这一演进彻底改变了软件交付的范式，让先进 AI 的配置门槛实现了数量级的下降。 - **构建周期缩短**：借助 Regal Copilot，智能体部署交付周期从过往的“数天专业团队支持”缩短至“单人一个下午自主完成”。 - **自然语言元监控**：支持通过自然语言直接下钻查询 AHT、归纳异常 Edge Case、一键微调底层 Prompt 与知识库切片。 - **组织形态预判**：企业未来的团队架构中，自主智能体将形成类似人类管理梯队的层级架构（Org Chart of Agents）。 --- ### [14:11](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=851s) 现场互动与问答：智能体从沙盒测试到生产落地的演进路径 在演讲尾声，Sivakumar 邀请现场参会者前往 1225 号展位，负责操刀 Regal Copilot 的产品经理也在现场提供实操演示；同时他还预告了 Regal 将于 9 月 17 日在纽约举办的 Regal Rise 年度用户大会（支持现场参会与线上直播），届时将有众多标杆企业登台分享落地经验。随后，演讲进入现场提问环节。 现场观众首先提问：“企业究竟该如何训练智能体，以及如何界定哪些场景归 AI、哪些场景归人工？”Sivakumar 回答称，这是 Regal 前线部署工程师（FDE）标准入场流程的核心。第一阶段双方会全面盘点现状，客户需提供现有的呼叫话术脚本（Call Scripts）以及当前人工呼叫的业务画像；第二阶段团队会进行优先级排序，精准锁定一个最适合通过语音 AI 优化的具体场景。 具体落地采取极其克制的沙盒渐进策略：先在沙盒环境搭建智能体原型，工程师团队通过海量模拟通话对其进行高强度的压力测试（"Prompt, reprompt, and basically test it, beat it against the wall"），不断逼出系统极限；直到双方对智能体的抗压表现完全放心后，才会将其切入真实生产环境，最初仅分配生产呼叫量的 5% 至 10% 进行金丝雀放量；只有当它持续稳定达成最低标准门槛后，才会逐步扩大该场景的放量比例，并横向拓展至更多业务场景。紧接着，现场观众提出了第二个关于“如何衡量解决率/拦截率（Containment）”的问题，Sivakumar 坦言 Containment 看似简单，但在不同客户的业务定义中存在巨大差异…… - **展位与大会信息**：展位号 1225；Regal Rise 年度大会定于 9 月 17 日在纽约举行并支持全球直播。 - **落地方法论**：拒绝全盘盲目上线，严格遵循“脚本与流程梳理 -> 挑选单一高优场景 -> 沙盒暴力压测 -> 5%至10%小流量灰度 -> 达标后规模化铺开”的阶梯式推进路径。 ### [16:30](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=990s) 成功指标的界定与定制化评估 在问答环节中，Lex Sivakumar 首先探讨了企业客户如何定义语音 AI 的成功指标。他指出，许多客户通常会采用行业标准的“拦截率”（Containment Rate），即在不转接人工坐席的前提下，AI 能够独立端到端处理的通话量占比。不过，Regal 在指标制定上保持了极高的灵活性，认为这应该是一个“双向沟通的协商过程”（a two-way street）。 每个企业客户的具体业务场景和关注点各不相同，常常出现不同客户关注不同核心指标，或者对同一个指标（例如转化或解决标准）持有完全不同的业务定义。因此，Regal 坚持针对每个客户进行个性化梳理与协同定义，确保技术落地的考核口径与客户的实际商业目标严格对齐。 ### [16:56](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=1016s) 成本模型：每分钟费用与全包式部署支持 当被现场观众问及语音 AI 的具体使用费率时，Lex Sivakumar 给出了明确的基准数据：使用成本大约在每分钟 15 到 20 美分之间（15 to 20 cents per minute）。值得注意的是，这一单价不仅包含了底层的模型调用与算力消耗，还全包了模型训练支持、Regal 团队的人力投入以及专属前线部署工程团队（Deployed Engineering Team）驻场协助的所有费用。 Lex Sivakumar 强调，从整体投资回报率（ROI）的角度来看，这种包含全套工程交付支持的 15 至 20 美分/分钟定价，“通常仅相当于人类坐席劳动力成本的 10% 左右”。这使得企业在极大地缩减服务成本的同时，还能获得高标准的工程实施保障。 - **每分钟综合费率**：0.15 美元 – 0.20 美元 / 分钟。 - **费用涵盖范围**：算法模型、系统训练、平台技术支持以及前线部署工程师驻场配合。 - **人力替代效益**：整体支出仅占传统人类客服/销售坐席人力成本的约 10%。 ### [17:17](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=1037s) Regal Copilot：基于通话数据的闭环自主迭代 针对客户是否能通过 Regal Copilot 构建具备自主反思并根据真实交互重新提示（Prompt & Reprompt）的智能体这一提问，Lex Sivakumar 给出了坚定的肯定回答。用户不仅能够对上线的智能体进行持续全生命周期的监控与调优，还可以直接向系统提问排查痛点，比如查询“智能体在哪些环节无法有效处理交互”、“知识库在哪类问题上失效了”或是“交互逻辑在哪里中断”。 在底层逻辑上，Regal Copilot 会主动调取通话录音、解析对话文本并进行情绪分析（Sentiment Analysis），精准定位异常与薄弱点，随后自动向管理员输出针对性的调整与提示词优化建议。运营人员只需轻点“接受建议”（Accept Recommendations），系统就会自动完成对 AI 智能体的更新迭代，实现无缝的闭环自愈与能力进化。 ```mermaid flowchart TD accTitle: Regal Copilot 智能体自主优化闭环流程 subgraph Monitor["通话采集与分析"] A["真实客户交互录音"] --> B["转录文本与情绪分析"] B --> C["识别断点与知识库缺陷"] end subgraph Improve["建议与自动更新"] C --> D["Copilot 生成调优建议"] D --> E["管理员确认 'Accept Recommendations'"] E --> F["AI 智能体完成更新迭代"] end F -.-> A ``` ### [17:51](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=1071s) 无限并发容量与“暖转接”（Warm Transfer）容错机制 在回应关于是否存在“容量受限导致 10% 到 20% 的电话无法承接”的问题时，Lex Sivakumar 明确澄清：在纯并发容量层面上完全不存在瓶颈，“AI 语音智能体具备瞬时承接无限路通话的能力，无论进线（Inbound）还是出呼（Outbound）都能同时处理”。 对于提问中实际隐含的另一个痛点——当 AI 能力不足以解答特定复杂长尾场景时该如何应对，Lex Sivakumar 指出标准解法是配置“暖转接”（Warm Transfer）机制。他幽默地类比了日常生活中拨打宽带客服的经历：“就如同你因为家里 Wi-Fi 坏了给电信运营商打客服电话，结果接通了错误的部门，坐席会在三方同时在线的情况下，以暖转接的方式把你平滑交接给另一位坐席。”AI 语音智能体同样遵循这一逻辑，在极少数无法独立消化的场景下，以三方通话的形式无缝交棒给人类专家。 ### [18:47](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=1127s) 转接率的业务差异：线索质检与客服支持的对比 关于“平均转接率究竟是多少”的问题，Lex Sivakumar 坦言很难给出一个单一的静态数值。一方面转接率直接取决于智能体的成熟度；另一方面，转接比例本身在不同场景下所代表的业务意义截然不同，甚至某些场景下高转接率本身就是核心成功标准。 他举例分析道，在线索资格初筛（Lead Qualification）场景中，AI 的核心目标就是精准识别意向客户并将高净值线索交接给人工销售，此时“100% 的转接率反而是理想目标”；而在典型的售后客户服务（Customer Service）问询中，高转接率则意味着自动化成效不佳，团队追求的是尽量压低人工介入比例。因此，Regal 倾向于根据具体业务场景，与客户共同定义专属的健康转接率目标，而非套用一刀切的行业通用数字。 ### [19:33](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=1173s) 商业模式与竞品差异化：按时长计费与红利让渡 面对语音 AI 价格战加剧以及大量竞品涌入的市场现状，Lex Sivakumar 深入剖析了 Regal 在定价策略与商业模式上的核心差异化壁垒。与市场上许多竞品采用的“按结果/成效计费”（Outcome-based pricing）截然相反，Regal 始终坚持“按实际使用时长（Per-minute）计费”。 Lex Sivakumar 犀利地指出了按成效计费背后的利益冲突：随着底层大语言模型（LLM）与语音 token 成本越来越便宜，如果服务商依然按固定成效收费，“这意味着技术降本带来的所有差价与超额利润全部被软件供应商赚走了”。而 Regal 采用纯按时长计费的底层逻辑，正是为了将模型成本持续下降的技术红利直接让渡给客户（pass on the savings）。