Brinks Home Cuts Call Duration 50% With Voice AI Agent

Brinks Home凭借语音AI智能体将通话时长缩减50% Brinks Home通过将平台主导权移交给业务团队并部署自动化测试框架以防范运营故障，成功扩展了其AI智能体的应用规模。 Philip Kolterman与Deepank Sharma深入剖析了将AI智能体从实验性试点转向企业级生产环境所需的关键策略。他们强调了赋能业务团队主导AI工作流、选择成熟平台而非冒风险自建API配置，以及利用自动化测试满足严格治理标准的重要性。Brinks Home的虚拟语音智能体“Veronica”取得了显著的实际成效，包括将处理时长缩短50%，并将转接率从30%降至8%。展望未来，他们强调要利用人机协同（Human-in-the-Loop）架构，并聚焦于长期的变革性业务指标。 主要行业分析机构预测，到2027年AI试点项目的失败率将达40%。 (0:04) 在选择供应商或技术之前，争取到跨组织的认同是成功的先决条件。 (2:02) 为了实现长期成功，应由业务团队而非IT部门来主导AI流程和工作流的维护。 (2:02) IT部门的职责应严格聚焦于提供API、基础设施和框架支持。 (2:02) 与成熟的供应商平台相比，利用基础通用LLM在内部自建AI智能体面临高得多的试点失败风险。 (5:23)

Deepank Sharma of Cresta Intelligence and Philip Kolterman of Brinks Home argue that avoiding enterprise AI pilot failure requires business-led operational ownership, specialized platform partnerships over raw commodity models, and rigorous automated testing. Kolterman demonstrates that combining automated evaluation with human-in-the-loop escalation allows contact centers to scale virtual agents safely while significantly cutting call minutes and transfer rates. Business Ownership and Focused Governance (2:02) Kolterman stresses that cross-organizational alignment must precede vendor selection. Rather than isolating AI operations inside IT, Brinks Home designed its architecture so that IT manages infrastructure, telephony, and APIs, while customer experience teams and business operators directly control workflows, configuration, and data maintenance. According to Kolterman, this separation eliminates operational bottlenecks and allows business teams to make iterative updates as artificial intelligence models rapidly evolve. To prevent paralysis, Kolterman cautions against oversized steering committees, advising leaders not to give every stakeholder a vote. Restricting governance and deployment decisions to an interdisciplinary core of essential leaders enables teams to pivot swiftly when operational snags arise during production rollouts. Platform Economics Over DIY Commodity Builds (5:23) Addressing why enterprises should avoid building purely in-house on raw model APIs, Kolterman recounts that Brinks Home previously attempted a DIY build using GPT-3.5 and Amazon services, but could not produce an agent sufficiently reliable for live customer interactions. He contrasts contact center economics across three tiers: human-handled calls averaging over $10.00 each, raw commoditized LLM calls costing roughly $0.10, and established enterprise platforms like Cresta operating in the $1.00 range per interaction. While raw APIs appear cheaper, Kolterman argues that the greatest enterprise cost is pilot failure. Paying mid-tier platform costs offsets risk by providing native fallback mechanisms to human agent interfaces, standardized agent desktops, and integrated quality assurance, helping companies avoid the 40% AI pilot failure rate projected by Gartner and McKinsey by 2027. Automated Testing and Foundational Data Architecture (10:56) Because conversational voice agents are non-deterministic, Kolterman notes that subjective 'vibe checks' are insufficient for enterprise safety. Over an 18-month span, Brinks Home introduced automated testing frameworks that evaluate virtual agents across thousands of synthetic test cases prior to release, allowing engineers to diagnose root causes objectively and prevent functional regressions during updates. This testing layer depends on pre-existing operational hygiene, including clean API lineage, normalized CRM data architectures, and baselines created by ingesting historical human agent interactions. Kolterman characterizes the resulting reduction in contact center minutes as a stair-step trajectory, where continuous A/B testing and incremental feature releases drive distinct drops in handling times. Measurable Outcomes and Human-in-the-Loop Collaboration (13:41) In production, Brinks Home's voice agent, 'Veronica,' achieved a 50% reduction in handle time minutes, lowered transfer rates from 30% down to 8%, and delivered a 30-point upside in Net Promoter Score (NPS). However, Kolterman acknowledges that the virtual agent's NPS still trails human agents due to a lack of natural conversational warmth, prompting A/B testing around nuance, tone, and proactive context, such as detecting upcoming technician dispatches. To close this gap, Brinks Home is implementing a Human-in-the-Loop (HITL) architecture with Cresta. When the AI encounters uncertainty, it signals a supervisor or Tier 2 agent who inspects the live call, injects guidance, and hands back control without transferring the customer, reserving human labor strictly for high-value friction points. Kolterman and Sharma conclude that enterprise AI requires shifting from immediate operational tracking to long-term transformational metrics, such as revenue generated 6 to 12 months post-launch. Citing industry estimates from Gartner, McKinsey, and MIT showing that 40% to over 60% of enterprise AI initiatives fail, they caution that agent deployment is too complex for commoditized tooling, making deliberate vendor selection and human-machine collaboration decisive factors in escaping pilot purgatory.

