OpenAI Unveils Workspace Agents and Frontier Platform to Automate Tasks

OpenAI发布工作区智能体与Frontier平台以实现任务自动化 金融服务机构可以通过部署由OpenAI Frontier平台管理的自然语言工作区智能体，弥合个人ChatGPT使用与复杂企业自动化之间的差距。 Lee Spacagna介绍了ChatGPT工作区智能体，以弥合员工个人使用与自上而下的大规模系统重构之间的运营缺失层。通过现场演示构建一个“幕僚长”（Chief of Staff）智能体，他阐述了自然语言、预置模板以及集成的Microsoft Teams、SharePoint和Salesforce等企业工具，如何让非技术业务用户实现复杂多步骤工作流的自动化。Spacagna进一步解释了封装的团队技能与OpenAI的Frontier平台如何使机构能够在金融服务等企业领域中治理、部署并持续改进数以千计的特定角色智能体。 OpenAI 推出基于 GPT-5.5 驱动的 ChatGPT 工作空间智能体（Workspace Agents），填补自下而上个人任务自动化与自上而下企业系统重构之间的“缺失层”，专注于团队及部门级的端到端复杂工作流。 (0:00) 业务人员无需技术或提示词工程背景，仅凭自然语言并借助元智能体（meta-agent）即可搭建智能体，支持连接 Outlook 与 Teams 并设定每日早 9 点自动推送简报等定时任务。 (2:07) 通过引入 SharePoint 和 Salesforce 等企业数据源连接及“技能（Skills）”机制，智能体能够沉淀组织内部知识与业务规范，实现会前调研及隔夜邮件/会议纪要的自动草拟。 (6:06) 为应对跨企业运行数千个智能体的治理难题，OpenAI 打造了 Frontier 集中化平台，助力金融机构在 KYC 审核、反洗钱（AML）调查及客户关系管理等场景中规模化部署与自我评估智能体。 (9:50)

OpenAI Solutions Engineer Lee Spacagna contends that enterprise AI has overlooked a critical departmental automation layer situated between individual user prompting and massive top-down system overhauls. To bridge this gap, OpenAI is introducing workspace agents and centralized governance platforms designed to execute multi-step workflows across everyday business software. Bridging the Enterprise Automation Layer (0:00) Spacagna identifies a persistent structural division in enterprise AI adoption. Organizations typically pursue either bottom-up experimentation—individual employees leveraging ChatGPT and Codex for discrete daily tasks—or top-down engineering initiatives that rebuild core functions like customer service and advisory systems from scratch. He argues that this leaves a missing middle at the team and department level, where complex multi-step coordination across shared tools remains largely unautomated. To address this operational layer, OpenAI introduced ChatGPT workspace agents powered by model advances such as GPT-5.5. Unlike static conversational bots or isolated custom GPTs, workspace agents unify shared applications, custom skills, and deployment environments into a single platform, shifting AI's utility toward executing end-to-end tasks that previously required hours or days of human effort. No-Code Agent Creation and Autonomous Scheduling (2:07) In a live demonstration of a "Chief of Staff" agent, Spacagna shows that creating operational agents no longer requires specialized programming or prompt engineering expertise. Instead, business users provide natural language prompts, and an underlying meta-agent automatically drafts the required system instructions and coordinates tool integrations. The demonstration connects the agent directly to Microsoft Outlook Calendar, Teams, and Outlook Email, configuring it to trigger autonomously at 9:00 a.m. every morning. Upon triggering, the agent reviews overnight emails, application updates, and schedules to synthesize an operational brief highlighting priorities, blockers, and decisions, which it then posts into a designated CFO team channel in Microsoft Teams. Capturing Institutional Knowledge with Skills and Connectors (6:06) To mitigate team burnout stemming from back-to-back meetings without adequate preparation, agents can be linked to enterprise repositories including SharePoint for internal notes and Salesforce for CRM history. Spacagna emphasizes that data access alone is insufficient; agents must also incorporate "Skills," which translate unwritten organizational habits and formatting conventions into structured, repeatable rules. By attaching a pre-built "Meeting Prep" skill, the agent determines how to structure executive briefings and which data sources to prioritize before launching a starter prompt inside ChatGPT. As a personal use case, Spacagna notes that he uses an overnight agent to process emails, call transcripts, and Slack threads into ready-to-send draft responses, which saves him the first hour of his workday. Centralized Governance and Scaling with Frontier (9:50) Spacagna frames this agent-driven model as widely applicable across specialized financial services workflows, including Know Your Customer (KYC) onboarding, Anti-Money Laundering (AML) investigations, and client relationship management. Rather than running isolated automation experiments, organizations are expected to deploy role-specific agents across every team. To resolve the governance challenges of managing thousands of operational agents, OpenAI developed Frontier. According to Spacagna, Frontier acts as a centralized management platform connecting siloed enterprise data warehouses and CRMs, allowing deployed agents to reason over enterprise data, execute code, access approved tools, and run automated self-evaluations to improve performance over time. Spacagna presents workspace agents and the Frontier management platform as an operating model capable of turning manual, multi-step departmental tasks into governed autonomous routines, though his brief presentation serves primarily as an operational overview before handing off to subsequent segments without evaluating potential technical failure modes or deployment constraints.

