Audible, State Dept Rein in AI Agents as Shadow Automations Mount

影子自动化激增，Audible与美国国务院加强管控AI Agent 企业领导者正通过将概率性语言模型封装在具有严格“人类在环”检查点的确定性工作流中，以抑制失控的成本和高企的安全风险。 专家小组探讨了企业如何从实验性的AI“狂野西部”阶段，转向结构化、集中式的企业治理以及限制概率黑箱模型的确定性工作流。来自Audible和美国国务院的高管阐述了他们如何借助托管沙盒、“人类在环”审查关卡以及基础设施层级的成本监控，在确保安全的同时保持生产交付速度。与会者指出了切实的运营障碍——从拥有过高权限且缺乏监管的桌面Agent，到沙盒内失控的Token开销以及AI生成的面条式代码。最后，发言嘉宾强调，推动技术普及依靠的是同行案例展示、众包功能投票和主动技能提升等文化赋能手段，而非自上而下的行政命令。 企业级AI正迅速从实验性的“狂野西部”阶段，转向集中式治理、编排以及信任优先的架构。 (0:04) 自主软件Agent往往在内部系统间以危险的高权限运行，若在缺乏IT监管的情况下于本地部署，将带来严重的运营风险。 (5:48) 成熟的组织通过采用配备凭据保管库、身份模拟和限定范围沙盒的托管连接器框架，以规避影子IT和脆弱的自动化。 (8:14) 为遏制未经授权的消费级AI的使用，美国国务院部署了直接连接内部外交档案、安全且契合特定任务需求的定制工具（如State Chat）。 (11:31) Audible通过允许业务部门“凭感觉编码”（vibe-code）快速构建原型，同时要求中央工程团队对生产代码进行加固与扩展，实现了部署速度同比提升50%。 (15:47)

As enterprise artificial intelligence moves beyond an uncontrolled experimental phase, organizations are shifting toward centralized governance and structured infrastructure to balance development velocity with operational safety. Panelists from Audible, Tray.ai, Nividous, and the U.S. Department of State argue that long-term enterprise value depends on encasing probabilistic models within deterministic architectures, managing administrative credentials, and cultivating cultural adoption through user-led enablement. Containing Elevated Permissions and Shadow IT Risks (5:48) Alan Hester warns that autonomous agents running on local employee workstations frequently operate with higher system permissions than human workers would receive, creating systemic operational vulnerabilities across organizations of any size. Drawing parallels to past waves of fragile spreadsheet macros and rogue scripts, Josh Noble notes that mature organizations are converging on standardized connector frameworks to contain shadow IT. This architecture relies on managed OAuth identity impersonation and centralized token vaults rather than raw API keys stored in local files. Without centralized boundaries, unmonitored experimentation can result in severe architectural and financial setbacks. Noble cites an advisory board case where an untrained employee operating in an unmonitored sandbox configured an automated script to poll internal systems every 30 seconds rather than running a daily synchronization, burning through $20,000 in API tokens within three days. Enforcing Deterministic Guardrails and Output Verification (18:49) To counter the risks of non-deterministic systems, Josh Noble advocates restricting probabilistic language models to discrete subtasks encased within deterministic state machines. In standard back-office workflows like invoice processing, routine transactions follow structured programmatic rules, engaging autonomous agents only to navigate complex exceptions such as missing purchase order numbers or unmatched vendor addresses through traceable, auditable branching logic. In mission-critical diplomatic environments, Amy Ritualo notes that the U.S. State Department couples technical export gates with an intern-style review model. Its Co-Drafter tool, designed to satisfy more than 300 style requirements, highlights classification headers in red and requires users to manually remove draft labels in Microsoft Word before documents can be distributed. Similarly, an internal search system indexing the 20,000-to-25,000-page Foreign Affairs Manual presents model summaries alongside direct source excerpts, training personnel to treat AI outputs as preliminary drafts requiring line-by-line verification. Balancing Development Velocity with Code Maintainability (15:47) AI-assisted tools can accelerate delivery, as demonstrated by Audible achieving a 50% year-over-year increase in production deployments while maintaining human-in-the-loop validation across code reviews and merges. However, accelerated vibe coding introduces technical debt and maintenance uncertainty. Josh Noble cautions that code generation tools default to bloated, monolithic files—such as single massive script files—unless guided by developer coaching and standardized context bundles that enforce modular architecture. Calculating return on investment requires moving beyond ad-hoc desktop metrics. Tim Martin explains that Audible governs financial exposure using AWS-level permissioning and spending alerts to prevent runaway token usage. Rather than tracking raw employee utilization rates, Audible anchors three of its six annual enterprise Objectives and Key Results directly to AI initiatives, evaluating success by increased engineering throughput and expanded content delivery for customers. Fostering Adoption Through Co-Design and Peer Enablement (32:36) Overcoming cultural resistance requires moving away from top-down mandates toward collaborative workflow integration. Amy Ritualo explains that the State Department crowdsourced its AI roadmap via an internal Thunderdome voting process, inviting staff across global posts to rank 20 feature proposals three times annually. To build trust, regional practitioners such as financial management officers lead peer-to-peer demonstrations of practical tools, supported by a global network of AI Champions sharing mission-level use cases. Audible applies a similar approach by hosting quarterly showcases across diverse business units and embedding dedicated AI Accelerators within functional teams. Alan Hester observes that broader workplace sentiment has evolved from initial fears that automation would eliminate jobs toward recognizing AI as an augmentative tool, requiring leaders to construct clear upskilling pathways. While generative tools significantly lower the barrier to automating enterprise tasks, panel participants emphasize that successful adoption requires strong engineering scaffolding rather than unrestrained autonomy. The ongoing operational challenge lies in sustaining rapid experimentation by non-technical teams while enforcing strict central controls to prevent unmaintainable code, runaway infrastructure costs, and compliance failures.

来自Audible、Tray.ai、Nividous以及美国国务院的技术领袖一致指出，企业人工智能应用正加速脱离自由试验的“狂野西部”阶段，全面转向由中心化治理、确定性系统与严格权限管理主导的新范式；转型的核心在于既要释放开发效率与自动化潜力，又要有效防范未受管束的自主智能体引爆系统性风险与长远技术债务。 权限失控与影子IT的收敛架构 (5:48) Nividous总裁Alan Hester强调，企业级AI面临的最隐蔽风险在于自主软件智能体往往被赋予远超同岗位人类员工的系统权限。若放任员工在本地工作站随意运行未经审查的智能体，将使企业陷入无法监管的真空，一旦缺乏统一编排与集中监控，极易引发灾难性的系统宕机。Tray.ai现场CTO Josh Noble将此现状比作十年前非官方脚本与Excel宏失控泛滥的影子IT危机，并指出领先企业正自发走向“趋同演化”：逐步摒弃将API密钥直接保存在本地配置的危险做法，转向通过集中式连接器框架实施统一治理。 成熟的架构模式要求通过OAuth等协议进行基于身份认知的权限模拟，使智能体严格继承特定用户的受限操作权限，而非直接获取全局管理特权。此外，业务人员应在隔离沙盒内快速验证业务原型，而生产环境的代码加固、API速率限制与合规审计则统一交由中央IT团队负责。Noble以其客户咨询委员会的真实案例警示缺乏防护的代价：一名未经训练的员工在未受监控的沙盒中编写了每30秒轮询三到四个企业系统的脚本，仅用三天就消耗了高达2万美元的API Token费用。 确定性工作流约束黑箱智能体 (18:49) 针对概率性大语言模型的不可控“黑箱”特性，Josh Noble指出企业走向成熟的关键是主动收缩授予AI的操作边界。虽然企业早期的冲动是尝试将全流程交给端到端的自主智能体，但高效能组织正在迅速纠偏，仅将大语言模型限制在不可或缺的离散子任务中，并将其严格封装在具备确定性状态机与可验证审计追踪的框架内。 在复杂的后台业务中，这种结合尤为重要。以发票处理为例，当发票信息与采购订单完全匹配时，系统遵循既定的确定性流程自动录入企业ERP；只有当遭遇无订单号或发件邮箱异常等非标状况时，才会触发自主智能体进行异常评估。即便处于异常分支，智能体也只能依据明确且可审计的规则链条执行特定操作（例如联络采购员或通知主数据团队），确保每一步决策都有迹可循，杜绝概率性决策在核心业务中自由脱轨。 严格的人机协同与“实习生”核查机制 (21:33) 在涉及重大战略与日常运营的高风险场景中，技术防线必须与明确的人机协同审查机制深度绑定。美国国务院首席数据与AI官Amy Ritualo指出，由于行业数据显示约97%的员工在缺乏官方工具时会转向未经批准的消费级AI，国务院选择主动在安全计算隔离区内推出State Chat以及公文起草工具Co-Drafter。为确保外交备忘录等文件严格符合包含300多条排版与格式规范的指引要求，系统在导出Word文档时强制将密级标头标记为鲜红色，并要求人工手动核验并删除“由Co-Drafter生成”的提示标签，以此强化使用者的直接责任。 面对厚达20,000至25,000页的《外事手册》（FAM），检索系统通过双栏界面并排展示AI摘要与对应原文及引用链接，并要求外交人员将AI输出视作“需要逐行严格复核的实习生初稿”，严禁不经审查直接采信。类似地，Audible首席技术官Tim Martin说明，虽然Audible的生产部署频次同比提升了50%，但在关乎核心听众体验的音频流生产链路中，从系统架构设计、代码审查到合并上线，必须全程保留严格的人工介入核准流程。 编码规范、投资回报与组织文化赋能 (27:07) AI辅助编程（vibe coding）的普及给传统的软件开发生命周期（SDLC）带来了长期的技术债务不确定性。Josh Noble指出，代码生成模型极易堆砌出难以维护的单体巨型文件（如臃肿的main.ts），而多数企业尚未针对大模型特性开展基础工程培训。他建议团队提供标准化的系统提示词与Markdown上下文文件包，将企业架构标准前置注入大模型，从源头确保生成模块化的高质量代码。 在投资回报（ROI）与财务管理方面，Audible通过AWS平台内置的权限与告警机制限制算力开销，并在公司6项年度OKR中将3项明确与AI赋能挂钩，促使评估重点从单纯的算力调用量转向业务交付成果。在文化推动上，美国国务院每年三次通过“Thunderdome”投票机制让全员参与筛选20项高价值功能，并依托驻法兰克福财务官等一线骨干建立“AI Champions”网络推广实际操作经验；Nividous总裁Alan Hester总结认为，企业文化已从早期的“机器人抢饭碗”恐慌逐步过渡为以人机协作为导向的角色升级，转型的核心最终取决于对内部人才的持续赋能与技能重塑。 整个讨论表明，企业推进AI转型的瓶颈已从单纯的模型能力转向系统架构与治理机制的重构。如何在鼓励非技术人员敏捷创新的同时，有效遏制快速生成代码累积的长远维护成本与权限失控隐患，并建立起兼顾算力成本与开发效益的治理框架，依然是各类机构规模化落地AI技术必须应对的长期考验。