随着模型 token 成本不断走低，Regal 会主动下调分钟单价，让企业客户的 AI 调用变得越来越轻量与实惠。 ## Notable quotes > [01:36](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=96s) 人类在沟通中对信任的门槛极高，但用户真正在乎的仅仅是以最快的方式达成诉求，至于沟通载体是面对面交谈、呼叫中心坐席还是语音 AI，用户其实并不介意。 — Lex Sivakumar > [01:36](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=96s) 面对语音 AI，客户的回答篇幅显著更长，甚至会出现大段的主动陈述。背后的心理机制非常微妙——人类会本能地愧疚于“浪费另一个人类的时间”，担心自己提出“愚蠢的问题”；但在语音 AI 面前，这种顾虑荡然无存。 — Lex Sivakumar > [06:13](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=373s) 人工坐席在经历一场充斥负面情绪或冲突的艰难通话后，往往难以立刻调整心态，通常需要经历至少 2 到 3 通电话的缓冲期才能恢复到基线生产力水平；而语音 AI 智能体则完全不存在情绪耗竭或服务质量波动。 — Lex Sivakumar > [12:06](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=726s) 管理人员可以在该对话界面中持续下达指令监控与优化智能体，例如直接询问“该智能体近期表现如何？”、“平均处理时长（AHT）是多少？”、“调出它出错的边缘案例（Corner Case）”或“我需要更新哪些 Prompt 和知识库？”。 — Lex Sivakumar > [12:06](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=726s) 企业的组织架构图（Org Charts）下方将出现层级分明的 AI 梯队，由上层的高阶调度智能体管理下层执行各项细分业务的子智能体。 — Lex Sivakumar > [14:11](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=851s) Prompt, reprompt, and basically test it, beat it against the wall — Lex Sivakumar > [16:56](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=1016s) 通常仅相当于人类坐席劳动力成本的 10% 左右 — Lex Sivakumar > [17:51](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=1071s) AI 语音智能体具备瞬时承接无限路通话的能力，无论进线（Inbound）还是出呼（Outbound）都能同时处理 — Lex Sivakumar > [17:51](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=1071s) 就如同你因为家里 Wi-Fi 坏了给电信运营商打客服电话，结果接通了错误的部门，坐席会在三方同时在线的情况下，以暖转接的方式把你平滑交接给另一位坐席。 — Lex Sivakumar > [18:47](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=1127s) 在线索资格初筛（Lead Qualification）场景中，AI 的核心目标就是精准识别意向客户并将高净值线索交接给人工销售，此时“100% 的转接率反而是理想目标” — Lex Sivakumar > [19:33](https://www.youtube.com/watch?v=1rzKiOIQwFs&t=1173s) 随着底层大语言模型（LLM）与语音 token 成本越来越便宜，如果服务商依然按固定成效收费，“这意味着技术降本带来的所有差价与超额利润全部被软件供应商赚走了”。 — Lex Sivakumar ## Mentioned Lex Sivakumar, Regal AI, Alex Levin, Rebecca Greene, Angie, Five9, Genesys, Honk for Help, AAA, Regal Copilot, Regal Rise, Gemini, ChatGPT ## Video description Lex Sivakumar - VP, Growth, Regal AI Regal has driven over 500 million calls for hundreds of customers, across both inbound and outbound use cases. In this session, Lex Sivakumar, VP of Growth at Regal, will share what that data actually reveals about high-performing Voice AI deployments versus stalled pilots. Lex will cover benchmarks behind the calls in production, walk through real customer examples, and dig into the analysis behind those 500 million calls, including what separates the deployments driving real revenue and efficiency gains from the ones that never get past testing. [Source](https://www.youtube.com/watch?v=1rzKiOIQwFs)

[1:36] Lex Sivakumar: 人类在沟通中对信任的门槛极高，但用户真正在乎的仅仅是以最快的方式达成诉求，至于沟通载体是面对面交谈、呼叫中心坐席还是语音 AI，用户其实并不介意。 [1:36] Lex Sivakumar: 面对语音 AI，客户的回答篇幅显著更长，甚至会出现大段的主动陈述。背后的心理机制非常微妙——人类会本能地愧疚于“浪费另一个人类的时间”，担心自己提出“愚蠢的问题”；但在语音 AI 面前，这种顾虑荡然无存。 [6:13] Lex Sivakumar: 人工坐席在经历一场充斥负面情绪或冲突的艰难通话后，往往难以立刻调整心态，通常需要经历至少 2 到 3 通电话的缓冲期才能恢复到基线生产力水平；而语音 AI 智能体则完全不存在情绪耗竭或服务质量波动。 [12:06] Lex Sivakumar: 管理人员可以在该对话界面中持续下达指令监控与优化智能体，例如直接询问“该智能体近期表现如何？”、“平均处理时长（AHT）是多少？”、“调出它出错的边缘案例（Corner Case）”或“我需要更新哪些 Prompt 和知识库？”。 [12:06] Lex Sivakumar: 企业的组织架构图（Org Charts）下方将出现层级分明的 AI 梯队，由上层的高阶调度智能体管理下层执行各项细分业务的子智能体。 [14:11] Lex Sivakumar: Prompt, reprompt, and basically test it, beat it against the wall [16:56] Lex Sivakumar: 通常仅相当于人类坐席劳动力成本的 10% 左右 [17:51] Lex Sivakumar: AI 语音智能体具备瞬时承接无限路通话的能力，无论进线（Inbound）还是出呼（Outbound）都能同时处理 [17:51] Lex Sivakumar: 就如同你因为家里 Wi-Fi 坏了给电信运营商打客服电话，结果接通了错误的部门，坐席会在三方同时在线的情况下，以暖转接的方式把你平滑交接给另一位坐席。 [18:47] Lex Sivakumar: 在线索资格初筛（Lead Qualification）场景中，AI 的核心目标就是精准识别意向客户并将高净值线索交接给人工销售，此时“100% 的转接率反而是理想目标” [19:33] Lex Sivakumar: 随着底层大语言模型（LLM）与语音 token 成本越来越便宜，如果服务商依然按固定成效收费，“这意味着技术降本带来的所有差价与超额利润全部被软件供应商赚走了”。