OpenAI 欧洲、中东和非洲地区金融服务主管解决方案工程师 Lee Spacagna 提出，当前企业 AI 落地在员工自发的零散使用与自顶向下的系统重构之间存在团队层级的“断层”；企业需要借助基于 GPT-5.5 的工作区智能体与集中式 Frontier 平台，以自然语言直接调度日常办公工具并实现规模化管理。 弥合企业自动化落地的中间断层 (0:00) 根据 Lee Spacagna 的观察，企业管理层在探索 AI 变革业务路径时，通常分化为两条路径：一种是由基层员工利用 ChatGPT 和 Codex 提升个人日常读写效率的“自底向上”模式；另一种则是从底层重新构建核心架构以改造客服、客户咨询与核心运营的“自顶向下”模式。然而，在这两种模式之间，团队和部门层级的自动化长期处于缺失状态。 为填补这一断层，OpenAI 依托 GPT-5.5 的能力跃迁推出了 ChatGPT 工作区智能体（Workspace Agents）。与仅能回答静态问题的传统交互不同，工作区智能体将共享应用程序、定制技能和部署工具整合至统一平台，能够直接调用电子邮件、日程表及日常办公软件，自主执行跨多步骤的复杂端到端工作流，将原本耗时数小时或数天的任务在日常工作环境中完成闭环。 无代码元智能体构建自动化幕僚长 (2:07) 在现场演示中，Lee 展示了通过 ChatGPT 侧边栏的“Agents”入口快速搭建“幕僚长（Chief of Staff）”智能体的过程。该方案旨在整合分散在各类邮件、会议和文档中的琐碎工作。构建流程无需用户具备提示词工程技能或编程背景，而是通过系统内置的“元智能体（Meta-agent）”以纯自然语言自动生成底层执行指令，使业务人员能够直接使用一个智能体去构建另一个智能体。 该演示智能体通过预设模板直接打通了微软生态工具，包括 Outlook 日历、Microsoft Teams 和 Outlook 邮件。用户设定智能体在每天上午 9 点自动运行，主动检索隔夜收到的邮件、应用更新及当天的日程安排，提炼出包含优先级、阻碍项与待决事项的每日简报，并自动推送至首席财务官（CFO）团队的 Teams 专属频道内。 融合多源数据与组织隐性知识技能 (6:06) 针对连续会议导致员工缺乏准备时间而产生职业倦怠的问题，智能体被进一步赋予主动会前调研的能力。系统支持挂接更多企业数据源，包括用于内部文档存储与共享笔记的 SharePoint、沉淀客户背景与 CRM 记录的 Salesforce，以及各类企业内部定制应用，从而打破系统间的信息壁垒。 此外，Lee 提出了“技能（Skills）”机制，其核心是将沉淀在员工脑海中的组织内部惯例与隐性知识转化为可复用的逻辑规则。例如挂接“会议准备（Meeting Prep）”技能后，智能体能够遵循既定的信息结构与排版规范整理会前摘要。Lee 提及个人实际应用时表示，他利用隔夜智能体梳理 Slack 讨论串、通话转录文本与邮件并起草回复，每天早晨仅需审核并点击发送，即可节省首个工作小时。 借助Frontier实现智能体集中治理 (9:50) Lee 指出，“幕僚长”模式不仅适用于个人助理场景，还可广泛拓展至金融服务领域的各类业务流，如了解你的客户（KYC）入职审核、反洗钱（AML）调查以及客户关系管理。这一转变意味着企业将告别孤立的单点自动化试验，转向每个业务团队皆部署专属角色智能体以承接手工劳作的新型运营模式。 针对未来企业内部可能并存数千个智能体所带来的管控难题，OpenAI 推出了名为 Frontier 的集中式平台。Frontier 负责将数据仓库、CRM 与内部系统进行统一治理与规模化部署，支持智能体在安全受控的环境下进行数据推理、代码执行与工具调用，并通过交互反馈与自动化自我评估实现性能的持续迭代。 该演讲阐明了企业 AI 正从被动的问答工具转向主动执行的团队协作智能体，但随着跨部门智能体数量的快速扩张，如何在打通异构企业数据孤岛的同时建立统一的治理体系与自主纠错机制，仍是规模化落地过程中的关键考验。

## Background 该视频由 OpenAI 于 2026 年 7 月 3 日发布在 YouTube 官方频道，主讲人为 OpenAI 首席解决方案工程师 Lee Spacagna。视频介绍重点展示了包括 Workspace Agents 与 Codex 在内的 OpenAI 最新企业级产品，解析其如何帮助金融服务等行业机构在实际业务工作流中实现 AI 的运营化落地。 ## TL;DR 针对企业在个人效率工具与战略级系统重塑之间长期存在的部门自动化断层，演讲介绍了通过 ChatGPT Workspace Agents 将 AI 升级为可直接委托端到端业务流程的数字协同者。现场演示了仅凭自然语言与模板零代码搭建‘数字幕僚长’的过程，通过整合 Outlook、Teams、SharePoint 与 Salesforce，实现自动汇总每日晨报与开展跨系统会前背景调研。该架构进一步将员工脑海中的隐性经验抽象为标准化的‘技能’（Skills），便于在 KYC、AML 等金融合规场景中横向复制角色型 Agent 矩阵。最终，依托 Frontier 企业级治理平台，组织能够打通底层数据孤岛，实现成规模 Agent 的安全纳管与持续自我进化。 ## Atmosphere & style 整场演讲呈现出沉稳务实、条理清晰且极具说服力的企业技术布道风格。主讲人以一线金融客户最真实的落地焦虑开篇，随后迅速切入流畅敏捷的零代码实操演示，由浅入深地拆解复杂的技术接入与治理架构。整体观感如同一场直击企业效率痛点的高效内训，不仅展现出大模型接管琐碎案头工作时的从容与优雅，更向观众传递出全新 AI 运营模式即将重塑日常协作的深远前瞻感。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=fAxlEcXiSts&t=0s) OpenAI 推出 ChatGPT Workspace Agents，填补了自底向上个人效率工具与自顶向下核心系统重塑之间的结构性断层，专注于部门级端到端任务委托。 - [02:07](https://www.youtube.com/watch?v=fAxlEcXiSts&t=127s) 借助 GPT-5.5 模型与内置的 Meta-Agent 机制，业务人员无需编写代码或提示词，仅凭自然语言即可接入 Microsoft 365、Salesforce 和 SharePoint 定制专属角色 Agent。 - [06:06](https://www.youtube.com/watch?v=fAxlEcXiSts&t=366s) “技能”（Skills）模块能够将团队脑海中难以量化的隐性经验固化为执行规则，实现自主跨源梳理会前情报与生成每日精炼晨报。 - [08:54](https://www.youtube.com/watch?v=fAxlEcXiSts&t=534s) 个人幕僚长 Agent 通过在夜间整合未读邮件、Slack 待办承诺与会议通话转录并拟定回复草稿，帮助员工每天早晨直接“赢回第一个小时”。 - [10:14](https://www.youtube.com/watch?v=fAxlEcXiSts&t=614s) OpenAI 推出企业级治理平台 Frontier，打破 CRM 与自研数据库孤岛，支持在合规受控环境中实现 Agent 的规模化纳管与自我学习进化。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=fAxlEcXiSts&t=0s) 企业 AI 落地在个人工具提效与大型数字化转型工程之间存在断层，部门级自动化因场景特异且缺乏研发资源长期处于空白状态。 - [01:09](https://www.youtube.com/watch?v=fAxlEcXiSts&t=69s) OpenAI 重新定义了 Agent：它并非对话式问答机器人，而是能够被委派实质性端到端复杂任务、主动调用日常工具的数字团队成员。 - [01:09](https://www.youtube.com/watch?v=fAxlEcXiSts&t=69s) GPT-5.5 的底座能力使得 Agent 能够接管以往耗费数小时乃至数天的多步骤复杂跨系统业务流程。 - [01:09](https://www.youtube.com/watch?v=fAxlEcXiSts&t=69s) Workspace Agents 将工具生态、技能库与组织级部署整合进统一的 Agent Builder，让自动化真正进驻到日常办公流中。 - [02:07](https://www.youtube.com/watch?v=fAxlEcXiSts&t=127s) 构建机制实现了彻底的零代码化，借助“生成 Agent 的 Agent”，业务人员用自然语言即可快速生成底层系统提示词与执行策略。 - [03:43](https://www.youtube.com/watch?v=fAxlEcXiSts&t=223s) 原生打通 Outlook 日历、邮件与 Teams 协作频道，数字幕僚长可实现跨渠道定时汇总并在安全确认后发布业务简报。 - [03:43](https://www.youtube.com/watch?v=fAxlEcXiSts&t=223s) 在执行向团队频道推送信息等对外动作时，系统严格内置授权审批机制，确保 Agent 的全自主行为安全可控。 - [06:06](https://www.youtube.com/watch?v=fAxlEcXiSts&t=366s) “技能（Skills）”机制的核心在于捕获并固化员工脑海中未成文的隐性业务经验，将其转化为可稳定复现的标准工作流。 - [06:06](https://www.youtube.com/watch?v=fAxlEcXiSts&t=366s) 通过挂载 SharePoint 内部知识库与 Salesforce CRM，Agent 可在背靠背密集会议前自动调取客户记录与商机上下文进行充分准备。 - [08:54](https://www.youtube.com/watch?v=fAxlEcXiSts&t=534s) 夜间全自动汇总未读信息、Slack 待办承诺与通话纪要并生成现成回复草稿，彻底改变了管理者的早间工作模式与效率瓶颈。 - [09:50](https://www.youtube.com/watch?v=fAxlEcXiSts&t=590s) 企业的真正机遇在于超越单点自动化项目，迈向允许各业务部门快速拉起专属角色型 Agent 的新型企业运营模式。 - [09:50](https://www.youtube.com/watch?v=fAxlEcXiSts&t=590s) 金融服务机构正加速将此架构复制到诸如 KYC 入职审核、AML 反洗钱调查等强合规、高复杂度的专业业务场景中。 - [10:14](https://www.youtube.com/watch?v=fAxlEcXiSts&t=614s) 平台级解决方案 Frontier 专为规模化治理而生，旨在打破数据仓库、CRM 与内部应用孤岛，赋予 Agent 一致的企业共享上下文。 - [10:14](https://www.youtube.com/watch?v=fAxlEcXiSts&t=614s) Frontier 支持 Agent 在受控环境中推理、执行代码和调用工具，并通过持续评估与交互反馈形成自我优化的进化飞轮。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=fAxlEcXiSts&t=0s) 企业 AI 落地的鸿沟与 Workspace Agents 的诞生 在服务欧洲、中东及非洲（EMEA）地区金融客户的一线实践中，OpenAI 首席解决方案工程师 Lee Spacagna 每天都会面对机构高管们提出的同一个核心诉求：“AI 究竟在哪里能真正改变我企业的运转方式？”（"where can AI actually change how my business runs?"）。当前企业在采纳 AI 时普遍存在两条割裂的路径：一条是自底向上的个人赋能路线，通过推广 ChatGPT 和 Codex 提升普通员工的“AI 素养”（AI literacy），帮助大家应对日常零散事务；另一条则是自顶向下的系统重塑路线，由高层推动重大的数字化转型工程，开发全新产品，以重塑客户服务、优化客户投顾咨询并加固后台运营支持。 然而，这两条路径之间存在一层长期被忽视的结构性断层——团队与部门层级的业务自动化。部门的日常工作往往高度依赖特定业务背景，难以通过单人桌面工具直接解决，同时又往往不及重大企业级项目那般庞大以至于能获得专属研发资源。为了填补这一关键的中间空白，OpenAI 推出了 ChatGPT Workspace Agents，旨在直接在团队工作空间中落地端到端的流程自动化。 ```mermaid flowchart TD accTitle: 企业 AI 落地架构与 Workspace Agents 定位 accDescr: 图表展示了企业自顶向下与自底向上两条传统路线，以及位于中间、由 Workspace Agents 填补的团队自动化层。 