## Background 该视频由频道 Ai4 于 2026 年 8 月 22 日发布在 YouTube 平台上。内容为一场高管圆桌讨论，汇聚了来自 Audible、美国国务院、Tray.ai 及 Nividous 的技术高管，重点探讨企业 CIO 和 CAIO 如何在告别无序探索后，通过系统治理、风险控制、ROI 追踪和组织文化变革来推动 AI 的生产级落地与价值实现。 ## TL;DR 圆桌探讨了企业 AI 落地正从野蛮生长的“西部荒野”转向强调系统化治理与人机协同的成熟阶段。嘉宾们分享了应对“影子 IT”、Agent 权限失控及不可预测“黑盒”风险的实战方案，提倡将概率性大模型限制在特定节点，并结合确定性工作流与强制人工审查。同时，各方还剖析了软件工程生命周期重构下的 ROI 量化难题、Token 预算失控惨痛教训，以及如何通过沙盒防护、标准化规范和自下而上的员工赋能机制化解组织文化阻力。 ## Atmosphere & style 整场讨论务实、克制且极具工程实战沉淀，完全褪去了早期对 AI 的盲目狂热与概念炒作。四位来自不同背景的高管在主持人引导下推心置腹，既坦诚拆解了 3 天烧掉 2 万美元 Token 的真实翻车事故，也生动描绘了从最初的失业恐慌走向人机协同赋能的文化转变。现场氛围既严谨专业又充满建设性，让观众在直面企业 AI 治理复杂性的同时，清晰感受到技术迈向确定性生产落地的成熟与稳健。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=4s) 企业 AI 落地正在经历关键拐点，从分散盲目、缺乏监管的“西部荒野”探索阶段，全面转向集中治理与系统化编排的新阶段。 - [06:04](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=364s) Agent 权限失控与新一代影子 IT 是核心隐患，成熟企业通过统一连接器隔离敏感凭证，让中央 IT 承接速率限制与并发底座以保障安全。 - [11:39](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=699s) 美国国务院主动推出专属 State Chat 解决员工私用商业 AI 的问题，并深度打通外交电报数据库与政策事前检验分析（Premortem Analysis）。 - [16:35](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=995s) Audible 在交付流程中坚定推行“人在回路”（Human-in-the-loop）审查，在确保生产级安全的前提下使发布吞吐量提升了 50%。 - [19:12](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=1152s) 化解大模型“黑盒”风险的核心是减少概率性操作的占比，仅在非结构化文本处理等环节调用 LLM，主体业务依托具备确定性审计轨迹的工作流。 - [21:33](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=1293s) 应对高风险公文输出时，国务院通过 Co-Drafter 强制植入人工审查“阻力”（如显式标签与手册段落左右双栏对照），杜绝盲目采纳 AI 初稿。 - [27:07](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=1627s) “氛围编程”（Vibe Coding）重塑了软件开发生命周期并增加了长期维护成本，ROI 衡量正从单纯的“工具采纳率”转向“全新工作模式与算力优化”。 - [32:36](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=1956s) 克服非技术人员的文化阻力需推行“与用户共建”，利用 Thunderdome 功能投票竞逐、同行答疑会与全球 AI Champions 标杆网络自下而上驱动转型。 - [40:31](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=2431s) AI 工具拥有巨大的“爆炸半径”，若在沙盒中缺乏工程调用规范指引（如高频轮询接口），极易造成数天内消耗数万美元 Token 的预算灾难。 - [41:48](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=2508s) 为避免模型生成无法维护的臃肿单体代码，企业可向 Claude/Codex 提供标准化的 Markdown 架构规范文档，在开发起点对齐工程规范。 - [42:45](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=2565s) 企业 AI 变革的最终落脚点在于组织文化引导，从最初被机器取代的失业恐慌，平稳过渡到以人为本的人机协同与员工能力升级。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=4s) 开场介绍：从野蛮生长到体系化 AI 治理 主持人 Grace DeGraaf（哈佛商学院美国中部研究中心研究员、前《商业内幕》科技编辑）开场指出，当前企业 AI 应用正在经历关键拐点——从最初各自为政、缺乏约束的“西部荒野”（Wild West）式盲目探索，转向需要更多系统化架构与集中治理的成熟阶段。随后，四位来自不同领域的企业及机构技术高管分别介绍了各自的业务背景与当前 AI 策略： Audible 首席技术官 Tim Martin 强调，如今借助各类 AI 工具，技术构建的门槛已降至历史最低点；但作为全球有声书领域的标杆品牌，Audible 最核心的底线是在加速创新的同时捍卫客户信任，确保所有进入生产环境的 AI 功能都具备高度的安全性和一致性体验。 Tray.ai 现场首席技术官（Field CTO）Josh Noble 分享了公司的技术演进历程：Tray.ai 最初专注于连接器框架（Connector Framework）以实现跨系统数据同步，随后拓展至自动化与业务工作流，并逐步融入自然语言处理与 AI 能力。当前他们最新的方向是帮助企业将前端通过 Claude 或 Codex 等工具“氛围编码”（Vibe-coded）生成的应用原型，无缝部署并锁定到后端更为严谨、确定性的工作流之中。 Nividous 总裁 Alan Hester 介绍，Nividous 是一家专注于智能自动化与智能体 AI（Agentic AI）的平台公司，支持在统一平台中协同管理 AI 智能体、业务流程与人类员工。他回顾道，如果说过去一年整个行业都在狂热讨论“智能体将接管世界”，那么现在的核心焦点已经演进为：“既然人人都能创建 Agent，我们该如何对其进行有效编排与治理，确保它们赋能业务而不是毁掉业务”。 美国国务院首席数据与 AI 官（CDAO）Amy Ritualo 负责领导国务院分析中心（Center for Analytics）。她指出国务院的 AI 战略聚焦于两大核心目标：一是利用 AI 消除行政杂务与繁重负担（Administrative toil），让员工工作更轻松；二是为派驻全球各地的外交官赋能，利用 AI 打造信息决策优势（Decision advantage）。 ```mermaid flowchart TD accTitle: 企业 AI 落地演进阶段 subgraph S1["第一阶段：狂热探索（Wild West）"] A1["门槛极低 / 自由构建"] --> A2["单点 Agent 爆发"] A2 --> A3["影子 IT 与权限隐患"] end subgraph S2["第二阶段：集中治理与协同（当前阶段）"] B1["连接器与身份隔离 (OAuth/RBAC)"] --> B2["确定性工作流 + 审计追踪"] B2 --> B3["赋能业务 / 生产级落地"] end S1 -->|"体系化收敛演化"| S2 ``` --- ### [06:04](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=364s) Agent 权限失控与影子 IT：企业治理的核心痛点 在讨论企业为何迫切需要收紧 AI 治理策略时，Alan Hester 指出了一个普遍被忽视的安全盲区：许多 AI 智能体在配置时被赋予的权限往往远高于执行同类工作的人类员工。当一家拥有数百甚至上万名员工的企业中，每位员工都能随手在桌面上配置数十上百个未受集中监管的 Agent 时，企业就面临着巨大的系统性与灾难性风险。 Josh Noble 将这一现状与十年前企业 IT 面临的困境进行了生动类比。当年业务团队追求业务流速，向中央 IT 申请自动化需求时，往往得到“排期需要 18 个月”的答复；为了自救，业务人员私下编写了大量不受控的脚本和宏，最终这些临时拼凑的“意大利面条式代码”反倒成了支撑关键业务的脆弱支柱。如今生成式 AI 领域正在重演这种“影子 IT”的历史：员工若无法通过合规渠道调用系统，就会直接将 API 令牌写进本地配置文件，让未经审计的本地模型在公司内网肆意操作。 Josh 观察到，当前业内最成功的企业正在不约而同地走向一种技术层面的“趋同进化”（Convergent evolution / 类似生物学中的蟹化现象）： - **底层凭证隔离**：中央 IT 搭建统一的连接器框架，将 API 密钥与系统凭证安全封存在数据存储区中，或利用 OAuth 让 Agent 仅以用户自身身份在限定范围内进行模拟操作。 - **划定沙盒边界**：赋予业务专家在安全沙盒内自由探索、构建 POC 的最大自由度，充分发挥其业务领域专长。 - **中央团队兜底生产要素**：由 IT 团队接管底层的速率限制（Throttling）、身份认证、并发控制等复杂工程问题，加速业务想法从原型向生产环境的转化。 --- ### [11:39](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=699s) 美国国务院的实践：State Chat、外交电报与事前检验分析 Amy Ritualo 详细分享了美国国务院在高度复杂且全球化分布的信息敏感环境下的 AI 治理与应用实践。国务院拥有遍布全球的 270 多个使领馆和代表处，日常高度依赖信息流转与人际沟通。面对行业调查中显示高达 97% 的员工在缺少官方工具时会私下使用商业 AI 的现实，国务院选择主动出击，在受保护的安全环境中推出了专属的“State Chat”。 为了超越简单的聊天对话并真正嵌入外交工作流，国务院对 State Chat 进行了深度定制与业务整合： - **打通外交电报系统（Cables System）**：将 AI 深度连接至使领馆与华盛顿之间最核心、最丰富的外交电报数据库，大幅缩短外交官查找、验证和汇总全球信息所需的时间。 - **事前检验分析工具（Premortem Analysis）**：专门研发了用于政策文件压力测试的 AI 评估工具。该工具改变了以往出现危机后才追问“哪里出了问题”的被动复盘模式，而是在多边谈判或政策发布前主动模拟演练“各种潜在失败场景与应对短板”。 - **协同专业团队共建合规边界**：以全球公共外交中的核心需求——“翻译”为例，技术团队并非直接开放通用 AI 翻译，而是与专业语言服务团队（Language Services Team）深度协作，细化不同语境与敏感级别下的 AI 翻译规范，确保关键涉外表述精准传达。 --- ### [16:35](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=995s) 创新与管控的平衡：Audible 的人机协同与生产交付提速 针对宽松实验与严格合规之间的张力，Audible 首席技术官 Tim Martin 阐明了 Audible 在客户体验与内部研发中采取的双重策略。由于 Audible 是数百万用户通勤和日常生活中的核心应用，平台必须保证生产环境交付的代码具备极致的可靠性与体验一致性。 在面向终端客户的软件开发流水线中，Audible 坚定推行“人在回路”（Human-in-the-loop）： - **全流程人类把关**：从前期的架构设计到底层的代码审查（Code Review）与上线发布，关键节点均由工程师严格把控。 - **生产发布提速 50%**：尽管保留了严格的审查机制，在 AI 工具的辅助下，Audible 的整体研发流速显著提升，生产环境部署量实现了同比 50% 的大幅增长。 - **打通上下游瓶颈**：随着编码环节的大幅提速，安全审查等传统瓶颈环节也逐步引入智能体机制进行初筛与辅助，确保从创意到客户价值交付的全局畅通。 在内部业务创新方面，市场与内容等业务团队正积极借助 AI 进行“氛围编码”（Vibe coding）来搭建提效工具。业务团队拿着自己调通的原型主动找到技术团队寻求生产化改造，形成了业务播种、IT 团队赋能落地并确保安全合规的良性共创生态。 --- ### [19:12](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=1152s) 破除“黑盒”风险：减少概率性操作与强化确定性工作流 关于如何化解大语言模型（LLM）“黑盒”机制带来的不可预测性风险，Josh Noble 总结了领先企业的演进规律：许多尚未成熟的公司仍处于“盲目狂热地让 Agent 处理一切”的阶段，而最成功的企业则在逐步精简 AI 的触点，主动减少系统中“概率性工作”（Probabilistic work）的占比。 成熟企业的治理共识在于明确大模型的能力边界： - **精准锚定 LLM 触发场景**：仅在非结构化文本解析、意图识别等必须依赖大模型的环节使用 LLM，其余环节全面依托具备明确逻辑的确定性工作流（Deterministic workflows）。 - **保留完整确定性审计轨迹（Deterministic Audit Trail）**：确保系统能清晰追踪每一次分支跳转的具体原因，以便于回溯排查。 - **规范智能体决策权限**：在发票处理（Invoice Processing）等典型企业业务中，严格限制 Agent 随意决定分支路径的权限，防止不可解释的黑盒决策对企业财务或核心业务流程造成实质损害。 ### [20:40](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=1240s) 确定性工作流与异常分支追踪 在讨论复杂业务流程的自动化时，演讲者以发票处理为例深入剖析了 AI Agent 与确定性规则的结合。常规流程中，系统提取发票信息、匹配采购订单（PO 号），并将数据写入 ERP 系统。然而一旦遇到没有 PO 号的发票，Agent 就需要根据具体上下文决定应对路径：如果是 "no-reply" 无回复邮箱发来的，Agent 需转而联系采购员寻找供应商对接人；如果是发票地址与 ERP 不一致，则需要对接主数据管理团队。通过构建这类确定性工作流，即便给予 Agent 在异常路径中自主选择决策的权力，企业依然能够完整回溯和审计 Agent 每一步操作的具体动因。 ```mermaid flowchart TD accTitle: Invoice Processing Exception Workflow subgraph S1["标准输入与处理"] A["收到供应商发票"] --> B{"是否有 PO 号？"} B -- "是" --> C["提取数据并写入 ERP"] end subgraph S2["Agent 异常决策分支"] B -- "否" --> D{"检查发件邮箱 / 地址"} D -- "无回复邮箱" --> E["联系内部采购员确认"] D -- "地址不匹配" --> F["转接主数据管理团队 (MDM)"] end ``` --- ### [21:33](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=1293s) 美国国务院的 AI 治理：双栏对照与强制人工审查 美国国务院首席数据与 AI 官 Amy Ritualo 针对 AI 输出的“黑盒”问题和高风险场景分享了治理经验。国务院从技术工具与人员培训两个维度构建治理闭环： 在技术层面，国务院推出了隶属于 StateChat 工具集的 `Co-Drafter`，用于起草信息备忘录、行动备忘录、谈话要点卡及背景清单等官方文件。这些公文面临超过 300 项格式规范要求。为了杜绝人员未经核实直接采纳 AI 生成的内容，技术系统被刻意设计了强制性“阻力”： - 导出到 Word 后，文档密级以醒目的**红色**标出，用户必须手动核验并修改分类。 - 文档顶部带有红蓝相间的 `Created by Co-Drafter` 标签，必须经过人工通篇审查后方可手动移除。 - 针对长达 20,000 至 25,000 页的《外交事务手册》（Foreign Affairs Manual），系统在检索界面提供**左右双栏并排展示**：右屏显示 AI 摘要，左侧直接列出原始手册段落与链接，确保用户可逐条比对。 在人员层面，国务院明确将 AI 定位为“实习生产出的初稿”，而非可以直接交付的终稿。通过下属的国立外交培训学院（National Foreign Affairs Training Center），AI 实践课程被系统性嵌入到经济官、管理官等各类职业技能全生命周期培训中。 --- ### [27:07](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=1627s) 软件开发生命周期重构与企业级 ROI 衡量 关于企业如何追踪 AI 的投资回报率（ROI）及控制 Token 预算消耗，Nividous 总裁 Alan Hester 指出，ROI 的衡量应划分为产品研发与后台桌面办公两个维度： 1. **软件工程层面的 ROI 模糊性**：AI 重塑了传统的软件开发生命周期（SDLC）。以往团队协同写代码、熟知每一行逻辑细节的模式，正逐渐被以“氛围编程”（Vibe Coding）为代表的人工智能辅助开发替代。这种转变给后续的系统维护、Bug 排查与代码生命周期成本带来了巨大的不确定性，导致产品端 ROI 的量化极为复杂。 2. **桌面办公分散化的风险**：如果企业内有 10 个员工做同一种工作，每人都用 AI 搭建了一套独立的工作流，就会产生 10 种不同标准、错误率各异的工具。这亟需在软件和 Agent 之上建立一层企业级统一治理框架，以实现可控的产出与成本度量。 Audible 首席技术官 Tim Martin 补充了 Audible 借助 AWS 体系进行的治理实践：通过严格的权限划分与告警监控，防止非技术人员无节制消耗 Token；对于技术开发团队，则设定“提高吞吐量的同时降低算力开销”的目标。在公司层面的 6 个年度 OKR 中，有 3 个直接围绕 AI 展开，推动评估重心从去年的“工具采纳率（Adoption）”向今年的“全新工作模式（Embracing new ways of working）”演进。 --- ### [32:36](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=1956s) 以人为本的 AI 落地：Thunderdome 机制与全球 AI Champions 在非技术人员占多数的组织中推行 AI 面临着显著的文化阻力。Amy Ritualo 透露，早期经常听到员工反馈“我主管说用 AI 是作弊”或是“我是财务管理官，这东西不适合我”。针对这一挑战，美国国务院采取了“与用户共同构建（Build with users）”的策略： - **Thunderdome（雷霆争霸）功能竞逐**：技术团队先筛选出 20 项候选功能，随后面向全机构用户发起投票，每年开展约 3 次，完全由一线员工票选最迫切需要开发的顶尖功能。 - **同行交流与 Office Hours**：员工往往排斥数据科学家或工程师的说教，但愿意倾听同行经验（例如法兰克福前线财务官分享实际工作流）。国务院为此设立了定期的答疑交流会（Office Hours）。 - **AI Champions（AI 冠军）网络**：由于国务院在外交层面的特殊敏感性，该项目无法命名为“AI Ambassadors（AI 大使）”，因而定名为“AI Champions”。这一机制在外派使领馆选拔积极探索 AI 的技术官与业务骨干，组建跨使馆的工作组并交流落地用例。 Tim Martin 同样分享了 Audible 举办的季度 GenAI 展示会（GenAI Showcase），为 HR、市场及工程师等处于不同采纳阶段的团队提供一个不仅展示成果、也能畅谈失败经验的安全交流空间。 --- ### [40:31](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=2431s) 沙盒失控与 3 天消耗 2 万美元 Token 的教训 Tray.ai 现场 CTO Josh Noble 从实战角度剖析了缺乏治理指导的负面案例。虽然企业内部提倡为员工提供沙盒环境自由创新，但在其主持的客户顾问委员会（Customer Advisory Board）上，各类由于缺乏最佳实践培训导致的“恐怖故事”屡见不鲜。 - **高昂的“学费”案例**：某客户企业的一名员工在沙盒环境中进行自动化探索，在短短 3 天内烧掉了 **$20,000 美元** 的模型 Token 费用。 - **事故根因**：该员工并未采用常规的每日单次同步（Daily Sync），而是设置了让脚本每隔 30 秒轮询三到四个核心系统。 - **治理启示**：功能强大的 AI 工具具备极大的“爆炸半径（Blast Radius）”，企业如果只发放工具权限却不提供基础的工程与调用最佳实践指导，极易引发严重预算灾难。 ### [41:48](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=2508s) 建立 AI 基础教育与开发规范：从单体代码陷阱到 Prompt 作弊码 在探讨研发团队使用 AI 辅助编码时，演讲者指出，如果缺乏规范指引，直接让模型生成代码往往会带来维护灾难——例如模型很容易把所有逻辑一股脑塞进一个庞大的单体 `main.ts` 文件中。团队需要通过小型的内部培训与指导来纠正这种习惯，明确告知工程师：“Claude 或 Codex 倾向于用最简单直接的方式生成代码，但后期面对这种庞大的单体结构将完全无法维护。” 目前几乎绝大多数企业在这一类“AI 基础工程素养与规范教育”上都存在缺失。其实解决这一问题的门槛并不高，团队完全可以提前准备一套标准化的 Agent Markdown 说明文档或设计规范模板，直接投喂给 Claude 或 Codex。这样一来，AI 在生成代码时就会自动遵循企业内部的设计标准与架构模式，工程师也无需时刻死记硬背每一条技术规范，相当于在开发起点就为员工和 AI 都装上了合规的“外挂作弊码”。 ```mermaid flowchart LR accTitle: AI 编码规范交付流程 subgraph Guide["企业标准输入"] A["标准 Markdown 规范库<br/>(架构/设计约束)"] end subgraph AI["代码生成阶段"] B["AI 编程模型<br/>(Claude / Codex)"] end subgraph Out["最终交付质量"] C["结构化、易维护代码"] D["避免巨型单体文件<br/>(如 main.ts)"] end A --> B B --> C B --> D ``` - **单体代码陷阱**：AI 默认倾向于生成最容易跑通的代码，常生成臃肿、难以拆解与维护的大型单体文件。 - **Agent Markdown 规范文档**：通过预设设计规范提示词包，让模型自主对齐企业工程架构规范，减轻开发人员的心智负担。 --- ### [42:45](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=2565s) 克服组织文化阻力：从“AI 抢饭碗”转向“赋能员工升级” 随着圆桌讨论接近尾声，主持人将话题交给了跨多个行业赋能客户的 Nividous 总裁 Alan Hester，探讨企业在落地 AI 时如何应对文化阻力以及员工对新技术的疑虑。Alan Hester 回顾道，最初大家的第一反应往往是恐惧——“机器人要来抢我们的工作了”（"the bots are coming for our jobs"）。当时很多员工陷入焦虑，甚至觉得“我的工作其实没那么难，AI 肯定能轻易取代我”，从而觉得自己随时面临被淘汰的风险。 如今大多数企业已经逐渐走出这种恐慌心理，并开始形成共识：AI 的真正价值绝非简单粗暴地裁员替代，而在于“让员工变得更优秀”（making people better）。当前的重中之重是为员工搭建完备的支撑与赋能体系，帮助他们实现职能转型，从原有的日常重复劳动平稳过渡到更有价值的理想工作状态。正如 Alan 所总结的，推动 AI 变革取得成功的终极关键，归根结底是要“深入了解你的员工”，并与他们并肩找到一条共同向前的转型之路。主持人随后致谢并宣布圆桌结束，邀请意犹未尽的观众在场外继续交流。 - **心态演进轨迹**：由早期的“替代焦虑与失业恐慌”，逐步转变为“人机协同与能力放大”。 - **成功转型的核心**：了解员工诉求，提供配套的培训与支持体系，协助岗位向更高价值层级平稳演进。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=4s) 既然人人都能创建 Agent，我们该如何对其进行有效编排与治理，确保它们赋能业务而不是毁掉业务 — Alan Hester > [06:04](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=364s) 底层凭证隔离：中央 IT 搭建统一的连接器框架，将 API 密钥与系统凭证安全封存在数据存储区中，或利用 OAuth 让 Agent 仅以用户自身身份在限定范围内进行模拟操作。 — Josh Noble > [19:12](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=1152s) 精准锚定 LLM 触发场景：仅在非结构化文本解析、意图识别等必须依赖大模型的环节使用 LLM，其余环节全面依托具备明确逻辑的确定性工作流 — Josh Noble > [21:33](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=1293s) 实习生产出的初稿 — Amy Ritualo > [27:07](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=1627s) 如果企业内有 10 个员工做同一种工作，每人都用 AI 搭建了一套独立的工作流，就会产生 10 种不同标准、错误率各异的工具。这亟需在软件和 Agent 之上建立一层企业级统一治理框架 — Alan Hester > [32:36](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=1956s) 我主管说用 AI 是作弊 — Amy Ritualo > [40:31](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=2431s) 功能强大的 AI 工具具备极大的“爆炸半径（Blast Radius）”，企业如果只发放工具权限却不提供基础的工程与调用最佳实践指导，极易引发严重预算灾难。 — Josh Noble > [41:48](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=2508s) Claude 或 Codex 倾向于用最简单直接的方式生成代码，但后期面对这种庞大的单体结构将完全无法维护。 > [42:45](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=2565s) the bots are coming for our jobs — Alan Hester > [42:45](https://www.youtube.com/watch?v=n1bPCzyz5s0&t=2565s) making people better — Alan Hester ## Mentioned Grace DeGraaf, Tim Martin, Josh Noble, Alan Hester, Amy Ritualo, Audible, Tray.ai, Nividous, 美国国务院 (U.S. Department of State), 国务院分析中心 (Center for Analytics), 哈佛商学院 (Harvard Business School), 商业内幕 (Business Insider), Claude, Codex, State Chat, Co-Drafter, Foreign Affairs Manual (外交事务手册), National Foreign Affairs Training Center (国立外交培训学院), AWS, Thunderdome, AI Champions ## Video description Alan Hester - President, Nividous Tim Martin - CTO, Audible Amy Ritualo - Chief Data and AI Officer, U.S. Department of State Josh Noble - Field CTO, Tray.ai Grace DeGraaf - Writer & Researcher, Harvard Business School How are today’s CIOs and CAIOs driving real AI value? In this executive panel, top enterprise tech leaders will share practical strategies for AI implementation—from aligning with business goals to managing cross-functional teams and vendors. We’ll explore how companies are governing AI use, tracking ROI, and overcoming cultural resistance. Hear firsthand how leaders are adapting their organizations for an AI-powered future, and leave with insights to better navigate infrastructure choices, compliance pressures, and strategic planning. [Source](https://www.youtube.com/watch?v=n1bPCzyz5s0)

[0:04] Alan Hester: 既然人人都能创建 Agent，我们该如何对其进行有效编排与治理，确保它们赋能业务而不是毁掉业务 [6:04] Josh Noble: 底层凭证隔离：中央 IT 搭建统一的连接器框架，将 API 密钥与系统凭证安全封存在数据存储区中，或利用 OAuth 让 Agent 仅以用户自身身份在限定范围内进行模拟操作。 [19:12] Josh Noble: 精准锚定 LLM 触发场景：仅在非结构化文本解析、意图识别等必须依赖大模型的环节使用 LLM，其余环节全面依托具备明确逻辑的确定性工作流 [21:33] Amy Ritualo: 实习生产出的初稿 [27:07] Alan Hester: 如果企业内有 10 个员工做同一种工作，每人都用 AI 搭建了一套独立的工作流，就会产生 10 种不同标准、错误率各异的工具。这亟需在软件和 Agent 之上建立一层企业级统一治理框架 [32:36] Amy Ritualo: 我主管说用 AI 是作弊 [40:31] Josh Noble: 功能强大的 AI 工具具备极大的“爆炸半径（Blast Radius）”，企业如果只发放工具权限却不提供基础的工程与调用最佳实践指导，极易引发严重预算灾难。 [41:48] Claude 或 Codex 倾向于用最简单直接的方式生成代码，但后期面对这种庞大的单体结构将完全无法维护。 [42:45] Alan Hester: the bots are coming for our jobs [42:45] Alan Hester: making people better

[0:04] Alan Hester: bring our business down [0:04] Josh Noble: bubble gum and duct tape [8:14] Josh Noble: crabification [11:31] Amy Ritualo: State Chat [21:33] Amy Ritualo: Created by Co-Drafter [32:36] Amy Ritualo: Thunderdome [32:36] Amy Ritualo: AI Champions [41:48] Josh Noble: giant main.ts file [42:40] Alan Hester: the bots are coming for our jobs

[0:04] Hi everybody. Um, welcome to leading the AI  transformation. I am Grace Stgraph. I am a [0:12] researcher with Harvard Business School's mid US  research center. So we cover the middle region [0:20] of the country, not California or the coast. Um,  and before that I was an editor for several years [0:27] at Business Insider where I covered all sorts of  industries including um, including tech and did [0:36] some AI coverage there as well. So I am excited  for um, this conversation we're about to have. Um, [0:44] for those of you who were at the um, keynote  presentations, I think um, there there was a [0:51] little bit of conversation about how right now  we've been shifting from this wild west of AI [1:01] implementation to a more centralized um, you  know, companies are realizing that there needs [1:08] to be a little bit more system, more systems,  more structure um in place um in order to have a [1:16] cohesive strategy around AI. So I'm excited to get  into that with everybody here today. Um so first [1:26] um I would each of you all to just give a  brief introduction about uh who you are, [1:32] your company or organization um and a brief  summary of your current approach to AI. Um [1:40] let's start with Tim. Great. Thank you, Grace.  