Brinks Home首席创新官Philip Kolterman与Cresta Intelligence现场首席技术官Deepank Sharma指出，企业AI智能体若要跨越试点失败陷阱，关键在于确立业务部门主导的权责划分、平衡商业平台与廉价模型的经济风险，并通过严密的自动化治理与人机协同机制实现可持续的生产级价值。 平台采购与自研风险的经济权衡 (5:23) 针对企业应自研大模型管道还是采购成品平台的争议，Kolterman透露，Brinks Home曾尝试使用GPT-3.5配合Amazon云服务自建虚拟智能体，但最终无法使其达到面向客户的高可靠标准。Sharma援引Gartner与麦肯锡的行业预测指出，约40%的AI试点项目预计将在2027年前走向失败；而在更广泛的统计中，甚至有超过60%的企业级AI计划难以走出试验阶段。直接调用大模型原始API虽然表面成本低廉，却蕴含极高的试点失败风险。 从成本维度分析，人工客服单次通话平均成本超过10美元，直接调用原始开源或商业模型API的成本在0.10美元左右，而采购Cresta等专业供应商平台的交互成本则约为1美元。Kolterman认为，尽管原始模型极具价格吸引力，但成熟平台自带内建的质检机制（QA）以及无缝回退至人类客服的原生系统界面。通过支付处于中间区间的平台费用，企业能够大幅降低落地阻力与试错风险，避免因系统不可控而导致项目全盘崩溃。 业务主导与精简决策架构 (2:02) 在组织协作层面，Brinks Home建立了明确的二元分工体系：IT部门负责底层基础设施，包括提供API接口、电话系统接入与安全框架；业务及客户体验（CX）团队则直接拥有并运行AI平台，全权负责数据维护、工作流配置及日常流程优化。Kolterman指出，由于底层AI技术每周都在快速迭代，让一线业务团队直接掌控平台能够消除层层流转的运维摩擦，实现对话策略的灵活敏捷调整。 在管理与治理机制上，Kolterman强调企业在推进AI决策时“切忌让所有人进入决策室”，无需赋予公司内部每个部门投票权。精简而高效的跨学科指导委员会能够在技术遭遇瓶颈时迅速调整策略，避免因过度协商陷入僵局。同时，他在分享中提醒企业管理者，在敲定具体工具与技术供应商之前，必须首先在跨组织层面达成统一的战略共识并获得坚定的业务支持。 自动化测试与阶梯式渐进上线 (10:56) 针对语音智能体的非确定性特征，Kolterman明确提出企业级治理不能依赖主观随意的“感觉测试”（vibe checks）。Brinks Home在过去18个月中建立了全面的自动化测试架构，在每次功能上线前均使用数千个自动化测试用例对语音模型进行客观评估。一旦用例未能通过，技术团队会直接追溯具体的客观根因而非凭直觉推测，以此严密防范系统迭代对既有对话能力造成回归退化。 Kolterman打破了AI可以“一蹴而就”部署的误区，指出联络中心通话时长的削减呈现典型的“阶梯式”规律——在缓慢的爬坡积累后，往往伴随新功能上线迎来断层式跃升。此类生产级落地高度依赖基础工程的完备性，包括早期标准化的人类客服工作台、完备的API文档、规范化的CRM数据架构，以及持续将历史交互数据注入系统建立基线。在此基础上，团队还需依托长期的A/B测试不断修正交互细节。 生产级成效与人机协同演进 (13:41) Brinks Home投入生产的语音智能体“Veronica”取得了显著的业务成效：呼叫转接人工的比例从原先的30%骤降至8%，净推荐值（NPS）提升了30分，通话时长与联络中心服务分钟数缩减达50%。除了初期的效率指标，Sharma与Kolterman指出企业还应关注长期转型指标，例如上线后6至12个月内创造的增量业务收入，以客观衡量AI对企业增长的真正驱动力。 不过，Kolterman坦言Veronica当前的NPS表现相较人类客服仍有差距，数字交互往往欠缺人情味。为提升交互体验与处理复杂语境（例如提前识别技术人员上门调度），Brinks Home正与Cresta联手开发新型人机协同（HITL）架构。当虚拟智能体识别到交互障碍时，可在不中断通话或生硬转接的前提下向二级人工客服或主管“举手求助”；人类客服实时介入注入指导后由AI继续服务。该架构旨在结合超个性化记忆能力，在保留AI高效率的同时让人工精力聚焦于高价值交互瞬间。 当前企业AI智能体在交互亲和力与复杂场景微调上仍面临考验。Kolterman与Sharma总结强调，企业级AI绝非标准化的即插即用商品，高企的行业失败率意味着挑选具备实际落地经验的战略伙伴至关重要；CIO唯有将清晰的业务所有权、严谨的自动化验证与深层业务流程深度绑定，方能摆脱陷入试点泥潭的宿命。

## Background 该视频由科技峰会频道 Ai4 于 2026 年 8 月 14 日在 YouTube 上发布。内容记录了 Brinks Home 首席信息官兼数字化转型高级副总裁 Philip Kolterman 与 Cresta 现场首席技术官 Deepank Sharma 的专题对话，重点揭示了企业如何破除 AI 概念演示的表象，构建自动化测试治理框架并将 AI Agent 真正规模化部署至生产环境的实战指南。 ## TL;DR 对话深入剖析了企业推动 AI Agent 从试点走向规模化生产的底层逻辑，直面大量 AI 项目陷入停滞的行业痛点。Brinks Home 结合自身实践，阐释了业务主导维护、复用成熟质检基座以及算清平台方案风险收益比的决策考量。通过构建批量自动化测试体系，团队成功攻克了语音智能体的非确定性治理难题，使智能体“Veronica”将转人工率由 30% 降至 8% 并大幅缩减通话时长。未来双方将聚焦“人机协同（HITL）”与超级个性化架构，持续重塑下一代联络中心体验。 ## Atmosphere & style 整场对谈节奏明快、务实坦诚，Deepank 的幽默串联与 Philip 的深厚实战经验形成了绝佳的互动张力。没有浮夸的技术概念包装，对话穿插着自研踩坑的坦率自嘲与严谨的成本账目剖析，从工程治理到管理决策层层递进，传递出一种硬核、严密且极具一线启发性的技术落地氛围。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=5fROEdYNMY0&t=4s) 行业分析预测到 2027 年将有高达 40% 的 AI 试点项目走向失败，核心症结往往在于陷入漫长停滞的指导委员会讨论，无法跨越到生产落地。 - [02:29](https://www.youtube.com/watch?v=5fROEdYNMY0&t=149s) AI 成功落地的底层逻辑在于全公司对明确战略达成强烈共识，技术必须服务于业务战略，而非为技术本身挑选工具。 - [02:29](https://www.youtube.com/watch?v=5fROEdYNMY0&t=149s) 将 AI Agent 的日常维护与运营权直接赋予业务部门，IT 仅负责底层 API、框架与通信集成，能最大程度消除交付和维护中的组织摩擦。 - [05:22](https://www.youtube.com/watch?v=5fROEdYNMY0&t=322s) 虽然底层大模型单次调用成本极低，但自研面临极高的失败与工程风险；采购成熟企业级平台能将成本降低一个数量级，同时大幅降低试错风险。 - [05:22](https://www.youtube.com/watch?v=5fROEdYNMY0&t=322s) 管理虚拟坐席与人类坐席需复用同一套基础设施与自动化质检标准，实现平滑转接与协同，其工程体量甚至不亚于构建虚拟坐席本身。 - [08:33](https://www.youtube.com/watch?v=5fROEdYNMY0&t=513s) AI 落地并非一蹴而就，而是在夯实 API 与数据基座后，通过小步快跑的阶梯式迭代与 A/B 测试逐步释放业务价值。 - [11:18](https://www.youtube.com/watch?v=5fROEdYNMY0&t=678s) 针对语音智能体的非确定性，依托 Cresta 自动化测试批量运行数千测试用例，从依赖主观判断转向由数据驱动审批发布。 - [14:03](https://www.youtube.com/watch?v=5fROEdYNMY0&t=843s) 敏捷决策应避免将所有人拉入会议室造成决策瘫痪，需聚焦核心领导层建立互信，遇到阻碍时快速果断纠偏。 - [15:10](https://www.youtube.com/watch?v=5fROEdYNMY0&t=910s) Brinks Home 的 AI 智能体上线后实现转人工率从 30% 降至 8%、NPS 提升 30 点、平均处理时长减少 50%，但绝对满意度仍需持续优化以追赶优秀人工代表。 - [16:38](https://www.youtube.com/watch?v=5fROEdYNMY0&t=998s) 下一代联络中心将采用人机协同（HITL）架构，AI 遇到复杂卡点时主动举手由二级坐席后台静默指导，避免直接转呼与客户重复阐述。 - [18:07](https://www.youtube.com/watch?v=5fROEdYNMY0&t=1087s) AI 智能体并非即插即用的大宗商品，选择具备成熟实战经验与成功落地先例的合作伙伴是 CIO 规避失败风险的关键领航策略。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=5fROEdYNMY0&t=4s) 破除“AI 演示”假象与行业痛点：从试点走向生产 对话在轻松的打趣中开场。Cresta 的现场首席技术官（Field CTO）Deepank Sharma 调侃道，联络中心（Contact Center）的人通常很有趣，而 IT 人员往往比较严肃。紧接着，他将话题抛向台下观众：本周在会场里大家都看过了无数酷炫的 AI Agent 演示，有多少人觉得它们令人惊叹？台下纷纷举手认同。然而，Deepank 随即揭示了一个残酷的行业现实——来自 Gartner 和麦肯锡等机构的行业分析预测，到 2027 年，将有高达 40% 的 AI 试点项目走向失败的终局（"see the doomsday"）。 Deepank 介绍，自己的团队在 Cresta 负责协助众多企业客户将 AI Agent 真正落地部署到生产环境中。在此过程中，他们亲眼见证过大量成功案例，但也看到无数项目由于陷入漫长且停滞不前的指导委员会（steering committees）讨论而夭折。为了剖析试点与实际生产之间的差距，Deepank 邀请了 Brinks Home 的 CIO 兼数字化转型高级副总裁 Philip Kolterman 同台分享。Brinks Home 已经让 AI Agent 在生产环境中稳定运行了相当长的一段时间，成功跨越了绝大多数企业尚未攻克的难关。 Deepank 将这场对话划分为三个核心议题： - **从何处着手（Where do you start）**：自研与采购的权衡、供应商与合作伙伴的选择。 - **如何进行治理（Governance）**：如何在保护品牌声誉的同时实施质量控制。 - **如何规模化扩展与未来展望（Scale and the future）**：步入正轨后如何持续创造业务价值。 ```mermaid flowchart LR accTitle: AI Agent 落地三步曲 subgraph S1["第一阶段"] A["明确起点与战略定位"] end subgraph S2["第二阶段"] B["建立严格质检与治理"] end subgraph S3["第三阶段"] C["规模化扩展与持续迭代"] end A --> B B --> C ``` --- ### [02:29](https://www.youtube.com/watch?v=5fROEdYNMY0&t=149s) 业务主导与无摩擦维护：Brinks Home 的战略起点 在探讨“从何处着手”时，Deepank 询问了企业在面临技术转型时最为纠结的“自研还是购买”（Build vs. Buy）以及“选择普通供应商还是深度合作伙伴”的决策心路。Philip 回忆道，Brinks Home 与 Cresta 的合作已经持续了整整 5 年。这意味着双方的合作始于 ChatGPT 诞生之前（"pre-ChatGPT"）。虽然这五年间 AI 世界经历了翻天覆地的变化，但技术落地的底层逻辑始终未变：任何 AI 实施方案能够成功，前提必须是全公司范围内对目标达成强烈的共识（strong buy-in），且所有技术动作都必须服务于明确的企业战略，而后再围绕该战略挑选工具、合作伙伴与实现路径。 谈及具体执行的关键，Philip 强调了两个核心要点：理顺底层数据（getting the data right）以及透彻理解人工业务流程。企业必须先将现有的人工服务流程标准化并形成完善的文档，明确自身的品牌标准。此外，Phil 极为看重 Cresta 平台所带来的一种根本性转变——从早期的“坐席辅助”（Agent Assist）演进到全自动“AI Agent”的过程中，平台将维护与运营权直接交给了业务部门。 在 Brinks Home，IT 团队只负责提供底层支撑，例如 API 接口、技术框架与电话通信集成（Telephony），而完全不参与日常业务流程与系统数据的维护更新。Philip 总结道，让业务人员直接上手管理工具，彻底消除了产品交付和后期维护中的内耗与摩擦（"take a lot of the friction out of the process"），从而打造出更为出色的最终成果。 --- ### [05:22](https://www.youtube.com/watch?v=5fROEdYNMY0&t=322s) 自研幻觉与数量级成本账：平台方案的风险对冲 Deepank 指出，当下很多 CIO 经常产生一种顾虑：如今基础模型（Foundation Models）层出不穷，在展览馆里随处可见，模型本身已经变成了大宗商品（commodities），企业内部的工程师似乎直接拿来调用就能自研，为什么还要花钱采购平台？ Philip 坦率分享了 Brinks Home 早期走过的弯路。在接触虚拟坐席（Virtual Agent）初期，他们曾尝试基于 GPT-3.5 结合 AWS 等云服务进行自主开发。然而，团队拼凑出来的系统在能力上根本无法达到直面客户的标准，完全不敢上线。为此，Philip 算了一笔清晰的“数量级”（orders of magnitude）经济账： | 服务模式 | 单次通话成本量级 | 核心特征与风险收益比 | |:--- |:--- |:--- | | **人工客服处理** | > $10 / 通 | 成本高昂，处理时长受限，但具备人工应变能力 | | **底层自研模型调用** | 几美分至十几美分 / 通 | 账面极度便宜，但自研失败率极高，面临 40% 的死局风险 | | **企业级 AI 平台** | 处于中间的“美元级”（Dollar range） | 成本相比人工下降整整一个数量级，复用成熟经验，风险极低 | Philip 直言，纸面上从 10 美元暴跌到 10 美分听起来非常诱人，但现实中最致命的风险是“花费巨额精力却一无所获”（the biggest risk is that you don't get anything）。选择具备成百上千次成功落地经验的合作伙伴，可以用可控的成本将支出降低一个数量级，同时将失败风险降至最低。在当前技术极其复杂且充满不确定性的阶段，这才是绝大多数企业的理性选择。 更关键的是，即使自研成功，企业依然要面临后续庞大的配套工程挑战：如何衡量效果？如何保证服务质量（QA）？