subgraph TopDown["自顶向下：核心系统转型"] A["战略级 AI 系统<br/>客户服务 / 投顾咨询 / 后台支持"] end subgraph MiddleLayer["中层断层：团队与部门级自动化"] B["ChatGPT Workspace Agents<br/>跨系统协同 / 流程委托"] end subgraph BottomUp["自底向上：个人效率工具"] C["ChatGPT 与 Codex<br/>日常办公协同 / 个人 AI 素养"] end TopDown --> MiddleLayer MiddleLayer --> BottomUp ``` ### [01:09](https://www.youtube.com/watch?v=fAxlEcXiSts&t=69s) 重新定义 Agent：从单点问答到全流程任务委托 在 OpenAI 的定义中，“Agent”绝非简单的对话式答疑机器人，而是“可以向其委托有实质意义任务的 AI 系统，而不仅仅是向其提问”（"AI systems that we can delegate meaningful tasks to, not just ask questions of"）。它需要像真正的团队成员一样，主动调用员工每天赖以生存的生产力工具——包括电子邮件、工作日历以及各类协作套件，并以人类员工熟悉的节奏和方式独立推进工作。随着 GPT-5.5 模型的推出，底座大模型的能力迎来了质的跃迁，使得 Agent 能够接管以往需要耗费数小时甚至数天才能完成的复杂端到端流程。 过去许多企业曾尝试搭建 Custom GPTs，但部门级协作往往需要更深度的工作流嵌入。Workspace Agents 将其进化为一套一体化的 Agent 构建器（Agent Builder），把共享应用生态、团队专有技能库（Skills）以及组织内部署统一在同一个平台之上，真正让 Agent 进驻到“工作本来就在发生的地方”。 ### [02:07](https://www.youtube.com/watch?v=fAxlEcXiSts&t=127s) 零代码实战：用自然语言构建“幕僚长” Agent 对于许多团队而言，自动化的难点从来不是缺乏工作场景，而是上下文极度分散——事实与决策线索被拆解在会议日程、往来邮件、碎片文档以及各种遗留系统之中。Lee 在 ChatGPT 界面左侧导航栏调出全新的“Agents”功能，现场演示如何为自己快速定制一名专属的“数字幕僚长”（Chief of Staff Agent）。这位数字员工的核心职责在于统筹跨部门工作、动态追踪业务优先级、提前筹备会议并确保整个团队高效推进。 通过选用预设的 Chief of Staff 模板，系统已初始化了基础指令、可用工具集及核心能力。随后，Lee 接入了 Microsoft 生态组件，包括 Outlook Calendar、Microsoft Teams 和 Outlook Email。整个构建过程彻底颠覆了传统的 Prompt 编写流程：后台配备了一个专门用于“生成 Agent 的 Agent”，用户完全不需要掌握提示词工程（Prompt Engineering）或编写任何代码，只需使用纯自然语言描述业务诉求，就能让 AI 协同生成底层执行逻辑，实现企业业务人员随时随地“用自然语言制造数字员工”。 - **开箱即用模板**：预置了幕僚长职责边界与通用协同规范，降低冷启动门槛。 - **Microsoft 生态原生接入**：直接打通 Outlook 日历、Outlook 邮件与 Teams 沟通渠道。 - **Meta-Agent 机制**：由内置辅助 Agent 将人类的业务意图转化为精准的 Agent 系统指令。 ### [03:43](https://www.youtube.com/watch?v=fAxlEcXiSts&t=223s) 定时触发与多源协同：自动生成每日晨报并同步 Teams 初步配置完成后，Lee 进一步将这名数字幕僚长个性化改造为符合团队晨会节奏的自动化助手。他向 Agent Builder 发送自然语言指令：“设定在每天上午 9:00 自动运行，全面检索当天所有日程会议、关联应用数据以及隔夜收到的所有电子邮件，汇总生成一份精炼的每日简报（Daily Brief）。”这样一来，管理者每天清晨步入办公室时，就能对全天的会议背景与关键事务了如指掌。 在测试阶段，界面自动生成了推荐的启动指令（Starter Prompts）。Lee 触发了“准备今日简报”任务，并明确要求简报必须提炼出“核心优先级、关键决策点、潜在阻塞点（Blockers）与后续跟进项”，最终直接推送到 Teams 的“CFO 团队日常准备”（CFO team - daily prep）频道。Agent 随即展开跨源上下文检索：从日历调取当日日程，交叉比对邮件中的往来记录，提取相关业务背景与联系人脉络。在首次执行对外发布动作时，系统严格执行安全审核机制，弹窗请求用户授权发送至 Teams 频道；一旦获得确认，结构化的业务晨报即刻同步送达，实现了日常工作流的无缝闭环。 ```mermaid flowchart LR accTitle: 幕僚长 Agent 晨报生成与推送流程 accDescr: 展示幕僚长 Agent 汇聚日历与邮件上下文、经安全审批后向 Teams 频道发送晨报的过程。 subgraph Ingestion["多源上下文获取"] Cal["Outlook Calendar<br/>当日会议排程"] Mail["Outlook Email<br/>隔夜往来邮件"] end subgraph AgentAction["幕僚长 Agent 核心处理"] Proc["交叉比对与提炼<br/>决策项 / 优先级 / 阻塞点"] Auth["安全确认<br/>获取 Teams 发布权限"] end subgraph Delivery["工作流触达"] Teams["Teams CFO 频道<br/>Daily Prep 晨报落地"] end Ingestion --> Proc Proc --> Auth Auth --> Teams ``` ### [06:06](https://www.youtube.com/watch?v=fAxlEcXiSts&t=366s) 进阶扩展：注入 CRM 知识与团队“隐性经验”构建会前准备技能 为了解决团队长期面临的痛点——成员在无缝衔接的“背靠背”（back-to-back）会议中疲于奔命、缺乏会前准备时间，Lee 继续对这位数字幕僚长进行了能力升级，希望它能在每次开会前主动完成背景调查，明确“参会人员是谁、最新业务进展为何、本次会议的核心目标是什么”。 为了给 Agent 补充更深度的业务上下文，Lee 进一步引入了另外两个核心企业系统： 1. **SharePoint**：调取组织内沉淀的大量会议纪要、内部文档与跨部门共享资料。 2. **Salesforce**：调取 CRM 客户关系数据，包括客户联系人历史、商机阶段及重要业务背景。 除了数据接入，系统还引入了“技能”（Skills）模块。Lee 将“技能”定义为捕获关键任务执行规则与行业经验的高效载体。