Uh, my name is Tim Martin. I'm CTO of Audible and, [1:47] uh, you know, speaking of or governance is comes  to mind immediately when we talk about the AI [1:53] transformation. Now that the barrier to build  with AI enabled tools has never been lower. [1:59] uh with a beloved brand like Audible, the  worldwide leader in audiobooks, we need to make [2:04] sure that things that are reaching production,  are preserving customer trust and providing new [2:10] and cohesive experiences for for customers. So,  it's all about speed for us, uh but doing it in a [2:17] very safe way for our customers. All right, this  Okay, um yeah, so Josh from Trey AI. Um so I'm [2:27] Trey's field CTO. I meet with our customers, kind  of bring new features, functions, capabilities [2:33] that we're we're working on to the forefront  with them, discuss, get their their feedback, [2:37] help them work through how to solve kind of their  their hardest automation problems with sometimes [2:44] bubble gum and duct tape. Um, and then bring that  that that feedback back to our team to to improve [2:49] our product. Um, Trey is a overall company. Um,  we started as a company that was building kind [2:54] of a connector framework. So, hey, I want to sync,  you know, data from multiple different systems. We [2:59] had that in place. Then we extend that to uh build  automation and workflows eventually bringing AI, [3:06] natural language processing into those workflows.  And our latest development has been, hey, [3:10] we've got a bunch of customers that now have  these automations and these workflows on the [3:13] back end that are running, but they also want  to have an improved, you know, app experience, [3:18] user experience, and why not use something like a  a cloud or a codeex or what have you to to build [3:23] out those experiences. So, we have a new um new  product in in place that lets you easily deploy, [3:29] you know, those types of vibecoded um apps  into production and lock those down into [3:34] more deterministic workflows that are running  behind the scenes. My name is Alan Hester. I'm [3:41] from Nitus. We are an intelligent automation and  agentic AI company. We have the capacity to from a [3:51] single platform allow you to manage your agents,  your workflows and your humans. The thing that [3:58] we're seeing happen with a lot of our customers  related to AI. If if you were here last year, [4:03] you probably remember that a lot of the talk was  about agentic and how everything was becoming [4:10] agentic and how agents were taking over the world.  And now we've kind of gotten past that point where [4:19] okay, everyone can make agents. There are a lot  of agents out there. How do we orchestrate and [4:25] govern all these agents so that they don't bring  our business down so that they're helpful, not [4:30] harmful. So I think that's that's what the current  focus is on. Good morning. I'm Amy Roualo. I'm the [4:38] chief data and AI officer at the US Department of  State. I lead the center for analytics uh at the [4:45] department and our mission really is to bring  data and AI to the mission of the department. [4:51] So I focus on data governance, AI governance  um deploying enterprise uh applications and [5:01] um data management uh structure uh and  and data and analytics infrastructure. [5:09] My focus right now for the department is really  um twofold. So we're looking at how can we use [5:16] um AI to reduce administrative toil? How can  we make it easier on our employees to do their [5:22] jobs? And then the second area of focus is  really how can we um support our diplomats [5:29] across the globe and using AI to bring a decision  advantage uh to to the mission of the department. [5:37] So I'd like to start with kind of the problem that  you all have seen when it comes to governance. Um [5:48] Alan, I'd like to start with you. What have  you seen in the organizations that you all [5:55] are working with that has prompted having  tighter guidance and governance strategies? [6:04] Yeah, I think as I was mentioning in the intro,  as agents get out there into your organization, [6:13] one of the things that not everyone is aware of  is that agents in many cases have more elevated [6:21] permissions than you would give to humans if  they were doing that same job, right? And so [6:26] what that means is not only are they out there and  essentially being unmanaged uh at least centrally [6:35] unmanaged but but they have the capability to do  things that could potentially be uh harmful or [6:42] even catastrophic for your organization if you  don't have some control over that. And I think [6:49] for organizations that that just think about  their size, no matter what it is, if you have a [6:54] 100 people, if you have a,000 people, if you have  10,000 people, if each one of those employees has [7:01] 10 or 50 or 100 agents on their desk available to  them and you have no control over that centrally, [7:10] then how as an organization can you ensure  that you're moving in the right direction? [7:15] But one other twist I'll throw on to the end of  that is that not only is it this problem of the [7:23] agents being out there and and kind of with  elevated permissions potentially, not always, [7:29] but potentially, but but also the the fact that  uh everyone has the ability to do this from their [7:38] desktop in in many cases without it buyin or  permission or approval. Many companies are just [7:46] starting now to figure out what kind of policies  they want to put in place and how they want to [7:52] go forward in terms of controls. And because of  that, as you said, Grace, it's kind of like the [7:58] wild west. Josh, could you share what you've seen  in terms of what companies have run into in ter [8:07] um you know, without having a cohesive governance  approach, what what issues are are you all seeing? [8:14] Yeah. So, I'd say if I look back over the last few  years, we're seeing a similar journey that we've [8:21] we've repeated this this thing before. So, we  think 10 years ago, we went and saw business teams [8:28] that wanted to run as fast as possible and they'd  go to it and say, "Hey, it I I need you to create [8:32] this automation or link these systems for things  to together for us." And it said, "Great, we'll [8:37] get to that. It'll be 18 months." And the business  needed to run faster. And so they were building [8:42] rogue scripts and macros that became these  loadbearing pieces of architecture across the [8:48] organization and people left and turned into kind  of the spaghetti code across the or organization. [8:52] We're seeing a lot of that same trend with an AI  right now. So a couple years ago if you're at this [8:57] conference it was let's all build you know build  with AI and you know then eventually last year was [9:03] agents are going to run everything for us. Now I'm  seeing what seems to be across our customer base [9:08] because nice thing is I have quite a few customers  we get to work with. So, I get to kind of survey [9:12] a lot of a lot of customers is it's it's um it's  kind of like with the um it's like crabification [9:18] or whatever where everybody the most successful  companies are actually starting to all kind of do [9:22] the same thing even though the fact they're not  even talking to each other about this. So, your [9:26] your most successful companies are starting to add  that that centralization and and in governance in [9:32] in place. So, for example, authentications. Um, I  don't want somebody going and grabbing their API [9:38] token and throwing that into, you know, a config  file on their local machine, having Claude start [9:43] doing a lot of things across the organization.  But if you don't give them a really easy ways [9:47] to connect to their systems, they're going to do  exactly that. So instead, you give them something [9:52] like a connector framework where you say, "Hey,  we're going to hide the OS in a secure data store [9:56] or we're going to use your OTS, your OOTH, and  have the agent actually impersonate you against [10:02] that system, but we're only giving you access to  access these certain systems with these certain [10:06] levels of permission, but go wild within that  sandbox and create and prototype and give us [10:12] kind of the best PC's that you can think of for  your job because you're the you're the subject [10:17] matter expert in your your field. We don't need it  going and building every BRD, every PDD thing out [10:23] there with you being able to do your job. Show  us what you want within these constraints and [10:27] then as a central team, a central team in it can  go and say, "Cool. Now, we're going to take care [10:32] of not having to make sure you know everything  about development. You we're going to take care [10:36] of those offs. We're going to take about take care  of um you know, throttling. Uh we're going to take [10:42] care of all those those other things so that you  can just go and create and give us the ideas and [10:45] we'll go and serve you up faster." So seen a lot  lot more of that. What was I was trying to think [10:50] of the word for convergent evolution. There we go.  