[0:04] Lex Sivakumar: Regal has driven over 500 million calls for hundreds of customers, across both inbound and outbound use cases. [1:40] Lex Sivakumar: humans feel bad about wasting another human's time, but they don't feel bad about wasting a Voice AI agent's time at all. [6:13] Lex Sivakumar: return to baseline productivity after a negative call [6:13] Lex Sivakumar: When a human agent undergoes a stressful or abusive conversation, it takes at least 2–3 subsequent calls to return to normal emotional baseline and performance. Voice AI operates with zero emotional variability, maintaining instant baseline standardization across every call. [12:41] Lex Sivakumar: AI building AI [12:41] Lex Sivakumar: beating it against the wall [16:30] Lex Sivakumar: Compared to traditional staffing models, this represents roughly 10% of the cost of human labor. [17:51] Lex Sivakumar: From a pure infrastructure and throughput perspective, voice AI imposes no capacity constraints: agents can handle an unlimited number of simultaneous inbound and outbound calls without degradation. [19:29] Lex Sivakumar: When vendors bill strictly per outcome while foundational LLM inference and token costs decrease over time, the vendor captures the entire margin spread.

[0:04] Hello everyone. Thanks so much for for having  me. Uh like she said, I am Lex Civic. I'm VP of [0:09] growth at Regal. I lead our marketing, business  development, and revops teams. Uh so just uh a [0:15] quick show of hands. How many of you have heard of  Regal before? Okay, I see a few hands, but not too [0:22] many. So I'm very glad I'm here. Uh Regal, we are  a voice AI agent platform. We were founded about [0:27] 6 years ago by Alex Leaven and Rebecca Green.  The two of them ran a 30,000 person call center [0:33] at a company named Angie, which is one of the  largest home services providers in the nation. Uh, [0:38] prior to starting Regal. Um, so we started off  actually as a CCast platform. So competing with, [0:44] you know, the 59s or Genesises of the world and  we foray into voice AI agents about 3 to four [0:50] years ago. So since then we've learned a ton. Um,  we've done a lot. We we've worked with a lot of [0:55] customers. Um so we've since hit a milestone.  We've driven over 500 million calls for our [1:00] customers. Uh so what I'd love to do today is just  talk to you about a few of the things that we've [1:05] learned along the way. Um go a little bit deeper  into a customer example that we have and then also [1:11] um allow time for ample Q&A at the end. Um so  again, I'd love to for this to be interactive. So [1:17] how many of you have already invested in voice AI  at your companies? Okay, nice. Uh, a good probably [1:25] 30 35%. Um, I'm very glad to be speaking to you  as well as the the other 65% that haven't yet. I [1:31] think there's a few things that we can pass along  that um can hopefully help you along your journey. [1:36] So, like I mentioned, the first thing that we'll  do is go over a few things that we learned. So, [1:40] one of the insights that I'll start with um  is regarding customer acceptance. Right? So, [1:44] this is always on the top of folks minds when I  speak to them, especially at events like this. [1:50] Usually the first objection that I get is people  are sometimes wary that their customer base that [1:55] they serve uh will actually want to talk to  voice AI agents um instead of just requesting [2:01] a human ASAP, right? Um and we've done a lot of  analysis on this and what we've learned is that [2:06] while humans have a super high bar for trust  for any conversation that you're having, all [2:12] they really care about is getting what they want  at the end of that conversation in as quick of a [2:17] way as possible. regardless of the vehicle that  gets them there which could be you know us in a [2:23] one-to-one human conversation like in person could  be a human contact center agent on the other side [2:28] or it could be voice AI um so one of the examples  that I'll actually talk about on this slide is [2:34] um an interesting one in payment collections so we  run a lot of payment collections type uh calls for [2:40] our customers and what we've learned when we match  up human calls versus voice AI calls is actually [2:46] that the consumer on the other end of the phone  uh opens up more with the voice AI agent than they [2:51] do with a human agent. So, if you think about that  use case, it's particularly sensitive, right? It's [2:57] a financial use case. Um someone's picking up the  phone asking someone when they're going to make [3:01] their