而在 Cresta 平台上，Brinks Home 管理人类坐席与虚拟坐席使用的是同一套基础设施——针对人类坐席打造的自动化质检（Automated QA）可以直接平移到虚拟坐席上；一旦虚拟坐席处理遇到瓶颈，系统能无缝将通话转接给使用原生界面的人类客服。Philip 指出，这套协同与转接体系的工程体量“丝毫不亚于甚至大于构建虚拟坐席本身”，是项目能够全面成功的核心支柱。 --- ### [08:33](https://www.youtube.com/watch?v=5fROEdYNMY0&t=513s) 夯实基座与阶梯式迭代：从标准化走向严苛治理 针对 Deepank 半开玩笑提出的“上线过程是否一帆风顺、无需准备基础工作”的假设，Philip 借用 Brinks Home 首席客户官的话调侃道：外界总以为一切都是“美满如蜜桃奶油”（peachy peaches and cream），但现实绝非如此。这项工程必须以小步快跑的方式“迭代式”（iteratively）推进。 Brinks Home 之所以能顺利落地，关键在于前期做足了功底： - 统一并规范了人工客服桌面体验与 IT 基础设施，构建了完善且文档清晰的 API。 - 提前将人类客服接入系统，沉淀了标准化的话术、知识库及呼叫应对流程。 - 人类坐席与虚拟坐席复用同一套管理分析工具，使得系统能直接继承成熟的质检标准。 在谈到成效体现时，Philip 形象地描绘道：如果观察虚拟坐席为联络中心缩减话务时长的图表，会发现它呈现出清晰的“阶梯式上升”（stair step）轨迹——在漫长的渐进爬坡期之后，随着新工具或新能力的上线实现一次跃升，随后再次平稳爬坡。这一过程需要不断进行 A/B 测试、修补系统漏洞并优化虚拟客服面对客户时的语气与沟通策略，使其更容易被用户接纳。 Deepank 对此深表赞同，他总结道，决定企业 AI 落地成败的关键分水岭，正是底层系统的就绪程度（API 完备度、CRM 数据架构、数据血缘与治理等）。随着基座与起跑线讨论的完成，对话随即过渡到了下一个关键板块：如何建立严格的**治理流程（Governance）**。对于以安防服务立命、承载着千家万户家庭安全信任的 Brinks Home 而言，在客户体验与满意度方面，容不得半点盲目的试错与冒险。 ### [11:18](https://www.youtube.com/watch?v=5fROEdYNMY0&t=678s) 摆脱主观判断：基于自动化测试的 AI 治理框架 在探讨语音智能体（Voice Agent）如何安全上线并承接真实生产流量时，Phil Kolterman 将治理流程划分为两大核心板块：跨部门战略治理与自动化测试验证。首先，治理并非 IT 单方面的职责，而是 IT 与业务团队之间的跨职能协作。团队必须聚齐关键业务角色，共同厘清系统边界、排查潜在逻辑漏洞，并在上线前确保业务逻辑严密无误。 在技术验证层面，过去 18 个月最关键的技术突破是自动化测试体系的建立。语音智能体天然具备非确定性（Non-deterministic）特征，加上语音流的复杂输入，过去极难进行高吞吐量的回归测试。Cresta 构建的自动化测试能力彻底改变了传统“听起来不错就上线”的主观评估模式，转变为完全由数据驱动的决策流程。团队每次部署更新时，都可以运行成千上万个测试用例，确保智能体行为与上次发布保持一致；一旦用例失败，团队能够精准下钻排查原因，评估其对生产环境的潜在风险。 ```mermaid flowchart LR accTitle: AI Agent 自动化测试与治理流程 subgraph S1["准备与设计"] A["业务与 IT 跨部门定义边界"] --> B["设计测试用例库"] end subgraph S2["自动化评估 (Cresta)"] B --> C["批量执行数千测试用例"] C --> D{"非确定性行为评估"} end subgraph S3["决策与发布"] D -- "测试失败" --> E["根因分析与风险评估"] D -- "测试通过" --> F["数据驱动审批上线"] E --> F end ``` - **非确定性测试突破**：彻底淘汰依赖人工拨打试听的偶发性验证，通过批量注入成千上万个自动化测试用例，提供可量化、可审计的治理数据支撑。 - **数据支撑决策**：坚守“以数据为决策基准，而非依赖个人观点”的原则，确保智能体迭代过程具备高度可复现性与稳定性。 --- ### [14:03](https://www.youtube.com/watch?v=5fROEdYNMY0&t=843s) 决策机制与敏捷纠偏：“不要把所有人都拉进会议室” 在谈到如何说服利益相关者并推进落地时，Phil Kolterman 提出了极具实践意义的管理洞察：“不要邀请所有人参会”。过多的表决权往往会导致决策瘫痪，企业应当聚焦核心领导层，在建立互信与战略共识的基础上迅速拍板。 Phil 强调，战略共识是所有敏捷决策的基石。在 Brinks Home 贯穿整个运营系统与 Cresta 合作的 6 次落地实施中，并非每一次尝试都一帆风顺。由于前期构建了清晰的战略对齐机制，团队能够在遇到阻碍和意外情况时快速调转方向、果断决策，使项目以最快速度重回正轨。 --- ### [15:10](https://www.youtube.com/watch?v=5fROEdYNMY0&t=910s) AI 智能体“Veronica”的业务成效与满意度挑战 Brinks Home 的语音智能体被命名为“Veronica”，取自公司首席客户官（Chief Customer Officer）Veronica 的名字（为了避免在日常沟通与会议中产生混淆，Phil 戏称智能体为“Ver-V”）。在全面上线后，智能体展现出了亮眼的运营数据： | 关键业务指标 (KPI) | 实施成效 | |:--- |:--- | | **转人工率 (Transfer Rate)** | 从 **30%** 大幅下降至 **8%** | | **净推荐值 (NPS)** | 提升 **30 个点** | | **平均处理时长 (Minutes Reduction)** | 减少 **50%** | 尽管指标表现抢眼，Phil 却坦诚指出了当前的技术局限：Veronica 的 NPS 绝对值依然不及优秀的人工客服代表。这一直是客户服务数字化产品（无论是 App 还是 Web 门户）长期面临的痛点——纯数字交互很难达到人类服务的满意度。为了突破这一天花板，团队正借助 A/B 测试持续优化智能体对用户情绪的镜像反馈（Mirroring）与复杂语境理解能力，力求让其表现真正变得“令人愉悦（Delightful）”。在实际通话中，Veronica 已经能够主动发现用户即将到来的设备上门排期（Dispatch）并提前介入，避免客户被迫经历死板的常规流程。 --- ### [16:38](https://www.youtube.com/watch?v=5fROEdYNMY0&t=998s) 下一代联络中心架构：人机协同（Human-in-the-Loop） 展望未来 6 至 18 个月的客户体验演进，Phil 与 Cresta 正在全力攻坚“人机协同（Human-in-the-Loop, HITL）”架构。在该机制下，当虚拟智能体在交互中遇到复杂或模棱两可的情境时，不再是粗暴地将电话转接给人工坐席，而是能够主动“举手”，请求主管或二级（Tier 2）人工座席介入评估。 ```mermaid flowchart TD accTitle: 人机协同 (Human-in-the-Loop) 实时指导工作流 A["客户拨打电话"] --> B["AI 智能体 (Veronica) 主动交互"] B --> C{"是否识别到复杂/潜在卡点？"