Like where where sloths somehow grew in different [10:55] areas of the world all at the same time. Um  yeah. So okay Amy, I know the US Department [11:03] of State is operating in a very different context  than you know the other panelists on stage. Um I [11:10] would love to hear how you all are approaching AI.  what what are you using AI for right now? And in [11:18] this context of of governance, there's always  going to be risks for when AI makes mistakes, [11:26] does things it's not supposed to do. So, how how  are you thinking about that? How is the State [11:31] Department thinking about that element of hero AI  implementation? Sure. Uh I'll try to cover all of [11:39] those things um and keep it relatively short.  So the state department is a is probably one of [11:46] the most uh globally distributed and information  intensive uh organizations. So we have over 270 [11:54] missions, posts, embassies uh across the globe  uh all who operate uh really on information and [12:03] engagements uh with people. So there was a real  opportunity with generative AI um for the state [12:10] department um and we were pretty uh we we leaned  in to the opportunity. Uh our folks could use [12:19] uh open-source um commercial uh AI tools if  they were doing it in an uncla on unclassed [12:27] information. Um but as we heard this morning  in one of the keynotes about um people you 90 I [12:32] think it was something like 97% of employees were  using AI even though they weren't supposed to. Um [12:37] we knew that that wasn't enough. So we needed to  provide tools that were in a secure environment [12:43] that allowed people to put secure uh information  uh into it. And so we we um really embraced that [12:50] opportunity. So, we have a tool. It's called uh  State Chat, which is our own um chatbot within a [12:57] secure environment. It's not uh we knew that it  it we didn't we weren't going to stop with just [13:02] a chatbot. Chatbot's super easy. Um right, like  it it's super helpful for our folks. They spend [13:07] a lot of time locating and um analyzing and uh  validating and then summarizing information. [13:17] And we need we knew that our employees needed to  connect to the state department's information. So [13:23] one of the biggest um sources of information at  the state department is is our cable system which [13:28] is a communication system be uh is a communication  uh between our embassies and uh Washington. It's a [13:37] extraordinarily rich uh set of information. So we  connected um our our uh chatbot to cables. We also [13:45] knew that we needed to adjust to our employees  workflows. So um one of the things that we [13:52] wanted to do in trying to give our uh diplomats a  decision decision advantage was really to provide [13:58] them with tools that they could use to analyze  and summarize tools. So uh summarize um policy [14:05] um papers or stress test a policy paper for a  negotiation or or something we were going to share [14:12] in a multilateral uh organization. And so we built  a tool uh specifically to help conduct premortem [14:19] analysis. So instead of uh instead of waiting for  an event to happen where we ask um what went wrong [14:25] and what could we have done better, this this  tool is really meant to to say um here's here's [14:32] what could go wrong in this type of scenario.  So so really we wanted to provide the tools that [14:39] people could use. We wanted to make them relevant  to their particular jobs and and workflows and and [14:46] at the same time just to just to touch on the  governance piece is as we were building this we [14:51] were thinking about governance. So um translation  is a big uh use for our our uh diplomats across [15:00] the globe our public affairs officers across  the globe. So but we needed to think about [15:05] what's appropriate use in translation. We have  professional translators. We have um we have their [15:11] uh information that is particularly important  to have um the right translation with the right [15:16] nuance in in language. And so we worked with our  language services team to develop a policy around [15:22] how we can use uh AI tools for translation.  So we think about uh governance uh as we're uh [15:30] ex as we're experimenting and we work closely  with our teams um across the state department [15:38] to ensure we have the the right governance in  place. Thank you Tim. Um one thing that we've [15:47] talked about is balancing experimentation and  innovation. you know that's one of the benefits [15:55] of having you know maybe a looser policy around  AI that's what a lot of companies say right is [16:03] um it gives you a chance to have more you know  innovation come out of those experiments um so [16:13] but then there's a tension between that and having  policies and and systems around ensuring AI is not [16:22] going to do something that's going to harm the  business, that's going to damage reputation. So, [16:29] how are you at Audible thinking about balancing  those that tension between experimentation and [16:35] governance with AI? Yeah, great. Thanks, Grace.  Uh, so I think of this in in two ways. So, [16:42] one is our customer experience. um where audible  is a a daily ritual for many of our customers part [16:51] of their commute. Uh we need to make sure that  the things that we're delivering to customers are [16:57] scalable and secure and coherent across the entire  experience. So we have um uh still have humans [17:05] in the loop for things that reach production.  They're at the front end with architecting these [17:10] things all the way to doing the code reviews and  pushes to production. We're seeing our velocity [17:16] significantly up in that workflow up about 50%.  Production deployments are up 50% year-over-year. [17:23] So things are getting through to customers uh  much more quickly and it's required uh other [17:28] bottlenecks as they appear in the system whether  it's what to build or the security reviews doesn't [17:34] matter how quick the build goes you still  need to go through the security review so [17:38] they're also becoming uh agentic in the ways that  they're doing things so it's all about increasing [17:43] that flow of idea to customer value um the other  part is internally uh folks are vi Josh what you [17:52] a vibe coding tool. They're looking at ways to  do their jobs more effectively. Um making sure [17:59] that that those things are also done securely.  And what we're seeing is that uh the the people [18:04] that are uh building these tools and vibe coding  these tools, they want to do their job better, [18:09] faster, so they can do more for customers.  