next payment. They're implying that you're  delinquent. Um so, there's a bit of judgment that [3:06] comes off on the phone. When a voice AI agent  actually makes that call, that sense of judgment [3:11] is gone. Um so, people actually on the other end  of the phone, they open up a lot more. So they [3:17] say things like, "Hey, it's been a tough month or  I'm getting paid next Friday or uh actually can [3:22] we work out an installment plan?" And that ends up  being good for both parties, right? The consumer, [3:26] they feel like they can, you know, open up,  say a little bit more about their situation [3:30] without any sort of judgment on the other end  of the phone. And then also for the business, [3:35] it's super important because then they get access  to information that they wouldn't have otherwise [3:39] gotten access to. Um, the other example that I'll  actually talk about is super interesting is lead [3:45] generation follow-ups. So, this is true across,  you know, many industries, healthcare, local [3:50] services, insurance, education. Um, but if you  think about it, you're a consumer and you want to [3:56] express interest in a product or service. Usually  what happens is you go to Google, you search [4:00] for whatever you're looking for, you get onto a  website, and then you fill out a form, right, that [4:05] collects basic information, name, age, location,  maybe some details about what you're looking for. [4:10] And then often times what happens is you either  need more information so you're calling in to that [4:14] website or they call you because you abandoned  your form or you didn't take the actions they [4:19] thought you were going to take. Um what happens on  that phone call actually is super interesting. So [4:25] when we lined up both human agent calls and voice  AI calls, what we learned is that people open up [4:31] more to voice AI. Um and you can see some of the  stats here, right? So more long form responses [4:36] and more extended customer monologues. So when we  started listening to these and tried to figure out [4:41] why, the main reason stuck out and it was it was  basically because humans feel bad about wasting [4:48] another human's time, but they don't feel bad  about wasting a voice AI agent's time at all. So [4:53] they're they're less likely to ask what they may  what they may perceive as stupid questions when [4:59] they're talking to another human. But when they're  talking to a voice AI agent, again, going back [5:03] to the payment collections use case, the judgment  isn't there, right? So they're opening up. They're [5:08] exposing a lot more about their situation, about  their context, uh about what type of service that [5:13] they want. If it's, you know, local services like  a remodeling project or if it's insurance, why [5:18] they want the insurance, they're providing a lot  more details upfront that that business wouldn't [5:23] have otherwise gotten. And again, it ends up  being super helpful for both sides of the of the [5:28] the equation there. So for the consumer, they're  able to open up, get a lot more information out, [5:34] and then receive more pertinent information to  them. And then for the business, they're able [5:38] to get a lot more information about the customer  upfront. So then they can tailor and personalize [5:43] the customer experience from that point onwards.  So all of you work with customers, you you know [5:49] customer discovery is the most important pro part  of the process in order to figure out whether you [5:53] lead in with someone or maybe someone doesn't  work out and you have to disqualify them and [5:57] focus on on folks that could be a better fit. So  we found that this uh otherwise counterintuitive [6:03] learning that people are actually accepting  AI and they're like talking more and more [6:07] uh is beneficial for both for both sides,  not just the brand but also for the consumer. [6:13] So the other thing that we've learned um and that  we see in voice AI is it allows you to improve [6:18] efficiencies without the traditional trade-offs  that you think of in the contact center. So if [6:24] I'm running a contact center, you know, 5 to 