它能够将那些“困在员工脑海中的隐性经验与不成文惯例”（"tribal knowledge and conventions that are currently trapped in people's heads"）固化下来，转化为稳定复现的自动化工作流。在保留原有幕僚长基础技能和排版格式技能的基础上，Lee 挂载了一套预先定义好的“会议准备技能（Meeting Prep Skill）”，用以规范信息提炼维度与输出结构。 最后，Lee 仅需在自然语言侧边栏轻描淡写地交待一句：“已添加 Salesforce 和 SharePoint，请新增一项能力，支持在 ChatGPT 中生成快捷会议简报，并根据当前日程主动准备下一场会议。”辅助 Agent 几秒内便完成了底层逻辑与技能的重写，并在界面上生成了全新的快速触发按钮。部署之后，该数字幕僚长即刻面向整个团队上线可用，在接收到“为下一场会议做准备”的指令后，自主穿梭于日历、SharePoint 知识库与 Salesforce 客户记录之间，完成高密度的跨源会前情报梳理。 ### [08:54](https://www.youtube.com/watch?v=fAxlEcXiSts&t=534s) 个人专属“参谋长” Agent：重塑日常工作流与会议准备 随着技能指令将各类分散信息核对并整理成一份结构紧凑的简报，工作协同的方式发生了质的转变。演讲者分享了他个人每天都在运行的一个专属 Agent：该 Agent 会在夜间自动审查所有新收到的邮件、全公司范围内的重要业务更新、他在前一天于 Slack 或电话会议中承诺跟进的事项，并结合所有的通话转录文本进行深度上下文分析。 这种高度自动化的流程直接为他“赢回了每天的第一个小时”。过去每天清晨需要耗费大量精力逐一排查消息，而现在，当他早晨步入工作状态时，所有待回复的客户邮件都已经根据最新的业务背景生成了精准的草稿。他只需快速浏览、点击审批并发送给客户即可，彻底颠覆了以往的工作节奏。同样的逻辑也延伸到了团队会议的筹划中——原本很多团队因人手紧张而根本无暇为密集的会议做充分准备，如今借助这一机制，每位员工在步入会场时，都仿佛配备了一位专属的“幕僚长”（Chief of Staff）Agent 为其做足了背景梳理与信息铺垫。 - **早间工作流自动化**：夜间自动抓取并整合未读邮件、全公司重要通报、Slack 待办承诺与会议通话转录记录。 - **即用型回复生成**：邮件草稿预先注入全公司业务上下文，用户仅需审核与确认发送，省去传统早间一小时的信息筛选时间。 - **会议准备降本增效**：让资源有限的中小团队或人手不足的部门同样享有资深幕僚级别的简报与筹备支持。 --- ### [09:50](https://www.youtube.com/watch?v=fAxlEcXiSts&t=590s) 从单点工具到全新运营模式：跨部门定制化 Agent 矩阵 会议准备与邮件起草仅仅是一个典型切片，其背后的架构模式具备在整个企业范围内横向复制的能力。在金融服务机构中，这一模式正在迅速拓展至诸如 KYC（了解你的客户）入职审核、AML（反洗钱）调查、大客户关系管理（Relationship Management）等高门槛且强合规要求的业务场景中。 演讲者强调，这项技术带来的真正机遇绝不仅限于交付某一个孤立的“单点自动化项目”，而是推动组织迈向一种全新的“企业运营模式”。在这一模式下，任何业务团队都可以快速拉起专属的角色型 Agent，将大量繁琐的案头工作剥离出日常职责，从而使整个业务链条以极高的速度敏捷运转。 ```mermaid flowchart TD accTitle: Role-Specific Agents Operating Model accDescr: 跨部门角色定制 Agent 运营模式流转图，展示从企业底层数据到垂直 Agent 再到业务价值的链条。 subgraph DataContext ["底层系统与上下文"] direction LR S1["邮件与聊天记录"] S2["通话转录与简报"] S3["业务数据库"] end subgraph RoleAgents ["角色定制 Agent 矩阵"] direction LR A1["幕僚长 Agent<br/>(会议/日常管理)"] A2["KYC 审核 Agent<br/>(客户准入)"] A3["AML 调查 Agent<br/>(合规风控)"] A4["客户关系 Agent<br/>(业务拓展)"] end subgraph BusinessOutcomes ["业务成果"] direction LR O1["释放高价值工时"] O2["规范合规流程"] O3["加速业务流转"] end DataContext --> RoleAgents RoleAgents --> BusinessOutcomes ``` --- ### [10:14](https://www.youtube.com/watch?v=fAxlEcXiSts&t=614s) Frontier 平台：规模化治理与自我进化的企业级生态 当企业内部涌现出成百上千甚至数以万计的 Agent 时，如何实施统一管理和治理便成了核心挑战，而这正是平台级解决方案 **Frontier** 应运而生的原因。Frontier 专为大规模部署和纳管 Agent 而设计，它打破了传统数据仓库、CRM 软件及各类内部自研系统之间的信息孤岛，赋予 AI 协同办公伙伴与人类员工完全一致的共享上下文。 在严格受控和合规的环境中，Agent 可以基于这些数据展开逻辑推理、运行代码、调用专属工具并安全地触发业务动作。更为关键的是，该系统具备与人类员工类似的持续进化机制：Agent 能够在交互中持续学习，并随着时间的推移评估并优化自身表现，实现“越用越聪明”。目前企业已能基于 ChatGPT、Codex 与底层 API 进行构建，未来还将进一步针对金融服务工作流推出更多开箱即用的特定 Agent、插件及专业技能包，以极低的定制成本无缝嵌入各类标准化业务流程，让组织能够更自信地向 AI 委派复杂工作流。 - **打破系统孤岛**：无缝对接企业数据仓库、CRM 以及自研核心应用，确保 Agent 获得完整共享上下文。 - **全流程合规与能力闭环**：在治理框架内支持复杂数据推理、代码执行与自动化工具调用。 - **自适应优化飞轮**：通过多轮交互反馈与持续的性能评估，实现 Agent 处理能力的迭代升级。 - **金融垂直领域适配**：依托 ChatGPT、Codex 与 API，提供面向金融工作流的即开即用型插件与专业技能。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=fAxlEcXiSts&t=0s) where can AI actually change how my business runs?