} C -- "常规问题" --> D["AI 独立完成闭环"] C -- "遇到卡点 (主动举手)" --> E["二级坐席 / 主管后台静默介入"] E --> F["人工向 AI 提供指导指令"] F --> G["AI 吸收指令继续服务客户 (无需转呼)"] ``` 在这一模式下，人工座席可以在后台实时提供指导，帮助智能体修正沟通方向，使整个服务过程始终由虚拟智能体主导，客户完全不需要承受转呼排队或重复阐述诉求的痛苦。Phil 预测，未来的联络中心绝大部分通话时长都将由虚拟智能体承载，人类座席则聚焦在最关键的“黄金分钟”，进行精准干预与体验升华。 --- ### [18:07](https://www.youtube.com/watch?v=5fROEdYNMY0&t=1087s) 超级个性化愿景与 CIO 的落地建议 Deepank Sharma 进一步补充了虚拟智能体在超级个性化（Hyper-personalization）上的巨大潜力。理想的体验应当像拨打航空公司电话时，系统已完全掌握用户以往偏好（例如默认锁定靠窗座位），免去所有冗余询问，未来甚至能直接关联到上线后的业务增收等转型层面的核心指标。 在对在场 CIO 与技术领袖的结语中，Phil 强调了合作伙伴选择的决定性意义： - **拒绝技术同质化思维**：当前的大模型与 AI 智能体技术绝非即插即用的“大宗商品（Commodity）”，底层包含极其复杂的工程细节与交互微调。 - **选择经验丰富的领航者**：必须选择像 Cresta 这样真正具备成熟实战落地经验、“此前已经做成功过”的硬核伙伴，才能在企业 AI 转型的深水区中确保 60% 以上的高成功率。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=5fROEdYNMY0&t=4s) see the doomsday — Deepank Sharma > [02:29](https://www.youtube.com/watch?v=5fROEdYNMY0&t=149s) strong buy-in — Philip Kolterman > [02:29](https://www.youtube.com/watch?v=5fROEdYNMY0&t=149s) getting the data right — Philip Kolterman > [02:29](https://www.youtube.com/watch?v=5fROEdYNMY0&t=149s) take a lot of the friction out of the process — Philip Kolterman > [05:22](https://www.youtube.com/watch?v=5fROEdYNMY0&t=322s) the biggest risk is that you don't get anything — Philip Kolterman > [08:33](https://www.youtube.com/watch?v=5fROEdYNMY0&t=513s) peachy peaches and cream — Philip Kolterman > [08:33](https://www.youtube.com/watch?v=5fROEdYNMY0&t=513s) stair step — Philip Kolterman > [14:03](https://www.youtube.com/watch?v=5fROEdYNMY0&t=843s) 不要邀请所有人参会 — Philip Kolterman > [15:10](https://www.youtube.com/watch?v=5fROEdYNMY0&t=910s) Delightful — Philip Kolterman ## Mentioned Philip Kolterman, Deepank Sharma, Brinks Home, Cresta, Cresta Intelligence, Ai4, Gartner, McKinsey & Company, ChatGPT, GPT-3.5, AWS, Veronica (AI Agent) ## Video description Philip Kolterman - CIO & SVP of Digital Transformation, Brinks Home Deepank Sharma - Field CTO, Cresta Intelligence Every CIO has seen the demos. Far fewer have live AI Agents handling real customer calls at scale. In this session, Phil Kolterman, Chief Innovation Officer at Brinks Home, joins Cresta's Deepank Sharma to pull back the curtain on the two questions that separate pilots from production: how do you decide where to start, and how do you continue to scale and drive real business value once you're live? Phil shares how Brinks Home navigated both — from choosing the right first use case to building the automated testing framework that gave their governance council something measurable to approve, not just a demo to believe in. A candid, practitioner-level conversation for CIOs who are done theorizing and ready to ship. [Source](https://www.youtube.com/watch?v=5fROEdYNMY0)

[0:04] Deepank Sharma: see the doomsday [2:29] Philip Kolterman: strong buy-in [2:29] Philip Kolterman: getting the data right [2:29] Philip Kolterman: take a lot of the friction out of the process [5:22] Philip Kolterman: the biggest risk is that you don't get anything [8:33] Philip Kolterman: peachy peaches and cream [8:33] Philip Kolterman: stair step [14:03] Philip Kolterman: 不要邀请所有人参会 [15:10] Philip Kolterman: Delightful

[0:04] Deepank Sharma: 40% of all AI pilots are projected to reach a "doomsday" failure by 2027 [13:41] Philip Kolterman: Don't invite everybody into the room [13:41] Philip Kolterman: you don't need to give everybody in the company a vote.

