Skan AI Maps Screen Activity to Ground AI in Operational Rules

Skan AI映射屏幕活动，将AI锚定于业务规则 企业级AI智能体在生产环境中往往会失灵，原因在于静态操作手册忽略了实际工作情况；这需要通过实时屏幕遥测来构建确定性的上下文图谱，从而降低Token成本并避免业务故障。 本次讨论探讨了为何上下文而非基础模型的选择正在成为企业级AI的核心持久优势。Avinash Misra解释了依赖静态SOP和RAG的传统方法为何会导致生产环境中的智能体失灵，而直接从员工屏幕捕获动态的“工作遥测数据”则能提供精确的业务图谱。通过将工作视为一种语言，并将确定性的业务规则与概率性LLM相结合，企业可以大幅降低Token成本，并在工作流程向自主智能体转型的过程中妥善应对“流程崩溃”。最后，讨论强调企业必须构建并掌握自身独特的竞争性上下文，而不是仅仅依赖租用的模型智能。 企业工作上下文图谱相当于高精度的地图，而基础模型则是行驶在这一版图上的驾驶员。 (0:08) 企业工作流具有严格的确定性，受规则、策略和合同约束；概率性AI必须立足于这种现实业务场景，而非凭第一性原理去盲目推测。 (0:08) 标准作业程序（SOP）和文档仅代表意图声明，而真实的业务运营则存在于实时的动态工作遥测数据和执行历史之中。 (3:08) 在生产环境中，通过传统RAG基于静态文档来落地AI往往会失效，因为它缺乏对现实中的异常情况、权宜之计以及人类决策细微差别的了解。 (3:08) 人类工作流数据包含大量噪声，但持续的屏幕特征化处理使AI系统能够过滤干扰并提炼出结构化的执行上下文。 (7:08)

Avinash Misra of Skan AI argues that enterprise AI agent performance depends primarily on the quality of an organization's operational context rather than frontier model intelligence alone. Because enterprise workflows are strictly deterministic, organizations must capture kinetic, real-world execution telemetry rather than relying on static documentation. Deterministic Business Rules Versus Probabilistic Generation (0:08) Enterprise operations such as loan approvals and insurance claim adjudications are governed strictly by regulations, contracts, and established corporate policies. According to Misra, these workflows are fundamentally deterministic, yielding binary outcomes like approval or denial. In contrast, large language models are probabilistic engines. Misra contends that enterprise systems should not rely on probabilistic models to infer core business rules from first principles, just as humans rely on memory rather than deriving arithmetic repeatedly. Grounding probabilistic AI within deterministic operational parameters prevents hallucinations and significantly reduces computational overhead. Misra highlights that constraining a model to select among validated deterministic pathways conserves electrical power and minimizes token consumption, addressing the unviable tokenomics that often arise when unconstrained models generate processes from scratch. Kinetic Work Telemetry Versus Declared Process Intent (3:08) A recurring barrier in enterprise AI deployment is the reliance on standard operating procedures, manuals, and system documentation. Misra characterizes these static documents as mere declarations of intent that depict how work ought to occur in theory, failing to reflect how work operates in practice. Traditional Retrieval-Augmented Generation systems grounded exclusively in static documentation break in production because they miss the real-world edge cases, workarounds, and nuances encountered by human workers. To bridge this gap, Skan AI advocates capturing the telemetry of work by directly observing employees' screens across disparate applications. While human workflows contain substantial ambient noise—such as open browser tabs and unrelated deviations across multi-step tasks—screen featurization isolates core decision-making variables, such as customer locations or claim loss attributes. Misra argues that as foundational models commoditize raw intelligence, an enterprise's primary lasting competitive advantage resides in owning this proprietary execution history. Bridging the Sandbox-to-Production Reliability Gap (9:41) Enterprise AI agents frequently perform well in controlled sandbox environments but fail when deployed into production. Amit Patel and Misra identify the sandbox environment as a misleading baseline because it relies on artificially sanitized workflows with minimal variations. When faced with live production across millions of operational variations, agents lacking deep execution context falter. Production readiness requires training systems on historical execution paths so that agents possess situational awareness. In this framework, an agent autonomously resolves scenarios with established institutional precedent and hands off anomalous edge cases to human personnel. Patel emphasizes that enterprise architectures must be designed for full-scale production from inception to prevent systemic failure and uncontrolled API costs. Work as Language and Agentic Process Collapse (16:51) Misra parallels natural language processing architectures to business execution by framing work as a language. In this model, individual screen interactions and user interface components function as words, while end-to-end business cases—such as onboarding or claim processing—serve as complete sentences. By predicting the next token across aggregated execution streams, models identify branching logic across jurisdictions and systems without manual process mapping. Historically, enterprise user interfaces were designed to accommodate human cognitive limits, displaying around 15 elements rather than hundreds because humans can maintain only a few details in working memory. As autonomous agents assume these tasks, procedural steps undergo process collapse, collapsing fragmented navigation across mainframes, web apps, and desktop tools into direct, agentic execution. The presentation concludes that while foundational models and generic intelligence can be rented, enterprises must capture and own their internal context graph. In the speakers' framing, the frontier model functions merely as the driver, while the competitive differentiator remains the resolution and accuracy of the enterprise's operational map.