（AI 究竟在哪里能真正改变我企业的运转方式？） — Lee Spacagna > [01:09](https://www.youtube.com/watch?v=fAxlEcXiSts&t=69s) AI systems that we can delegate meaningful tasks to, not just ask questions of（可以向其委托有实质意义任务的 AI 系统，而不仅仅是向其提问） — Lee Spacagna > [01:09](https://www.youtube.com/watch?v=fAxlEcXiSts&t=69s) 把共享应用生态、团队专有技能库以及组织内部署统一在同一个平台之上，真正让 Agent 进驻到‘工作本来就在发生的地方’。 — Lee Spacagna > [02:07](https://www.youtube.com/watch?v=fAxlEcXiSts&t=127s) 后台配备了一个专门用于‘生成 Agent 的 Agent’，用户完全不需要掌握提示词工程或编写任何代码，只需使用纯自然语言描述业务诉求。 — Lee Spacagna > [02:07](https://www.youtube.com/watch?v=fAxlEcXiSts&t=127s) 实现企业业务人员随时随地‘用自然语言制造数字员工’。 — Lee Spacagna > [06:06](https://www.youtube.com/watch?v=fAxlEcXiSts&t=366s) tribal knowledge and conventions that are currently trapped in people's heads（困在员工脑海中的隐性经验与不成文惯例） — Lee Spacagna > [08:54](https://www.youtube.com/watch?v=fAxlEcXiSts&t=534s) 这种高度自动化的流程直接为他‘赢回了每天的第一个小时’。 — Lee Spacagna > [08:54](https://www.youtube.com/watch?v=fAxlEcXiSts&t=534s) 每位员工在步入会场时，都仿佛配备了一位专属的‘幕僚长’ Agent 为其做足了背景梳理与信息铺垫。 — Lee Spacagna > [09:50](https://www.youtube.com/watch?v=fAxlEcXiSts&t=590s) 这项技术带来的真正机遇绝不仅限于交付某一个孤立的‘单点自动化项目’，而是推动组织迈向一种全新的‘企业运营模式’。 — Lee Spacagna > [10:14](https://www.youtube.com/watch?v=fAxlEcXiSts&t=614s) Agent 能够在交互中持续学习，并随着时间的推移评估并优化自身表现，实现‘越用越聪明’。 — Lee Spacagna ## Mentioned OpenAI, Lee Spacagna, ChatGPT, Codex, ChatGPT Workspace Agents, GPT-5.5, Agent Builder, Chief of Staff Agent, Microsoft, Outlook Calendar, Microsoft Teams, Outlook Email, SharePoint, Salesforce, Slack, Frontier ## Video description How OpenAI’s newest enterprise products, including Workspace Agents and Codex - can operationalize AI across financial services organizations. [Source](https://www.youtube.com/watch?v=fAxlEcXiSts)

[0:00] Lee Spacagna: where can AI actually change how my business runs?（AI 究竟在哪里能真正改变我企业的运转方式？） [1:09] Lee Spacagna: AI systems that we can delegate meaningful tasks to, not just ask questions of（可以向其委托有实质意义任务的 AI 系统，而不仅仅是向其提问） [1:09] Lee Spacagna: 把共享应用生态、团队专有技能库以及组织内部署统一在同一个平台之上，真正让 Agent 进驻到‘工作本来就在发生的地方’。 [2:07] Lee Spacagna: 后台配备了一个专门用于‘生成 Agent 的 Agent’，用户完全不需要掌握提示词工程或编写任何代码，只需使用纯自然语言描述业务诉求。 [2:07] Lee Spacagna: 实现企业业务人员随时随地‘用自然语言制造数字员工’。 [6:06] Lee Spacagna: tribal knowledge and conventions that are currently trapped in people's heads（困在员工脑海中的隐性经验与不成文惯例） [8:54] Lee Spacagna: 这种高度自动化的流程直接为他‘赢回了每天的第一个小时’。 [8:54] Lee Spacagna: 每位员工在步入会场时，都仿佛配备了一位专属的‘幕僚长’ Agent 为其做足了背景梳理与信息铺垫。 [9:50] Lee Spacagna: 这项技术带来的真正机遇绝不仅限于交付某一个孤立的‘单点自动化项目’，而是推动组织迈向一种全新的‘企业运营模式’。 [10:14] Lee Spacagna: Agent 能够在交互中持续学习，并随着时间的推移评估并优化自身表现，实现‘越用越聪明’。

[0:00] Lee Spacagna: where can AI actually change how a business runs? [2:07] Lee Spacagna: You don't need prompt engineering skills, you don't need any technical skills at all... as a business user, you can now use an agent to build another agent for you just with natural language.