[0:04] I think what I said was that uh usually we  do these they're contact center people and [0:08] contact center people are fun and it people are  not fun. So I kind of think about it. All right. [0:16] So I hope like all of you have seen a lot of AI  agent demos out here since Monday. How many of [0:23] you think they are just fabulous great work? Raise  your hands. Awesome. Yet there is a a a fact that [0:32] keeps staring on you know in our faces from all  of the experts in the industry from Gartner from [0:37] McKenzies of the world to say 40% of all of the  AI pilots are going to go and see the doomsday [0:42] in 2027. I am Dupang Sharma. I am field CTO at  Cresta. Uh we work with a lot of our customers [0:49] to deploy these AI agents live in production. I  have my team here who go build build build out [0:54] these AI agents for our customers and we've seen  a lot of pilots go successful and we have seen [1:00] a lot that have got stuck in the steering  committees to learn and talk more about it [1:06] uh please uh you know uh join me in welcoming Phil  Coleman who is the CIO at brings home his team has [1:13] AI agents running live in production for quite  some time now and he has figured out what maybe [1:20] most of us haven't so we're going talk through  the the process of what are those decisions that [1:25] that leads to success in these scenarios where  what goes down you know the bottom versus what [1:33] goes actually into production and we are able to  scale those of course I feel Phil like I would [1:39] try to divide this this this chat into like three  sections one where do you actually start two uh [1:45] once you've started it how do you govern it three  would be fine you know we have good governance [1:51] around around our our AIS and deployments. Now,  how do you actually scale these and what should [1:56] be the future look like starting with like let's  say where do you start? Right? Of course, whenever [2:03] there's a new technology initiative or or or a new  business transformation project that that happens [2:08] in in any of the companies, what we've heard is  there are a lot of build versus buy decisions. [2:12] There is lot of thinking about whether we should  go with this vendor versus that vendor, whether [2:17] we should look for a vendor versus a partner.  like help us walk us you know walk us through [2:22] your thought process and how did you come up with  a strategy around around these aspects. Yeah. Um, [2:29] you know, I I often tell people we've been with  Crest for 5 years, and you have to remember that [2:35] was pre-Chat GPT. So, the world has changed a  lot over that period of time. And the how do you [2:42] go about it now versus then is is probably a lot  different, but I think that there are some things [2:47] that are are consistent. Um, I think the most  important thing and it sort of transcends AI or [2:56] any other implementation is making sure that you  have uh strong buyin across the organization for [3:02] what you're trying to accomplish and that you  have a strategy in place and that what you're [3:06] doing is part of that strategy and then um then  you're picking tools and partners and approaches [3:13] around that around that strategy. Um but um you  know beyond that if you kind of move into this [3:20] particular implementation and and and what it  takes to make that successful I think there are [3:25] some really important keys to that. One of them is  um getting the data right and making sure that you [3:32] understand that you know what you're building  towards. Um, and a lot of that comes down to [3:38] understanding your your human process and making  sure that that you have that well documented, [3:44] that you know your brand standards and that um  that that's well understood. Um, one of the things [3:50] that we really liked about the Cresa product  really from the beginning and has continued [3:56] as the product has evolved as we've moved from um,  agent assist type functions to AI agent is the um, [4:03] the ability to really put the technology in place  and it is there's a there's certainly a technology [4:10] program aspect to it but then being able to turn  it over to the business and let the business [4:13] really run the tool. um my teams um from an IT  perspective have zero involvement in maintaining [4:21] the processes um you know maintaining the data  in the system um we provide APIs we provide um [4:30] technology frameworks we provide the telefanany  but um all of that is managed by the business and [4:35] I think that giving them a platform that they can  do that in effectively really makes for a better [4:42] end product because you take a lot of the friction  out of the process in terms of of uh getting the [4:49] uh getting the product out the door and continuing  to maintain it. That that that's perfect. So [4:54] So what you're saying is put the business into  the heart of that process, understand what the [4:59] business needs are and and create a frictionless  process of building that iterative value with that [5:04] system. And of course with you know the like the  way AI is evolving on a weekly basis now not just [5:10] you know uh monthly or yearly it it becomes easier  and easier u and and I totally you know 100% [5:17] resonate with that sentiment. Now the other other  thing that that that we hear from the CIOS the [5:23] push back that that usually comes is models are  marvels right u I I I've seen like so many models [5:31] in the in in the exhibit today and and and they're  all available as commodities that everybody can [5:35] use um what what is your thought process around  I could go and just build these things myself in [5:42] my set of engineers because anyways they are they  are you know available but what's the what's the [5:47] push to go towards a platform versus just building  things which are available as commodities you know [5:52] yourself. Yeah. I mean I will tell you that when  we first looked at virtual agent so we were uh we [5:58] were crested partner long before that but when we  first looked at virtual agent we tried to build [6:03] it ourselves and this was I believe it's GPT 3.5  and we were using Amazon services and we tried to [6:10] stitch something together um and we couldn't get  anything that was competent that we were willing [6:14] to put in front of customers and um as this space  has evolved um you know I've kind of I sort of put [6:22] it on a on a spectrum from um and it sort of works  out nicely on like orders of magnitude. Over on [6:28] one end you've got the cost of a call. So the cost  of a a human call handled call it's more than $10, [6:35] you know, on average for um our length of calls  to be able to handle a human- handled call. And [6:40] then if you go way over to the other side and  you sort of think about it on a model basis or [6:44] um maybe even running your own models, you know,  you can do that for pennies, tens of cents, [6:49] 10 cents. Um, and