在 Ai4 大会上，Skan AI 联合创始人兼首席执行官 Avinash Misra 与 Plug and Play Ventures 合伙人 Amit Patel 提出，基础大模型在企业端的落地成效取决于其所处的业务上下文。通用大模型的智能正在迅速商品化，企业竞争的关键不再是获取更聪明的“驾驶员”，而是构建高精度的企业业务“地图”，即将概率性的人工智能严格约束在确定性的业务规则与真实动态执行历史中。 确定性业务与概率性模型的边界 (0:08) Avinash Misra 指出，企业级核心工作流（如银行贷款审批或保险理赔结算）在本质上是确定性的，严格由成文政策、监管法规与合同条款支配，其决策多为二元结果。在现实中，人类无需每次都从算术公理重新推演“2 + 2 = 4”，而是直接调用既有经验与记忆；同理，作为概率性计算引擎的大语言模型（LLM），不应在缺乏约束的情况下凭空猜测或推演企业核心业务规则。 Amit Patel 强调，企业领导者必须清晰界定这一技术边界。将概率性 AI 坚实锚定在企业现有的合规框架、合同限制以及过往实际执行历史中，才能确保系统选择合规的执行路径，避免因模型随机发散而产生不可控的运营与合规风险。 动态执行遥测取代静态意图声明 (3:08) Avinash 区分了“意图声明”与“动态执行遥测”的本质差异。企业内部的标准作业程序（SOP）、操作指南与系统日志仅代表理论上“应该如何工作”的意图声明，并未反映运营现实。真实的日常业务是由员工处理的“鲜活历史”构成的，充斥着各种例外情况、临时替代方案与边缘案例。企业若仅依靠过时静态文档构建传统的检索增强生成（RAG），AI 智能体在面对生产环境的真实波动时极易失效。 为构建真实的企业上下文图谱，Skan AI 通过静默观察技术在数百名员工的操作屏幕上实时捕获跨系统交互行为。Avinash 认为，当前沿大模型抹平了通用智能的差距后，企业唯一能够建立并长期拥有的核心竞争护城河，正是日常实际业务在动态执行中所沉淀的高分辨率遥测数据与历史经验。 沙盒陷阱与规模化部署的令牌经济学 (9:41) Amit 援引美国前五大银行之一的首席人工智能官的观点指出，企业级 AI 工具必须能够经受住数百万种工作流变体的考验才能在生产中存活。然而，大量在概念验证沙盒中表现出色的智能体上线后便迅速崩溃。Avinash 将纯净的沙盒测试称为脱离现实的“陷阱”，因为沙盒往往剔除了边缘场景与干扰，赋予团队虚假的安全感，反而成为走向规模化生产的阻碍。 企业必须从架构设计之初就面向全组织生产级规模进行规划，沃尔玛等大型企业的成功正源于严格的前期技术尽调。此外，将大模型限定在经过验证的确定性流程边界内，能有效遏制“令牌经济学”的成本膨胀。相比于让模型从零开始自由生成方案，在确定性路径中引导大模型进行判断，可大幅减少 API 令牌消耗、削减运维成本并降低能源消耗。 工作语言化与界面认知限制下的流程坍缩 (16:51) Avinash 将 Skan 的核心方法论类比于大语言模型的预训练过程：OpenAI 通过海量文本训练模型预测下一个标记（Token）以理解语言；Skan 则将企业的实际业务视作一种语言，员工屏幕上的 UI 组件、点击步骤和交互操作是“词汇”，而端到端业务案例（如从接案到裁决的完整理赔）则构成“句子”。通过海量学习此类执行序列并预测下一个行为步骤，系统能够掌握组织特有的运作模式，如自动识别不同地理区域（如加州与得州）在合规流转路径上的分支差异。 与此同时，传统企业软件（涵盖大型机、网页与桌面客户端）受限于人类的生理与记忆带宽，界面设计通常只能展示约 15 项信息以配合人类 3 至 5 个细节的短期工作记忆。随着工作流向自主智能体迁移，将引发“流程坍缩”：智能体无需通过破碎繁复的人机界面跨系统搬运数据，而是直接在底层贯通流程。绘制人类操作软件的遥测路线图，既优化了当下的员工效率，也为智能体实现流程坍缩提供了结构化执行路径。 租用通用模型与掌控私有上下文资产 (13:36) 针对业界流行的“租用模型，拥有智能”观点，Avinash 提出了关键修正：企业所租用的是基础模型及其底层的原始智能，而企业自身必须全力构建并掌握专有的“竞争性上下文”。