10  years ago, I'm going to largely oversimplify, [6:28] but I have two main goals. One is to deliver  whatever outcome it is. So it could be you know [6:34] driving revenue for my business or uh resolving  you know inbound customer service inquiries [6:39] uh at a managed predictable cost. So otherwise  an efficiency metric. And the second thing I'm [6:45] responsible for is providing a customer experience  that I am proud of and that the brand proud is [6:50] proud of and represents that brand. Um and even  the way that I've phrased it, you can tell if [6:54] you're thinking about this in the past oftentimes  you have to make decisions or run tests. uh when [6:59] you're optimizing for one, it's at the cost of the  other. So they're almost uh dependent variables. [7:05] What voice AI allows you to do is kind of separate  them and make them independent variables. So you [7:10] can actually influence both at the same time. So  if you're managing humans in a contact center, [7:15] often times a lot of the things that's on your  mind, um you could be a seasonal business, [7:19] right? You have to hire um performance manage,  you have to manage turnover. Average turnover [7:24] in a contact center is about 30%. Um and then  you have to constantly uh monitor calls to make [7:29] sure everything is at a certain uh standard  quality. Uh with voice AI obviously a lot of [7:34] those things become easier slashgo away right  you can flex up and down at the at the snap of [7:40] a finger. You can also uh automate a lot of the  call monitoring that you have automate QA. So, [7:47] uh, we found that it kind of shifts this paradigm  on its head where you're able to think of running [7:52] tests to influence both of those variables instead  of thinking about one of them as a trade-off. So, [7:58] one of the nifty metrics that we found in in  our calls is basically the return to baseline [8:04] productivity after a negative call. So, if you  think about AI, not only uh does it handle the [8:10] call that you're in, but you should think about  the call after that and the call after that. Um, [8:14] we all have tough conversations. Uh, in a contact  center, uh, job oftentimes the agent is having, [8:20] you know, tough conversations a lot more than they  probably like to. Uh, and it's tough, right? It's [8:24] emotionally tough. Uh, a lot of times people  don't hold their composure on the call. It not [8:29] only affects the call that they're on, but also  takes a while for them to return to a baseline [8:33] state to serve customers the way that you want to  serve them. So, we found that it takes at least [8:38] a couple calls for folks to return back to that  productivity level. Whereas you can imagine with [8:43] voice AI um that happens instantaneously, right?  There's no variability in the behavior or the [8:49] standardization of your voice AI agent. So now the  the last thing that uh we've learned and that I'll [8:57] talk a little bit about is that AI is going uh way  beyond automation, right? So, if I was here a year [9:02] or two ago, I would tell you, you know, AI could  handle conversational Q&A, could probably, you [9:08] know, link up to an FAQ and give you a lot of the  answers that you're looking for. What we're seeing [9:13] now, especially with our customers, is that voice  AI is taking decisions. Um, it's updating CRM, [9:19] it's updating systems, it's taking o over more and  more of the human workflow in a way where you can [9:25] trust it and only escalate for the super important  human critical use cases. So, one of the examples [9:31] that I'll talk about is actually a customer of  ours. So, Honk, uh, Honk for help is a roadside [9:37] assistance vendor. Think of them as kind of like  a AAA, right? So, first I'll go through kind of [9:42] what their customers were going through before  they used us, and I'll talk a little bit about [9:46] how they're using us to automate a lot of their  calls. So, if you think about it, you know, you're [9:50] driving on a on a highway. Um, your car breaks  down, you pull over to the side of the road. [9:55] Uh what happens today or before this is you pick  up the phone, you call your roadside assistance [10:00] vendor, you tell them where they're where you're  at, and uh an agent looks up where you are [10:04] probably on Google Maps or whatever system that  they're using. They find nearby service providers. [10:10] So mechanics, auto shops, whatever it is, they  put you on hold. They call those mechanics, [10:15] those auto shops, probably call about five of  them, three to five depending on where you are. [10:19] Try to get competitive bids. And then they call  you back while you're waiting on hold. deliver [10:24] the information that they just collected, tell  you which one, tell you about all of them. You [10:28] choose one of them with a certain price. They go  back on hold. They call that service provider back [10:33] and say, "Hey, this person wants to engage your  services. Uh they get get you back. They send the [10:39] service provider to you. Service provider comes to  your car, fixes your car, and you handle payment [10:43] with them. So before we were helping Honk for  help, that whole process took about three to four [10:49] hours. So obviously not great for you who side  who's stuck on the side of the the highway with a [10:54] flat tire or busted carburetor or whatever it is.  Um so using us our voice AI agents we've been able [11:00] to compress a lot of that workflow into the agent  itself as opposed to using the human. So which [11:06] then frees up that human for higher value tasks.  So what what our agent does is when you call in [11:12] you can say what landmark you're near. It'll ping  our agent will ping a maps API automatically find [11:17] the nearest service providers close to you.  While you're still on the phone, it'll engage [11:22] another agent to make five outbound calls all at  the same time because you can make infinite calls [11:27] inbound and outbound at the same time. Collect the  information that the human agent was previously [11:32] collecting and then deliver it back to you while  you're still on the phone without going on hold. [11:35] and then you make a decision, figure out who you  want to engage. Uh, and then it can also handle [11:40] payment on the phone while the service provider  is on the way. So, you can have a lot of these [11:45] different things going at the same time, which  eventually gets you off the side of the highway in [11:50] much less time. So, we've seen that with Honk for  help specifically for its consumers, they've been [11:55] able to cut that endto-end time down by as much  as 50% in in a lot of these use cases, which is, [12:00] you know, overall just really good for that end  consumer um who's uh just trying to get home. [12:06] So, uh, we talked about a few things just now.  Um, we talked about customer acceptance. We talked [12:11] about, um, AI kind of shifting the way that you  think about running tests without the traditional [12:16] trade-offs. And then the last thing we spoke about  is, you know, how much, uh, how how much better [12:21] AI is getting, the decisions that it's handling,  the parts of the human workflow that it's handling [12:25] that it previously couldn't. Um, so it has us  constantly thinking about how can we compress [12:31] the time to value for our customers so consumers  can see the type of benefit that I just I was [12:37] just mentioning. Now, um, so we're constantly  thinking about how to accelerate that and how [12:41] to get it quicker and quicker. So, um, what we've  done a lot this year is invest into what we're [12:47] calling AI building AI, right? So, um, this is a  new product that we have that we released about [12:52] 6 months ago. It's called Regal C-pilot. It is  basically our version of using AI to build those [12:57] AI agents. So whereas previously we had you know  a human team we still have that human team in our [13:03] building the agents that I just mentioned you  can actually today if you wanted to go onto our [13:09] website and try out Regal Copilot and you can see  it's a conversational interface and you can set [13:14] up voice AI agents in the span of an afternoon.  