then uh I'll let Cresta talk to  you about pricing, but you know, to put it on an [6:55] order of magnitude basis, you know, it's in that  that dollar range in the middle. And so it fits [7:00] kind of right in the middle of that space. And um  so you know, you you kind of look at you go, I go [7:06] from $10 all the way to 10 cents. That that sounds  awesome. But the reality is that to your point, [7:13] your statistics earlier, the biggest risk is that  you don't get anything. That you have a you have [7:18] a pilot that fails. And here you can go all the  way in order of magnitude down with relatively [7:25] little risk with someone who's done it hundreds  of times. That's pretty good, you know. And [7:30] um and I would argue that at this moment in time  that it is hard enough and there is enough risk [7:37] in the process that that is for the majority  of companies the way to implement this thing. [7:43] Um I would also point out that even if we would  have been successful rolling out our own agents [7:50] uh you know the next question would have been  how do you measure it? How do you ensure that [7:54] it's that the quality is there? How do you keep  tabs on it? And um one of the great things that we [8:00] have in the process we have in place where we use  the same system to manage the human agents as we [8:04] use to manage the virtual agents. All the QA, all  the work we did to automate uh quality assurance, [8:10] all of that applies to the virtual agents. All the  technology to handoff calls if the virtual agents [8:15] not successful. The human agents working in their  native experience. Um there's a ton of value in [8:21] that. That's that's every bit as big a project as  building the virtual agents, if not more so, and [8:27] uh super important to the overall success of your  project. Thank you. Thank you for sharing that. [8:33] Now I'm hoping like when you started building  these virtual agents, everything was a breeze. [8:38] You had all systems ready. You did not have to do  any homework or or were there like you know a lot [8:43] of foundational work that your team had to do.  Our chief customer officer often says that about [8:48] it that everything's just peaches peachy peaches  and cream. Um no I mean I think that you have to [8:55] approach these projects iteratively and you have  to um we had a lot of advantages because we had [9:01] we had spent a lot of effort building our human  agent desktop experience and normalizing the IT [9:08] infrastructure around that making sure we had  well doumented APIs um and then also uh we had [9:14] a lot of advantages because we had put the humans  into the Cresta system first um all of the work [9:20] that that um the teams had done to normalize  the knowledge and um standardize our processes, [9:28] document how we want to approach calls. Uh the  system that we use is great because um all the [9:34] tools that we apply to human agent management,  all of that applies to virtual agents. So, [9:39] a lot of advantages to getting there, but it's  still an incremental project. And it's still, [9:43] if you look at the graph of um the minutes that  she takes out of the contact center, you know, [9:49] it is still a stair step with slow climbs and  then big jumps as we put something new in and [9:54] then and then stair steps. Um, and a lot of that  is around enabling tooling. It's um it's AB test [10:01] work to uh work out kinks in the system, improve  her approach with customers so that they're more [10:07] accepting with her. Um, all of that applies here  and I think an incremental approach just like with [10:12] most IT development is the right way to to go  about putting this in place. Get your get your [10:18] key wins early and then keep moving. Makes perfect  sense and that's exactly a theme that we see with [10:24] most of our customers that that they spend a lot  of good work and that's actually what what stands [10:30] in between you know success and failures that if  your systems are ready or you've spent enough time [10:37] in in making your systems API ready CRM data  architectures uh data lineage rolling out AI [10:45] agents becomes a little easier task as compared to  you know if you you don't have anything. Thank you [10:51] for for for sharing that. Now, we've built out  the systems. We learned where to start. Let's [10:56] roll into what does what does governance look  like? Uh I'm I'm pretty sure you have like a [11:02] tight governance process. Brings home security  is one of the most reputed brands in terms of [11:06] u a company which we trust to to protect our  homes. Now if and I'm sure like there's no room [11:13] for experime exper experimenting on the customer  experience customer satisfaction for rolling out [11:18] a voice agent but you have a voice agent that  you rolled out confidently in production which [11:23] is taking live traffic uh still the seat is is  pretty good and we're going to see the numbers in [11:28] some time what does your governance process looks  like what makes you confident and and how are you [11:33] evolving that governance process like what kind  of testing protocols or testing sequences that [11:38] you have in your in your in your life cycle of  of rolling out these changes. Yeah. You know, [11:43] I'll put it I'll put it kind of in into  two spaces. One is sort of governing [11:47] um what we're going to do and how we're going to  do it. And that is really an interdisciplinary [11:52] project between the IT organization and the  business organization. And you have to have [11:57] the right voices in the room to make sure that  um what you're doing is appropriate and that [12:03] um that it's well understood that you've found the  edges and the holes and understand where there can [12:09] be problems and making sure that it's it's tight  before you move forward. Um but then you mentioned [12:15] and it's one of the one of the things that we've  seen real growth in the in this in the product [12:20] over the last um 18 months is automated testing.  