他做出历史类比：二十年前企业雇佣的人类员工本质上也是运行在颅骨内的 20 瓦特生物模型，新员工入职必须经过系统培训以适应企业专有业务，外部 AI 模型进入企业同样需要这种深度的上下文磨合与业务落地流程。 Amit 提醒管理者防范盲区：仅将模型部署在私有云容器中并不能完全杜绝洞见流失，模型厂商依然可能借交互推断出企业的决策与评估逻辑。Avinash 强调，业务决策背后的具体推理逻辑（如客户准入属性、地域条件、收入层级及具体理赔参数）必须作为核心资产留在企业内部。Skan 借助长时间系统观察与屏幕特征化分析，将员工脑中的隐性决策逻辑转化为企业专属的数据基础设施。 Avinash Misra 与 Amit Patel 最终向企业高管重申：在人工智能深度融入业务的进程中，企业效率的最终分水岭并不取决于底层通用大模型“驾驶员”的智商高低，而取决于企业能否为驾驶员配备一份详尽描绘自身实际动态运作的“高精业务地图”。

## Background 该视频由 YouTube 频道 Ai4 于 2026 年 8 月 22 日发布，记录了 Skan AI 联合创始人兼 CEO Avinash Misra 与 Plug and Play Ventures 合伙人 Amit Patel 的深度对谈。内容主要探讨为何前沿大模型在进入企业真实场景时容易因缺乏上下文而失效，并阐述了如何通过持续观测人机交互的真实工作流构建“企业工作上下文图谱”（Enterprise Context Graph of Work），以此作为企业 AI 落地中不可复制的持久竞争壁垒。 ## TL;DR 对话深入探讨了当前企业 AI Agent 在沙盒演示中表现出色、一旦进入真实生产环境便迅速崩溃的根本原因——传统 SOP 文档与系统日志仅代表管理意图，无法反映实际业务中包含数百万种细微变体的真实执行路径。Skan AI 提出“将工作视为一种语言”，通过在屏幕执行平面上并发捕获并特征化员工的真实操作，让模型从端到端的业务流中自发涌现对业务规则与分支逻辑的深层认知。对谈强调，在基础大模型与通用智能逐渐大宗商品化的时代，决定企业 Agent 成败的关键不在于“AI 驾驶员”本身的智商高低，而在于其手中是否握有一张详尽、动态且反映真实工作现状的“企业级高精上下文地图”。 ## Atmosphere & style 整场对谈兼具轻松幽默与高密度的商业技术洞察，两位对谈嘉宾作为长期合作伙伴配合默契，从疲惫换鞋的调侃自然切入硬核的企业架构思辨。现场交流氛围务实而犀利，嘉宾用极其通俗且具画面感的生活比喻（如“驾驶员与高精地图”、“20瓦人脑大模型入职”）直击当下企业盲目追求大模型底座的认知误区。随着对谈深入到“流程坍缩”与隐性知识捕获等前沿议题，现场情绪层层递进，既有对技术落地痛点的冷峻剖析，又洋溢着对重塑未来企业人机协作范式的坚定信心。 ## Key takeaways - [00:08](https://www.youtube.com/watch?v=2kjXV50-iiA&t=8s) 基础模型的能力上限取决于其背后的企业上下文支撑；模型如同“驾驶员”，而企业上下文则是指引其行车的“高精地图”。 - [01:36](https://www.youtube.com/watch?v=2kjXV50-iiA&t=96s) 企业决策本质上是受政策与合同约束的确定性系统，AI 的核心任务不是从零发明路径，而是将概率性模型锚定在既有规则与执行历史中做出抉择。 - [03:08](https://www.youtube.com/watch?v=2kjXV50-iiA&t=188s) 企业 SOP 和规章制度仅代表“工作开展的意图声明”，真实的商业运作充斥着大量例外与变体，必须依靠对屏幕实际操作的“被观测工作模型”来还原真实状态。 - [07:08](https://www.youtube.com/watch?v=2kjXV50-iiA&t=428s) 沙盒环境中干净的数据与极少的变体是扼杀生产落地的“毒药”，仅靠文档 RAG 构建的 Agent 进入拥有数百万变体的真实生产环境必然崩溃。 - [13:44](https://www.youtube.com/watch?v=2kjXV50-iiA&t=824s) 大模型与通用智能正迅速沦为大宗商品，企业无法垄断模型本身，真正能构建核心竞争壁垒与护城河的是其独有的“竞争性上下文”。 - [16:51](https://www.youtube.com/watch?v=2kjXV50-iiA&t=1011s) Skan AI 将企业工作流程视为一种“语言”，将屏幕 UI 操作与跨系统流转作为“词汇”，案件生命周期作为“句子”，从而通过行为预测自然涌现出对业务流程的认知理解。 - [18:02](https://www.youtube.com/watch?v=2kjXV50-iiA&t=1082s) 现有 UI 界面本质上是受限于人类狭窄上下文窗口的妥协产物；随着 AI Agent 接管业务，多系统间繁琐的人机交互将发生彻底的“流程坍缩”。 - [20:28](https://www.youtube.com/watch?v=2kjXV50-iiA&t=1228s) 在各家底座大模型智力水平逐渐收敛趋同的背景下，决定企业 AI 落地成败的关键壁垒在于模型手中所拥有的上下文地图质量。 ## Detailed walkthrough ### [00:08](https://www.youtube.com/watch?v=2kjXV50-iiA&t=8s) 开场介绍：Plug and Play Ventures 与 Skan AI 的碰撞 对谈在轻松幽默的氛围中拉开序幕。Amit Patel 开场打趣说，经历了一整天冗长的会议，双脚早已疲惫不堪；从上一次参会吸取的血泪教训就是“千万别穿皮鞋”，于是这次换上了舒适的 Toms 帆布平底鞋。随后他介绍了自己的背景：作为全球风险投资机构 Plug and Play Ventures 的合伙人，他们搭建了全球最大的企业创新网络，常年合作的巨头包括 Nike、Walmart 以及 Mercedes-Benz 等，帮助这些大企业对接生态网络中的初创技术，并利用这些真实互动反馈来做出投资决策。Plug and Play 投资十余年，已孕育出 38 家独角兽企业。 随后，Skan AI 的联合创始人兼首席执行官 Avinash Misra 登台致意。他回顾道，Plug and Play 正是 Skan 多年前创业之旅最早的投资机构之一。Misra 引用了 Skan 的核心座右铭——“基础模型的能力上限，取决于它背后的上下文支撑”（A foundational model is only as good as the context behind it）。为了让听众更直观地理解，他给出了一个极其生动的比喻：“当全世界都沉迷于寻找马路上最聪明、最强悍的驾驶员时，我们却专注于为你需要行驶的区域绘制最详尽、最精准的高清地图。”在企业级环境中，真正决定自主代理（Agency）与智能（Intelligence）能否落地的，恰恰是驱动它们的具体上下文（Context）。 - **对谈嘉宾**：Amit Patel（Plug and Play Ventures 合伙人）与 Avinash Misra（Skan AI 联合创始人兼 CEO）。 - **核心业务生态**：Plug and Play 旗下拥有庞大的企业创新网络与 38 家独角兽投资战绩；Skan AI 早期即获得其投资。 - **核心类比**：基础模型与 AI 代理如同“驾驶员”，而企业上下文则是那张指引行车的“高精地图”。 --- ### [01:36](https://www.youtube.com/watch?v=2kjXV50-iiA&t=96s) 确定性与概率性 AI：企业决策的基石 Amit Patel 将话题切入技术底层，提问在构建企业工作上下文时，为什么理解“概率性 AI（Probabilistic AI）”与“确定性系统（Deterministic System）”的区别至关重要。Avinash Misra 解释道，在任何真实的企业场景中——不论是银行还是保险公司——一笔理赔申请要么批准、要么拒绝，一笔款项要么支付成功、要么失败。这些行为绝对不是凭空摸索出来的，而是受到严格的组织政策（Policy）或合同（Contract）所约束。