you know, whereas it would have taken a few days [13:20] for our team to do that um in the prior life, now  you can do all of that by yourself in a you know, [13:26] Gemini chatgpt type interface. And then not  only that, you can also continuously monitor and [13:32] improve all of your agents within this interface.  So you can query it and ask, hey, how's this agent [13:37] doing? Um what's the average handle time? Or uh  give me a corner case where it didn't go right. [13:42] Uh what prompts do I need to update? Where where  do I need to update the knowledge base? and it'll [13:46] give you all of that. So then you're basically  managing an agent that's managing an agent, [13:51] right? And we're getting kind of into the into  a meta world where you know in the future you're [13:55] going to have many agents that you that you rely  on that have org charts and layers underneath them [14:00] for other agents that they handle. So it's a very  very exciting time and uh really exciting time for [14:05] us to be able to try to bring this value to you.  So um that was all of the content that I have, [14:11] but a few things that I'd like to mention. So  we are at booth 1225. We'd love to show you [14:16] Regal Copilot. Our product manager that actually  developed the product is here with us right now. [14:21] So she can give you demos. Uh and obviously I've  only able been able to talk at surface level about [14:26] some of these things. So please come to the booth.  And then also we have a conference called Regal [14:31] Rise. It's our annual user conference on September  17th where we're going to have a bunch of our [14:35] customers come in and speak about how they've  implemented voice AI at their companies as well as [14:40] some industry peers and partners. So, um, that's  September 17th in New York City. So, if you'd [14:45] love to come in person, that'd be great. Um, but  if you can't, we're also live streaming it. So, [14:50] uh, at the end of this, uh, we'll have members  of our team handing out one pagers about Regal [14:54] Rise and you can sign up and we'd love to see you  there. So, uh, with that, I have five minutes for [14:59] questions if any of you have any. Booth 1225.  Yeah. Okay. So the question is about training [15:08] the agent and how many use cases you know go to  the agent versus the human. So um this is all part [15:14] of the kind of the onboarding experience with our  four deployed engineers. So what we do on day one [15:19] is we try to assess your current state right and  that's with your help. You give us you know your [15:23] call scripts. You tell us what kind of calls your  humans are handling versus you want AI to handle. [15:27] Um and like a full description of what's going  on today. What we do then is then we prioritize [15:32] the ones that we want to optimize via voice AI  and we'll pick one use case where the goal is [15:37] to start with the small percentage of the calls  that you handle in production with the ultimate [15:42] outcome being you know more and more of those  calls. So, we'll start with building a voice AI [15:46] agent in a sandbox uh environment um to set that  up and then our team will go prompt, reprompt, [15:53] and basically test it, beat it against the wall,  make sure that it's handling all the the simulated [15:58] calls the right way. And then once we we're  comfortable, both of us are comfortable that [16:02] it's performing good enough to handle production  calls, we'll start with something like 5 to 10% [16:06] of production calls. And then as they keep uh  meeting the minimum standard requirements, we'll [16:11] go and expand more and more into um that like more  of that use case and then also other use cases. [16:20] Yeah. And the the second question is how do you  measure containment? So basically it depends on [16:24] containment seems simple but it really depends  on every customer. Um so short answer is whatever [16:30] however you you uh define it today. uh I'll say  what we do for a lot of customers is basically [16:35] the standard uh definition of containment. So  how many calls can I handle before without going [16:39] to a human so but we're flexible on the success  metrics and that part of the process is a two-way [16:45] street and we handle that with every customer  individually. So a lot of times it's