you know, when um voice agents initially came out, [12:29] they're they're very difficult to test. Um you  know, they're non-deterministic in in general, [12:34] but then you add the voice component on top of it  and um it's hard to feed volume test cases. Crest [12:41] has done a great job building out that capability  so that it's a datadriven discussion. It's not a [12:47] well, you know, called it, it sounds good and and  move on. Um you we can run thousands of test cases [12:53] against it. we can make sure that the last time  we deployed it, she still works the way that she [12:58] did the last time. And if a test case fails, we  can go evaluate that test case and understand why [13:03] it failed and if it's acceptable to to continue  on. Um I think moving from particularly in this [13:11] where it's a shouldn't say the same thing every  time it's non-deterministic. It's important that [13:17] that you've got real data to back up your your  governance decisions and understand what you're [13:21] trying to do. I think very important point that  that don't ground your judgment on opinions but [13:27] ground it on data. I mean something that we  we we've been saying for years. Uh and that [13:31] thing still applies no matter what like if if  in case you're building things on AI non AI [13:36] u if you have data to back up your decisions that  always works. Perfect. Now like let's say that's [13:42] that if that's the governance process u you you  you touched a little bit upon the the right you [13:49] know voices in the room. How do you how do you how  are you able to convince like everybody? Is it is [13:54] it again like you know you have the right data to  to prove out the use case? You have the right data [13:58] to say this is how it works. Uh or are there like  some some gotchas that that you want to share with [14:03] the audience? Um you know I guess the first thing  I would say when you when you ask that question is [14:09] don't invite everybody into the room. I I think  that um you need to have the right leaders in in [14:16] the discussion so that um you can make decisions  quickly, but you also don't need to give everybody [14:21] in the company a vote. And I think it's important  that you establish those relationships and have [14:26] that as foundational. Again, back to how you need  to start with the right strategy in place and make [14:31] sure that that um everybody is aligned with what  you're trying to accomplish because if you can do [14:36] that, then that makes those governance discussions  a lot easier. Um, and it makes it a lot easier to [14:42] move quickly and pivot quickly because there will  be you we've done six implementations with Cresta [14:50] over our entire um operational system space and  not all of those have gone perfectly, you know, [14:58] and it's important that you're able to pivot and  move quickly when something doesn't go smoothly [15:04] and make decisions um to get back on track. So  cool. Let's let's look at some numbers and and [15:10] you know I I hear the the the name of the AI agent  that you have is Veronica and Veronica is live. Um [15:18] is that is that something that you named after  one of your teammates? Yes, she's right there. [15:22] Chief customer officer Veronica. That's it's very  confusing. By the way, I tend to call her Veriev [15:28] because not this Veronica, the other one because  uh it's confusing if uh if she's on a call and [15:34] you're talking about the call. So that's awesome  and we we we do see these numbers that were shared [15:40] by your team uh that that you see uh you know  transfer rate reduction from 30% to 8%. You see an [15:47] NPS uh upside of 30 points and you see reduction  in minutes of 50%. That's all great, right? That's [15:55] all all of the results that that you want to see.  Now if I have to ask you what is it that you want [16:00] to achieve next? What do you think that that your  CX would look like? maybe six months from now, [16:06] what is it that your team is now aiming to do? Um,  you know, I think there are a couple things. Um, [16:12] first of all, uh, when I look at those numbers,  they're they're great. I will tell everyone in [16:17] the room, uh, Ver's C, uh, MPS score is not  as good as our human agent MPS scores. And, [16:23] and this is sort of like I feel like bedeled my  entire career, um, working on technology products [16:30] for customer service, the digital experiences. It  doesn't matter if you're talking about the portal [16:34] or the app, they're just not as satisfying for  customers. And I do think that the virtual agent [16:40] holds that opportunity. But she has to be better  at mirroring customers. She has to be better at [16:46] handling nuance. Um, she has to, the word I use  a lot is delightful. And she's got to become [16:51] delightful. And we focus on that a lot with  AB testing and trying to move her her forward. [16:58] She's getting much more nuanced in terms of  experience. And you know, I I I run into calls [17:05] all the time and I I [ __ ] my head and I'm very  impressed with her where she will notice that a [17:11] customer has a dispatch coming up and bring that  up early and not force the customer to go through [17:17] some wrote process. Um but she has a ways to go on  personality. So I think we're working on that. Um, [17:23] the other thing that I'm really excited  about that we're working on with CREA is um, [17:28] what I characterize as human in the loop. So, it's  the ability for the uh, virtual agent to raise a [17:35] hand and ask for a supervisor or a tier 2 agent  to evaluate the call and then provide guidance in [17:42] the call and get the call back on track so that  the customer doesn't have to have a transfer and [17:46] doesn't have to leave the virtual agent. I believe  that that's where the contact center um over the [17:52] next 6 months, 12 months, 18 months ultimately  I think that's where the minutes in the contact [17:57] center go is that more and more minutes will be  handled by the virtual agent. Humans will still [18:02] be involved in the process but only in the minutes  that matter to uh keep a call on track and delight [18:07] the customer. That's actually quite powerful.  I think I think we we we did talk about this [18:12] uh yesterday and and a week ago which is virtual  agents able to hyperpersonalize the experience. [18:19] Um and I would wait for that day when like let's  say when I call up my airlines and I say I want [18:22] to book a flight and they know exactly last time  I actually booked a window flight window seat and [18:28] they don't ask me that question. They should just  know. uh but yes I think I think I think u and I [18:33] feel like these matrix are are kind of strategic  ma matrix but but there are transformational [18:38] matrix that that we can define something like you  know revenue that we are getting after deploying [18:44] the voice agent after like 6 months 10 months down  the line or 12 months down the line but but those [18:48] are some some great learnings Phil thank you so  much for for for uh sharing your experience what [18:54] the brinks team is doing this is this is something  all of us can definitely learn from are there any [18:59] words for the CIOS out here that you want to  share to say if they want to still stay on that [19:05] six you know 60% successful rate that that MIT's  Garters and Mckenzies of the world keeps keeps [19:11] uh you know touting about what is it that they  should or should not do like one I mean I I I [19:16] said it earlier I think at this moment it who  you're working with really matters um Crest [19:21] has been a great partner for us um but it's very  important to have strong partners and people who [19:27] have quote unquote done it before techn technology  is not a commodity at this point. There's there's [19:33] a lot of uh nuanced work to be done in this  space. So, thank you so much. Yep. [applause]