正如人类在做基本算术时，算“2 + 2 = 4”根本不需要每次都退回第一性原理去推演一遍，而是依赖既有的记忆与既定规则；企业组织内部的大量行为，同样早已被编码成确定性的章程与条款。 Misra 尖锐地指出，当前许多企业误以为必须让基于大语言模型（LLM）的概率性引擎从零开始重新发明轮子、自己去琢磨该怎么做决定。但正确的做法截然相反：我们必须把这些概率性引擎“锚定”（Ground）在企业政策、合同的运营现实以及组织真实发生过的执行历史中。AI 的职责不是天马行空地重新发明路径，而是在经过验证的既有正确路径之间做出精准抉择。Amit 总结道，这意味着 AI 必须真正理解业务决策发生时的各种细微差别（Nuances）与变体（Variations）。 ```mermaid flowchart TD accTitle: 概率性模型在确定性上下文中的锚定架构 subgraph S1["企业确定性基石 (Deterministic Reality)"] A["政策与合同条款<br/>(Policies & Contracts)"] B["真实执行历史与记忆<br/>(Lived Execution History)"] end subgraph S2["概率性推理引擎 (Probabilistic Engine)"] C["前沿大模型 / AI Agents<br/>(LLMs)"] end subgraph S3["落地决策成果 (Production Decision)"] D["在合规路径间筛选抉择<br/>(Choose Between Right Paths)"] E["业务结果：批准/拒赔/转账<br/>(Deterministic Outcome)"] end A --> C B --> C C --> D D --> E ``` --- ### [03:08](https://www.youtube.com/watch?v=2kjXV50-iiA&t=188s) 从“工作遥测”到“上下文图谱”：真实执行与书面文档的鸿沟 Amit 强调了 Misra 的前瞻性。虽然近半年以 Foundation Capital（特别是 Jaya Gupta 的博文）为代表的风投界才开始大谈“企业上下文图谱”（Enterprise Context Graphs），但早在 2019 年两人初次会面时，Misra 就提出了类似的理念，当时将其称为“工作遥测”（Telemetry of Work）。在与无数首席转型官（CTO）、首席信息官（CIO）和首席运营官（COO）的沟通中，Amit 发现大家都渴望在 AI 时代重构工作流，但行业内充斥着浮于表面的流程挖掘工具（Process Discovery Tools），根本无法捕获真实的决策变体。而 Skan 却凭借对细微业务差异的深度捕获，赢得了全美前五大银行和顶级保险公司的信任。 面对 Amit 提出的核心问题——“为什么靠企业现成的 SOP 和文档无法构建真正的模型，唯有依靠‘被观测的工作模型’（Observed Model of Work）？”，Misra 做出了深刻的剖析。他指出，SOP 和规章制度并非毫无价值，但它们只是“企业对工作应当如何开展的意图声明（Declarations of Intent）”，根本不是“工作实际发生的真实状态（How work is done）”。真实的商业运作充满了例外、排列组合与权变细节，全部隐藏在组织动态的“动能工作”（Kinetic Work）与活着的生活史中。 因此，如果仅仅靠传统的检索增强生成（RAG）从文档中拉取知识并喂给 AI，这样的 Agent 投入生产环境必然崩溃。AI 必须直面真实的“执行平面”（Execution Plane），学习人类在处理复杂业务时所做出的实时决策。Skan 的底层技术通过在数以百计的员工屏幕上进行并发观察，将屏幕发生的行为进行特征化（Featurizing），理解跨系统、跨人员的真实操作，从而将这种蕴含企业竞争优势的动态上下文注入模型。 --- ### [07:08](https://www.youtube.com/watch?v=2kjXV50-iiA&t=428s) 破解沙盒与生产断层：从个体决策到千人千面的规模化 Amit 进一步分享了前一天与一位财富 50 强（Fortune 50）企业 AI 负责人的私下对话。