And it's allowing them to express things to [18:14] the IT partners uh at Audible is the tech team and  hey, I have a need here. Here's a tool. Here's a [18:20] prototype. here's vibe coded. Uh can you help  me do this better? Um so we're very much in [18:26] collaboration with other parts of our business,  the marketing team, the content team, uh to to to [18:32] build the tools that they they plant seeds with to  actually bring them to to to production quality. [18:38] So one element of governance that you know some of  us have have spoken about before is how to address [18:49] the blackbox nature of AI where maybe you get an  output and you're not sure what exactly happened [18:57] between your input and that output. Um, and  that poses risks in a lot of different settings, [19:04] especially when you're not sure if something went  wrong or how to check that. So, um, Josh, I'd [19:12] like to ask how you all are thinking about from  a governance perspective that blackbox element. [19:20] How how are you how are you thinking about that  in terms of you know protecting protecting the [19:26] company making sure that nothing um that you  don't want an output that is problematic is is [19:33] not going to cause damage in some way. [clears  throat] I think it comes down to I mentioned [19:38] earlier many of our customers the most successful  ones are converging on a on a similar way of of [19:43] doing things and some of the least less successful  companies out there haven't gotten quite to that [19:47] point. It's very much we're all vibe curious and  everybody go build as fast as possible. It's just [19:52] encouraging that versus the okay, let's really  think about this. And when those companies start [19:56] to mature up, what I'm realizing is they're  actually reducing the touch points of AI. So [20:01] reduce the amount of probabilistic work going  on where early on it was let's just get agents [20:05] to do everything. This is this is awesome. Um,  and now it's wait no. If we want to scale this, [20:10] we actually should only use LLMs in the exact  spot that they should they should be used and [20:16] have as much where we can do deterministic  workflows, have a deterministic audit trail, [20:20] be able to trace and track exactly why things  went down different branches. Now you have agents [20:26] in there that you've now built out workflows  for and tools for and they can go in certain [20:31] instances pick and choose their their own journey.  So you take an example like invoice processing. [20:36] I get an invoice into my inbox. I don't want an  agent going and pick and choosing random ways to [20:40] go and process that invoice. It should extract the  information off the invoice, match it up to the PO [20:45] number, punch the data into my ERP system where  necessary. But what happens if an invoice comes [20:51] in that doesn't have a PO number? Now I've got a  little different problem. I have I can build out [20:57] tools that says, "Hey agent, depending on what  that email is, um maybe it's a no reply email, [21:03] so you need to go reach out to the buyer if that  happens." and find out who we should contact at [21:07] at the supplier that sent us the invoice or um  maybe you need to reach out to our master master [21:12] data ma management team because the ERP doesn't  have address information that matches what's on on [21:17] the invoice. In that point, I can build really  deterministic workflows. And yes, I'm giving [21:23] putting an agent in place to pick and choose what  happens in those exception, you know, pathways, [21:28] but I can still trace that back to why it did  the thing that it that it did. [snorts] Um yeah. [21:33] So Amy, I'd like to ask the same question to you.  I know you all have several different use cases [21:41] that you're implementing AI in. Um are but how  are you thinking about this this kind of blackbox [21:49] um question and within your own approach within  the State Department's approach to governance, [21:55] how are how are you trying to account for um  and and manage some of some of that? you know, [22:01] we're not sure what happened there and making  since you know, your outputs are so much more [22:07] important. Um the the risks are a lot higher  with the work you all are doing. Um so yeah, [22:14] same question to you. Sure. Um I think about  it in a couple of ways. So one is there's a [22:21] technology aspect to this. So um we have built  tools that help people take uh responsibility [22:30] for their work products. So we have a tool it's  called co-drafter. It's part of our suite of state [22:37] chat tools. Um and and it is really intended to  help people quickly um draft uh common department [22:46] uh documents. So we have as you can imagine we  have um multiple formats for documents. So we have [22:52] information memos, we have action memos, we have  talking point cards and background checklists and [22:58] all of those have a style style guide associated  uh with it. And um there are probably over 300 [23:06] requirements for that style guide. So when we  built this tool to create uh the the end uh goal [23:13] is is the tool will help you draft one of those  documents. when you actually output it to word and [23:20] um you have to to review it. We have built  into the technology um the requirements for [23:27] people to review that document. They have to at  least check the classification of each of the do [23:33] of the document. They have to check because  it's in red and so they have to change it. [23:38] um we've built a a little um label at the top  that says created by co-drafter and it's in red [23:45] and blue and so you have to review the document  you have to remove that that um label on that [23:51] document. In other examples, we built a the  department has a 20,000 or 25,000page foreign [23:59] affairs manual that which policy manual which  which um people use to to get guidance um from [24:08] anything from uh looking at adjudicating adoption  cases to um providing awards to employees. And so, [24:19] um, we when we built a tool that helped people  access that information, we built it so that not [24:25] only did it have the the links and connections  to the actual extracts of those documents, [24:30] so you could open up um the the actual document  and see what the the output of match what the [24:37] output of the the AI tool shared with with the  actual document. Um, we also built it to show it [24:44] side by side. So, so you have on the right screen  you have um what the AI summary is of that uh of [24:49] your question with the actual uh listing and uh  segments of the actual foreign affairs manual. [24:57] On the technology side, we we really have thought  about how we can make it easier for employees to [25:02] um engage with AI. And then on the on the  human side, which is the person employee side, [25:09] um which we're very um person centered in the  way that we do uh AI is is really focused on [25:17] helping people understand the and train people  around the use of AI. So, um, always, you know, [25:25] human review, always, uh, people understanding  that this is a tool. This isn't, um, this isn't, [25:33] uh, the end point. You don't just get to to create  something and then hand it in. It's we really t [25:39] we really help people understand that this is this  is sort of an output from say an intern or uh that [25:46] you then have to to review and ensure the accuracy  of the document. And we have uh fortunately for at [25:53] the state department we have a national uh foreign  affairs training institute where we um train uh [25:59] our employees on uh if I'm an economic officer I  go through a tradecraftraft course on on uh being [26:07] an economic officer. If I'm a management officer,  um there are trainings uh in in management officer [26:14] training from from um initial uh entry into the  foreign service to to more advanced um training [26:21] and we have we are working with that institute  to in um to integrate AI training into the all [26:28] of those tradecraftraft courses. So they're  uh not only um having we're not only building [26:36] tools that help them to to do the human uh review  that encourage them to to do the human review, [26:43] we're also providing training throughout their uh  life cycle uh employment of the state department. [26:50] So I'm hearing a lot of humans in the loop at  really critical points and making sure that um you [26:57] know there's that human layer before anything gets  gets pushed um externally. Um one other element of [27:07] governance is ROI. um you know the whole budgeting  element of this. There's been countless news [27:16] articles about companies that because they don't  have oversight into what their employees are doing [27:22] with AI, they end up burning through AI budgets  really quickly um and aren't able to have that [27:28] visibility in into token costs. Um Allan, how  are you all thinking