[0:00] [music] [0:05] >> Hi everyone. My name's Lee and I'm a [0:07] lead solution engineer here at Open AI, [0:09] working very closely with financial [0:10] services customers across EMEA. [0:12] Today, we've got some really exciting [0:14] things to show you, some of which we [0:15] only launched last week. [0:17] Every day when I work with financial [0:18] services institutions, they always have [0:20] one question, where can AI actually [0:22] change how my business runs? [0:25] Today, there are two paths for AI [0:27] adoption. First, giving ChatGPT and [0:30] Codex so that employees can use AI in [0:32] their daily work. [0:34] Second, there's systems. This is where [0:35] companies are building entirely new [0:37] products, they enhancing their customer [0:39] service, they're improving client [0:41] advisory, and they're working on [0:43] operational support. [0:45] So, it looks like this. We've got [0:46] ChatGPT and Codex working for the bottom [0:49] up using employees as they get more AI [0:51] literate. [0:52] And then we've got AI systems from [0:53] top-down, and these are those major [0:55] transformational initiatives. [0:58] But there's a missing layer in the [0:59] middle. [1:00] The automation at the team and the [1:01] department level. And this is a gap that [1:03] the new ChatGPT workspace agents is [1:05] designed to close, and that's what I'll [1:07] be demoing for you today. [1:09] When we say agents, we mean AI systems [1:12] that we can delegate meaningful tasks [1:13] to, not just ask questions of. [1:16] And we can do that by using tools that [1:17] we already rely on, like email, [1:19] calendar, and those productivity apps [1:21] that we're using every day. [1:23] And we want these agents to complete [1:24] work in the same way that people do. [1:26] And in the last few months, there's been [1:28] a huge jump in capabilities. And we had [1:30] another leap last week with GPT 5.5. [1:33] Today, agents take on complex work that [1:35] used to take hours or days, and they can [1:37] handle them from start to finish. [1:40] But what happens when you need to build [1:41] something custom, and you need to [1:42] delegate something that your team or [1:44] your departments are currently working [1:45] on? [1:47] Many of you have already built custom [1:48] GPTs. [1:49] With workspace agents, we evolved that [1:51] into something much more powerful. We've [1:53] got a new agent builder, which brings [1:55] shared applications, skills, and [1:56] deployment all to one platform. And this [1:59] allows these agents to work in the same [2:00] place that work already happens. [2:02] So now let's jump into the demo. [2:07] This is the is the standard chat GPT [2:09] interface that I'm sure you're all [2:10] familiar with. [2:11] But down the left-hand side, you can now [2:12] see we've got an agents option that we [2:14] can start with. [2:17] And now for many teams, the challenge [2:18] isn't a lack of work to automate, it's [2:20] that the work is spread across meetings, [2:21] documents, emails, and other systems. [2:24] And the decisions all depend on specific [2:25] context. [2:27] So today, I'm going to show you how I [2:28] can quickly spin up an Agenty co-worker. [2:31] In this case, I want to build my own [2:32] chief of staff agent. [2:34] I want it to help me coordinate work, [2:36] track priorities, prepare meetings, and [2:38] help keep the team moving. Every [2:40] function can delegate meaningful work to [2:41] agents, and these can understand the [2:43] role, use the right tools, and they can [2:45] operate with how the team already works. [2:47] Here, I'm going to use one of the [2:49] templates we've got already for the [2:50] chief of staff agent here, and you can [2:52] see this already has a set of [2:53] instructions. It already has a set of [2:55] tools that it's able to work with, and [2:57] it already has capabilities that I can [2:59] start using. [3:02] Next, you can see here I want to start [3:05] connecting some of those tools that I [3:06] mentioned to make sure it's correct for [3:07] my workflow. In this case, I'm going to [3:09] use the Microsoft set of tools here. So [3:12] we've got things like Outlook Calendar, [3:14] Teams, and then my Outlook email. [3:20] Now, we can see that the instructions [3:21] are going to be automatically written by [3:23] another agent. So you don't need prompt [3:25] engineering skills, you don't need any [3:27] technical skills at all. [3:29] But what it means is that as a business [3:30] user, you can you now use an agent to [3:32] build another agent for you just with [3:34] natural language. [3:37] And that's it. That's the initial [3:38] version. I haven't written any code, and [3:40] I've got the first version of the agent [3:41] ready to go. [3:43] But now I want to customize this to my [3:45] own requirements. I want it at 9:00 [3:47] a.m., I want it to run every day. I want [3:49] it to look at all of my meetings, look [3:51] at all of the applications, look at my [3:53] emails that came in overnight, and [3:54] generate a daily brief so that I could [3:56] arrive and be prepared for all of the [3:57] meetings for the day. [4:01] So, once again, here I just give the [4:02] instructions again in natural language, [4:04] telling it I want it to run at 9:00 a.m. [4:06] No technical skills needed here at all. [4:09] And within a matter of seconds, we can [4:10] see that it's able to customize my [4:11] agents to my team's requirements and [4:13] work exactly how I like to work every [4:15] day. [4:19] So, once this is done, we can go and [4:21] test our agent. [4:29] And we can now see there's two starter [4:30] prompts underneath to get me started. If [4:32] I want to, I can give it its own set of [4:34] instructions to perform for me, but you [4:35] can see that there's two already there [4:37] to go. And you can think of these as [4:38] capabilities that have already been [4:40] built into this agent. [4:42] So, let's ask it to prepare the today's [4:44] brief. [4:45] Here, you you can see I'm asking it to [4:46] do a concise brief using all of the [4:49] available information that I mentioned [4:50] earlier, highlight priorities, [4:52] decisions, blockers, and follow-ups, and [4:54] then post these in the CFO team channel [4:56] in the daily prep channel. [5:00] And now we can see the agent spinning [5:01] up. [5:02] It's going to start grabbing all of [5:03] those details, connecting into my email, [5:05] connecting to my calendar, and all those [5:07] other sources that I mentioned. It's [5:09] going to check all of those meetings [5:10] that I've got for the day. It's going to [5:12] cross-reference that with information [5:13] that might be in my emails. It's going [5:15] to pull all of the contacts needed from [5:17] all of these sources. [5:19] And the first time it runs, it's going [5:20] to ask me to uh give permission to post [5:22] to the Teams channel, so it's just going [5:23] to set that up now. And we'll go and [5:25] give it the the approval and see how [5:26] that's worked. [5:30] And that's it. That's now posted to [5:31] Teams. So, let's now go and have a look [5:33] at what it was able to generate for me. [5:37] So, now over in Teams, we can see in the [5:39] daily prep channel, we can see that [5:40] there's an update, so let's go and have [5:42] a look what it posted. [5:43] And we can see our chief of staff agent [5:45] from ChatGPT has gone and collected all [5:47] that information and then it has posted [5:48] that daily brief for me inside Teams, [5:51] exactly where I want the information to [5:52] be for my daily work. [5:54] So within a couple of minutes, we've [5:55] built an agent from scratch. We've [5:57] connected it to tools that I use every [5:59] day. We've given it some customized [6:00] guidance and we now have a running chief [6:02] of staff agent for my whole team. [6:06] But let's go back to the agent and take [6:08] it a step further. [6:09] This This week my team have been burning [6:11] themselves out running from meeting to [6:13] meeting and they haven't had any time to [6:14] prep in between. [6:16] So now let's add a new capability. I [6:18] want the agent to proactively research [6:20] before every meeting, like having an [6:22] expert chief of staff who's the telling [6:24] me who's there, what's the latest and [6:26] then what's the goal of that meeting. [6:28] So for this we need some additional [6:29] context for some other tools. So now [6:31] let's go and add some more that are [6:33] available. [6:34] I'm going to start off with SharePoint. [6:37] This is where I'm storing all of the [6:38] company information, all the information [6:39] that I've taken as notes that's shared [6:41] across the organization. So we're going [6:42] to add that. [6:44] And next I want to add Salesforce all [6:46] that CRM and all that kind of rich [6:47] information about all the contacts from [6:49] that customer. [6:50] So we're going to add Salesforce as [6:51] well. [6:55] And now that's done. We've got those two [6:56] apps connected. [6:58] You can also add other apps that you use [6:59] every day here as well or even custom [7:01] applications that you just you have [7:02] inside your business. Next is skills. [7:05] Skills are a way of capturing snippets [7:07] of information instructions to perform [7:09] critical tasks. [7:10] Think of these as an amazing way to [7:12] capture all of that those tribal [7:13] knowledge and conventions that are [7:14] currently trapped in people's heads and [7:16] we can turn those into repeatable [7:18] workflows. [7:20] You can see that there's two skills [7:21] already in use here. We've got the chief [7:22] of staff skill and we've got a final [7:24] brief formatting skill. [7:27] But now let's go and add another one [7:28] that I've already been using across my [7:30] team. So I've already got a skill here [7:32] for meeting prep. This tells chat GPT [7:34] the way that I want this information to [7:35] be structured, the key information [7:37] that's needed, the source of this [7:39] information and where I want that [7:40] information to be posted. So let's go [7:42] and add that to my agent as well. [7:46] Finally, let's save our changes. [7:50] And now I want to give the agent some [7:51] more instructions about what to do with [7:52] these new applications that I've gone [7:54] and connected. So, again, we'll use the [7:56] agent on the side to have a natural [7:57] language conversation and give it this [7:59] additional context to go and update the [8:01] agent. [8:04] So, here there's all the information. [8:05] I've just added Salesforce and [8:06] SharePoint, add a new capability. I want [8:09] it to be able to generate these quick [8:10] meeting briefs in in ChatGPT. And all [8:13] the information I want to give it is [8:14] just give me the information for the [8:15] next meeting. [8:16] So, it needs to go through here and make [8:17] the updates. Um and then it's it will [8:20] also um add a new starter prompt there [8:23] for me to use in a second. [8:25] So, now again, let's go and update this. [8:30] And now we can go and deploy this and [8:31] it's now available for the whole team. [8:34] So, let's go and use it. [8:36] And here is my completed agent, deployed [8:38] and ready. And now you can see we've now [8:40] got a third starter prompt underneath as [8:41] well to prepare me for my next meeting. [8:47] Now it's going to run, pull all of those [8:49] contacts from those different sources [8:50] including Salesforce [8:52] um and SharePoint that I went and added. [8:54] It's going to go and um to check all of [8:56] the information, put that together into [8:58] a concise brief in the way that the [9:00] skill gave the instructions to go and [9:01] represent that. And personally, I have [9:03] one of these running every day. [9:05] Um I have my own agent that checks all [9:07] of my emails that come in overnight, the [9:08] important updates from across the [9:10] business, the things that I said I would [9:12] do on Slack yesterday or on calls and [9:13] all of the context from the transcripts. [9:15] And now it means I get that first hour [9:17] of my day back cuz I come into the into [9:19] work in the morning and all of my emails [9:21] have a draft ready to go with all of the [9:23] context from across the business. And it [9:25] means that I can just go through all of [9:26] those emails and click send and just [9:28] approve those and get those out to my [9:29] customers. And it's completely [9:31] transformed the way that I work. [9:34] So, So we can see here we're all ready [9:36] to go for that next meeting. [9:38] Before the team was understaffed and [9:39] they couldn't prep for those meetings, [9:41] but now we've enabled everyone to turn [9:42] up as if they've been prepped by their [9:44] own chief of staff agent. [9:50] But that was just one example there, but [9:52] this is a pattern that you can apply [9:53] across the business. [9:55] We've seen examples of agents like KYC [9:57] onboarding, AML investigations, [9:59] relationship management, and more. [10:01] The opportunity here isn't about one [10:03] single automation project. It's actually [10:06] about a brand new operating model. Every [10:08] team can spin up a role-specific agent [10:10] to take manual work off their plate and [10:12] help the business move faster. [10:14] But the next question is is what if [10:16] we've got thousands of these agents? How [10:17] do we manage them? [10:19] And that's where Frontier comes in. [10:21] Frontier is our platform for deploying [10:23] and managing agents at scale. [10:26] It connects to systems usually in silos, [10:28] things like data warehouses, CRM, and [10:30] internal applications. [10:32] It gives AI co-workers the same shared [10:34] context that the teams currently rely [10:36] on. [10:38] And from there, agents can reason over [10:39] data, they can run code, they can use [10:41] tools, and they can take actions all in [10:43] a governed environment. [10:46] And the key thing here as well is as [10:47] they work, the system improves. They [10:49] will learn from interactions, they will [10:51] evaluate their performance over time, [10:53] and it means that the more they do, the [10:54] better they get, just like human workers [10:56] in the business right now. [10:59] So today it's possible to build in [11:01] ChatGPT, Codex, and the API, and we want [11:04] to make it easier to deploy [11:05] out-of-the-box agents, plugins, and [11:07] skills all specific for financial [11:09] services workflows. [11:11] This matters because it moves the system [11:12] towards much more automation. [11:14] We can use purpose-built agents that [11:15] plug directly into work, and they can [11:17] handle all of the repeatable processes [11:19] with even less lift and customization. [11:22] With all those foundations in place, [11:24] agents become incredibly powerful, [11:26] allowing you to delegate more workflows [11:27] to AI over time. And next, Stephanie is [11:30] going to show you how teams are using [11:31] them to create transformative impacts [11:33] across the workforce. Thank you. [11:36] >> [applause]