该高管正全力打造其运营模型的“数字孪生”（Digital Twin），但他提出一个致命痛点：企业级记忆（制度与意图）只是宏观框架，真正驱动业务运转的是“个人层面的决策记忆”（Individual-level Decision Making）。要实现真正自主的 Agent，仅靠机构级汇总数据远远不够，必须抓取到微观个体的判断。 Misra 完全赞同这一观察。他指出，机构层面的信息是意图，而地面的执行才是该意图的具象化。然而，地面执行的数据极其嘈杂——任何人在执行包含 $N$ 个步骤的任务时，中途都会产生大量无关的分心操作与路径偏离。Skan 的核心技术就在于从极其嘈杂的真实屏幕数据中提炼出纯净的上下文，以此锚定模型，让原本依赖人工干预的复杂流程得以在 AI Agent 手中被“折叠”（Collapse），再交由执行层运转。 谈及当下企业 AI 的困境，Amit 指出业界正刮起一股浮夸的“全员 Agent 化”之风，但现实却是：“Agent 在沙盒（Sandbox）里表现亮眼，一进生产环境就彻底崩盘。”几天前某前五大银行的首席 AI 官也向他直言，银行内部的业务工作流包含“数以百万计的变体（Millions of Variations）”，无法 scale 的产品毫无价值。 Misra 犀利地点破了沙盒的陷阱：“沙盒中干净的数据与少量的变体，正是最终毒死生产落地的毒药（The poison to getting into production）。”在 Pilot 试点中轻松跑通的演示，因为严重低估了现实的尺度与复杂度，直接成了上线的拦路虎。要跨越这道鸿沟，唯一的办法是在设计之初就引入详尽的执行路径与被观测到的真实业务变体，使模型在面对历史异常时具备真正的感知能力；一旦遇到从未见过的情形，能够安全地将其转交人工处理。 ```mermaid flowchart LR accTitle: 沙盒陷阱与生产环境现实对比 subgraph Sandbox["沙盒试点环境 (Sandbox / Poison)"] direction TB S1["流程高度理想化"] --> S2["极少变体与例外"] S2 --> S3["Demo 演示成功<br/>(虚假繁荣)"] end subgraph Production["生产环境落地 (Production Grounding)"] direction TB P1["数百万真实路径变体"] --> P2["注入被观测工作图谱<br/>(Enrich with Observed Realities)"] P2 --> P3["自主处理历史变体 +<br/>遇未知转交人工 (Human-in-the-loop)"] end S3 -.->|"盲目上线直接崩溃"| P1 ``` - **生产落地的设计原则**：企业（如沃尔玛）必须从“第一天”（Day One）就以全企业规模化为目标进行架构设计，考量真实运行中的 Token 经济学（Tokenomics）与缓存机制，杜绝盲区。 - **计算与能耗优势**：若能利用确定性上下文将 LLM 的输出空间约束在有限的确定性选项中，其 Token 消耗和能源开销将远低于让模型自由发挥。 --- ### [13:44](https://www.youtube.com/watch?v=2kjXV50-iiA&t=824s) 租赁智能，坐拥上下文：企业独有的竞争护城河 随后，对话转向了近期科技界由 Satya Nadella 和 Alex Karp 等人引发的热门论调——“租赁模型，拥有智能”（Rent the model, own the intelligence）。模型正迅速沦为无差别的大宗商品（Commodity），但 Misra 对此给出了更深一层的修正： > “你不仅可以租赁模型，你其实连智能本身也是租来的。企业真正需要拥有并誓死捍卫的，是你自己的‘竞争性上下文’（Competitive Context）。” Misra 用一个生动的历史类比阐释了这一点：20 年前，企业其实就已经在使用“大模型”了，只不过那些模型装在消耗 20 瓦能量的人类大脑（Cranium）里。