about about ROI um either [27:37] internally or with you know what you've seen  with the companies that that you're working with? [27:43] Yeah, I think it's really interesting question  and I think if we think about it in the way that [27:49] Tim was describing in his organization kind of the  way the development efforts were breaking down in [27:55] terms of some are product development efforts  some are more back office efforts and I think [28:02] if you look at both of those things separately  you see different dynamics come into play as it [28:08] relates to AI and thinking about ROI. So one is  just more from a product development standpoint. [28:16] The software development life cycle has changed  quite a bit because of AI. I think everyone who's [28:22] in software development knows that and is still  kind of coming to grips with what it ideally will [28:31] look like going forward. But one thing is for  sure it's not going to be the way that it has [28:37] always looked before. And so I think one of the  big challenges that we face is as that software [28:43] development life cycle changes, how does the way  that we measure the ROI of the product that's [28:50] produced get changed to match that as well because  what we what we know is that the way that it was [28:58] done before where a single team was was building  a product together and they all knew all the ins [29:06] and outs and the nuances of all the code and the  functions that are involved doing essentially [29:12] vibe coding for a product, that same thing will  not hold true. And so we don't know what that [29:20] will mean in the long term in terms of the the  amount of maintenance that that's required and [29:27] the upkeep that's required, what that all will  look like, and that clouds the ROI situation. [29:33] So that's kind of from the product standpoint,  but if you think about it from more the the back [29:38] office or or desktop productivity standpoint,  it becomes even more muddled because as Tim [29:46] was describing in his organization, you you  want as a company, you want to be able to put [29:52] tools in the hands of your people that are  going to help them to do their work better. [29:58] But what happens if across the enterprise you  have 10 different people that are doing the [30:04] same job? They all come up with a way to do that  job better, but those ways are all different. So [30:11] there's 10 different things out there that are  built to do that same job and they might not [30:17] all be as effective. They might not all have the  same propensity for errors being introduced into [30:24] the system. And so all of these things kind  of highlight the need for having a governance [30:31] layer that's that's kind of one step above the  the software and the agents and the tools that [30:37] you produce so that you can bring together all of  these things in a in a unified way that makes it [30:45] measurable at the enterprise level. Tim, how  are you all thinking about ROI at Audible? [30:52] Yeah. So, um two two aspects of that. One,  uh being part of Amazon, we're uh we're [31:01] leveraging AWS quite a bit. There's lots of uh uh  availability uh and governance within the tool. [31:08] So, if if non-technical users, they're not going  to be burning through tokens uh at an incredible [31:16] rate. Monitors will go off, alarms will go off.  Permission permissioning is a big deal as Shash [31:21] was talking about uh for our developers. We want  to increase throughput while actually reducing our [31:28] um our compute spend. So these are goals that  our teams have um and and from a corporate [31:34] perspective, we have uh objectives in key result  framework. Uh three out of our six um uh OKRs for [31:41] this year are framed up around AI. So we're trying  to tie back the goal to deliver more customer [31:48] value to grow to deliver more stories to people uh  have them uh h have our employees uh look at u the [31:56] objectives that we want as a company embracing AI  moving faster giving customers new way to discover [32:02] uh their next listen and bring more books to  customers ears um from the top and then finding [32:09] ways for them to connect it and and engulinging  them on that, measuring them on that and providing [32:14] an environment where uh we last year was about  adoption for us. There was adoption metrics are [32:20] people using these tools, what percent and we  saw that go up and it's it's far less about [32:25] adoption this year that it really is embracing new  ways of working. So we set the the the standard [32:31] from the executive team about what we expect uh  through those frameworks that we already have. [32:36] So another layer of AI implementation, it's not  just about the, you know, the technical system, [32:43] but it's also about the people side of things  and the people management side of things. Um Amy, [32:51] I'd like to start with you. Um since you're an  organization that is not full of technical people [32:58] um how are you all thinking about um just  the people side of of an AI rollout an AI [33:06] implementation getting people on board with new  systems managing people who may be using them in [33:13] a way they shouldn't be using how are you how  are you all thinking about the the the human [33:17] side and of of your AI roll out yeah so we're  as As I mentioned earlier, we're very people [33:25] centered when we think about AI uh adoption and  enablement. And so um we have heard things like I [33:34] have heard things uh like my supervisor says that  it's cheating using AI is cheating. Um this isn't [33:43] for me. Uh this how I'm I'm a financial management  officer. This isn't for me. how am I supposed to [33:50] use these um these tools? So, we've you know heard  all all kinds of um uh challenges associated with [34:00] with AI implementation and we have um focused  in in a few different areas when we think about [34:08] uh AI adoption. Um we uh build with users. So um  one is we when we built our chatbot and when we [34:19] build the custom uh tools associated with it we  are engaging with the users all the time. So we [34:28] have um we have we collect information uh from  them uh regularly about tool use and and in our [34:36] first year of development we have um worked with  uh users to provide us with feedback on feature [34:45] rankings. So we went out and did a c we so my team  had uh what they um dubbed Thunderdome which was [34:53] uh getting the technical uh and development  teams together to to to come up with a list of [34:59] 20 features that that they um think is important  for us to build. Then we sent that out to our [35:05] users across the department and said what are your  what are your top features that you want to see [35:12] developed? And we did that probably um three times  throughout the year. And we get their feedback [35:19] and they and then they're part of the development  process. They see that we're listening to them and [35:24] that we're uh developing the tools that are built  for them and that has created a a lot of trust [35:31] um in in the use of AI uh at the department and  it has um generated a lot of uh engagement with [35:39] uh our folks as well. The other thing is we  build we build to their workflows. So, so like [35:45] Alan mentioned um we're looking we're not looking  to deploy technology. We're looking to change the [35:52] way the department does its work. And that means  we have to help people understand how to use AI in [35:58] their workflows. So if I am a financial management  officer, I we want to help them think through [36:05] how they can use AI in their individual daily  workflows. And one of the things that we have [36:11] learned very early on is that they don't want  to hear from us. They don't want to hear from [36:16] um our data scientists, our engineers, our product  team. They want to hear from the FMO that's out [36:21] in Frankfurt on how they're using AI. And so  we've created a lot of opportunities through we [36:27] call them office hours um where people can can um  showcase we ask people to showcase