different [16:49] metrics or it's the same metric with different  definitions. Uh we work with you on that. Yes. [16:56] Yeah. So the second question is what is the  cost per minute of using the voice AI? So, [17:01] it's anywhere from, you know, 15 to 20  cents per minute. That includes all of the [17:06] uh training, the cost for our people, the for  deployed engineering team that comes out. Um, [17:11] but that's usually about 10% of the cost  of human labor. Yes. So, the question is, [17:17] if you're using Regal C-Pilot to build an agent,  will you be able to prompt and reprompt by itself [17:22] using customer interactions? The answer is yes.  So uh you can use it continuously monitor and [17:27] improve all of the agents that you have but  specifically you can prompt it and say hey [17:32] where is the agent not able to uh handle the  interactions or like where's the knowledge base [17:36] failing or where's the agent failing then you can  get call recordings uh it can analyze those call [17:41] recordings analyze the transcripts the sentiments  suggest recommendations to you and then you can [17:46] simply just say accept recommendations and then  it'll go and improve the agent itself. Do you [17:51] mean from a capacity perspective? Oh okay.  So the question is if there are 100 calls [17:56] uh are there are there any types of calls where  you know we can't handle 10 to 20% of the calls [18:01] on a purely capacity basis. The answer is no. Um  so AI voice agents can handle infinite calls all [18:07] at the same time um both inbound and outbound. I  think maybe a second part of the question you're [18:12] getting to is what if you know the agent isn't  equipped to handle those calls. So in those you [18:17] in those use cases we suggest building in a warm  transfer functionality to a human agent. So that's [18:23] basically the same way, you know, if you're on  the call with Spectrum because your Wi-Fi went [18:26] out and you're in the wrong department,  the agent connects you to another agent, [18:30] does a warm transfer while all three parties are  on the phone. We do the same thing with a with a [18:34] voice AI agent. So we recommend for all those  times where, which is a minority of the time, [18:39] where a voice agent can't handle it, you do  a warm transfer to a human that can't. Yeah. [18:47] Yeah. The average the question is what is the  average percentage of warm transfer? I'll say [18:52] it's hard to give you a number on this because  it's across all of our customers. It depends on [18:56] how good the agent is like how mature it is. Uh  when when we start out the there and sometimes [19:02] there's a success metric for a certain percentage  of transfers because in sometimes that could be [19:07] a good thing right lead qualification that's  a great thing if you qualify a lead as an AI [19:11] agent and then you hand it to a customer. We aim  to you know uh figure out a success metric with [19:16] our customers and make sure we're hitting that.  So it's I I hate to avoid the question but I can't [19:20] give you a number because it varies widely like  with lead qualification 100% is the goal right you [19:25] want to qualify all the leads that you get in and  then give them off to your human customer service [19:29] inquiries probably not 100%. So yeah yeah yeah  yeah. So the question is the price of voice AI [19:35] is going down and uh a lot of competitors are  in the space. How do you differentiate all of [19:39] that? I'll answer it but two ways. So one is the  pricing. So uh we price on a usage basis on a per [19:45] minute basis. A lot of our competitors price on  an outcome basis. So we price specifically on a [19:49] minute basis in order to pass on the savings  to our customers. So if you think about it, [19:54] if the AI and LLMs are getting cheaper and cheaper  and you're paying per outcome, that means the [19:59] vendor reaps all the the difference, right?  Reaps all the margin on that. Um, what we aim [20:03] to do is price per minute. So then we pass those  savings on to you. So when tokens become cheaper, [20:09] we reduce the usage price so then it becomes  cheaper and cheaper. So, I hope that answers your