their use cases [36:35] uh in AI to share and exchange information um  across teams. We've also created an AI champions [36:42] program. Some people some organizations call them  AI ambassadors programs but we can't call it that. [36:47] Um so we have a AI champions program where the um  there's a uh it can be a technology officer, it [36:56] can be um uh someone uh in in a particular mission  or uh post who is particularly AI leaning forward [37:04] uh forwardleaning. And so uh those champions  put together at their post and at their mission [37:10] working groups to share experiences on how they're  using AI. They're they're looking at the use cases [37:17] for the different workflows that the department is  engaging on um to uh to to share that information [37:25] um with other not only people within their embassy  but also people across embassies. So we we gather [37:31] the AI champions together to um not only do  training for them and and share with them what [37:36] what kind of tools and technology are available  but also for them to exchange more importantly I [37:41] would say for them to exchange information on what  they're doing uh at their mission. So so those are [37:47] those are a few of the areas that we have really  focused on on being people uh centered and and [37:52] really um uh letting employees understand how the  technology is being built. let them feel part of [37:59] uh the build and and and help uh prioritize what  they want to see uh built as well. Tim, I'd like [38:08] to ask a similar question to you. How are you all  thinking about the people side of AI rollout? Um, [38:16] how are you, um, how are you thinking about, you  know, people being being reluctant to use it or [38:24] encountering, you know, push back that that maybe  has has merit where they're seeing the tools have [38:32] um, you know, get things wrong um, and so they're  nervous about about using them. Yeah, for sure. A [38:37] couple of the things that Amy mentioned resonated  with me. So we also have a quarterly genai [38:42] showcase with many different tracks of work where  uh novice AI users for example our HR department [38:49] they'll you know they're at a different uh stage  of adoption than our engineers and marketers are [38:55] probably somewhere in the middle so a safe space  for people to share what they've done but also [39:01] talk about the things that didn't go well because  everyone's uh you know in this together they're [39:05] they're they're also using AI in their personal  lives so they know for example if in a chatbot [39:11] uh if they're using chat JPD uh GPT it doesn't get  it right all the time and they're used to steering [39:16] things. So just providing a safe space for people  to connect and Allan like you said if 10 different [39:22] people or or teams are doing the same things  hopefully that provides an opportunity to say [39:27] oh I'm doing that let's collaborate um we we used  AI accelerators uh Amy so you can feel free to [39:34] use that as well uh but uh a group that's really  out on the forefront and is embedding themselves [39:39] um uh within these teams to help them um you know  really embrace AI and and and Grace I think that [39:46] the heart of your question is is really uh where  I mentioned last year was about adoption, this [39:52] year uh is is about embracing AI for us and so uh  people want to do their jobs better and faster. [39:58] They don't want to mechanically do the same things  over and over again. So using AI to help them do [40:04] their jobs better and then framing it up in terms  of uh you know Audible's goals to grow uh uh uh [40:11] to to grow our cate category to bring more people  into uh into books and to provide more books and [40:18] more languages uh to our listeners connecting  that up and down has has proven to be uh [40:24] pretty valuable and resonates with our employees.  Josh, I'd like to ask a similar question to you. [40:31] what have you seen um as some some downsides  to AI use and how are you all thinking about [40:39] um you know the people side of this. So internally  because the type of organization we are we we like [40:46] to give our folks a lot of room to sandbox  and and play so they're rapidly building and [40:51] iterating with their customers kind of as fast  as possible just to throw things at the board [40:54] and see what's what's interesting. Um, but  I also host a customer advisory board. So, [40:59] kind of like your your internal meetings. And  that's where the real horror stories start to [41:04] come out that get really interesting across the  organizations. Um, one uh one of our our customers [41:09] was telling us that they had an individual that  was doing that sort of sandboxing stuff and spent [41:14] $20,000 in tokens in three days. Um, a couple  minute conversation with it would have gotten [41:20] around that. um he was pulling three or four  of their systems every like 30 seconds instead [41:26] of just doing like a once a day sync and just he  didn't know any better. They didn't have programs [41:31] in place that were helping to teach these folks  that have access to these very powerful tools that [41:36] could have a very big blast radius just some basic  best practices. It's like I I bet you if a lot of [41:43] you have probably built something with like a in  a in a clawed code or a codeex and if you go and [41:48] look at what what that built for you right now,  there's probably this giant main.typescript file [41:53] that's that's through there. Um you know, I can  easily go and coach somebody and say, "Hey, by the [41:58] way, this is how Claude or how Codeex is going to  want to build that thing to be easy to begin with, [42:03] and it's going to be impossible to manage when  you have this giant monolithic thing later on." [42:08] and we just go through a little training sessions  through some of those things to help everybody do [42:11] a bit a bit better. But I I think that's an area  that almost every company I work with hasn't done [42:17] a great job really building that foundational  ed education. And that could be as easy as, hey, [42:23] here's a little bundle of some, you know, agent uh  markdown files and things that you can give to a [42:29] claude or a codeex so it just knows here's what  our design standards are. And so you don't have [42:34] to always rely on the human remembering every  single little thing of what to do. You kind of [42:39] give them the cheat code out the out the gate.  So great. Well, we're almost at time. Um Allan, [42:45] I'd like to give you a chance to, you know,  address the same question. What what your I know [42:51] you guys operate in a lot of different industries  and you're working with a lot of different kinds [42:55] of companies. what what have you seen um in terms  of this question of cultural resistance and you [43:03] know humans using AI in a way that that causes  problems and and how you're kind of thinking about [43:10] about that layer. Okay. Yeah. I think the big  thing used to be the bots are coming for our jobs, [43:18] right? That was everyone's first response to  uh well we're going to implement AI. everyone [43:25] started thinking, well, I'm one foot out the  door now because my job isn't that hard, so [43:30] I'm sure AI can do it. And I think now we've kind  of moved on from that. And most companies realize [43:37] that the that the real advantage to AI isn't in  replacing people, it's in making people better. [43:45] And so the the key issue is putting the support  around them that allows them to transition their [43:51] job from where it is now to where ideally it could  be. And so there are a lot of different aspects of [43:58] that that you have to bring into play to make it  successful. But I think the the key thing is to [44:04] to know your people and to work to find that way  forward. Great. Thank you all. That's our time. [44:11] Um I know a new panel is starting here in just a  few minutes. So, um, if anybody wants to continue [44:17] conversations, um, let's do that outside.  But thank you so much. Thank you, Chris.