当招募一个新人进公司时，企业必须对他进行流程培训、灌输企业内部无数的业务细节与微妙规范，他才能产生生产力。今天，哪怕换成了不装在头骨里的外部超大 AI 模型，这一套“入职入轨”（Onboarding）的上下文注入逻辑没有任何改变。 Amit 补充道，许多企业合规部门天真地以为“只要模型部署在自己的本地私有环境里，数据就不会泄露”。但他们根本没意识到，前沿模型实验室所沉淀的，远不止是静态数据，更是每一次决策的判定、每一次例外的处理逻辑、何为正确答案与何为错误答案的深层洞见。未来企业采购中，对于“上下文归属权”的博弈将变得尤为激烈。 在问答环节，现场观众立刻提问：“你们究竟是如何捕捉决策背后的推理过程与约束条件的？”Misra 回答道，这依靠两步关键机制： 1. **长时间全量暴露**：让模型持续、全面地暴露在企业长周期的真实执行环境之中； 2. **屏幕全维度特征化（Featurizing the screen）**：无论界面如何变化，系统会把屏幕上的所有关键属性实时特征化——例如客户入职时的所在州、住址、收入水平，或者保险理赔单中的损失类型与报案发生州。通过在底层将成千上万个连续行为的输入属性与最终决策进行对齐分析，AI 便能真正解析出深藏在日常操作中的决策约束与业务逻辑。 ### [16:51](https://www.youtube.com/watch?v=2kjXV50-iiA&t=1011s) 将“工作”视为一种语言：从屏幕行为中涌现对流程的认知理解 Avinash Misra 用一个精妙的类比阐述了 Skan AI 的底层技术逻辑：OpenAI 将全球海量文本（单词、句子、段落、n-grams）输入 Transformer 架构，通过不断训练模型“预测下一个 Token（predict the next token）”，最终涌现出了对语言的认知理解与智能；而 Skan AI 则是“将工作视为一种语言（treat work as language）”。在企业实际环境中，屏幕上的 UI 交互、操作步骤、跨人员与跨系统的往复沟通，就相当于一个个“词汇”；而一个案件从最初录入（Intake）到最终裁决（Final adjudication）的完整端到端生命周期，就构成了一个“句子”。 当系统观察并收集了成千上万个这样的“句子”，再通过模型来预测下一步行为时，模型就会针对该企业、该业务流程自然涌现出对“工作语言”的深层理解。例如，模型在海量操作中会自动推断出规律与心智模型：“每次遇到加利福尼亚州（California）的数据时，流程就会走向路径 A；而一旦出现德克萨斯州（Texas），就会走向路径 B”。通过这种规模化推断，系统能够精准重构企业内部复杂且动态变化的实际执行路径。 ```mermaid flowchart LR accTitle: 将工作建模为语言的技术逻辑 subgraph LLM["传统大语言模型 (LLM)"] T1["词汇 / 字符 / N-grams"] --> T2["完整文本段落 / 篇章"] T2 --> T3["预测下一个 Token"] T3 --> T4["语言智能与理解涌现"] end subgraph Skan["Skan AI 工作语境图谱"] W1["UI 交互 / 点击 / 协作步骤"] --> W2["完整案件生命周期 (Intake 到 Adjudication)"] W2 --> W3["预测下一步工作行为"] W3 --> W4["企业工作流程认知涌现"] end ``` - **词汇与句子的工作隐喻**：将屏幕 UI 元素、操作点击与系统间流转视为“词（Words）”，将端到端的业务案例视为“句子（Sentence）”。 - **规则与模式的自动沉淀**：无需人工预先编写死板规则，模型通过观察海量执行路径自动推断出基于特定条件（如不同州、不同业务类型）的分支逻辑。 --- ### [18:02](https://www.youtube.com/watch?v=2kjXV50-iiA&t=1082s) 用户反馈、界面局限与工作流程向智能体的“流程坍缩” 现场观众提出疑问：当用户在使用智能体或应用时，实际上是在验证它能否达成目标，如果用户反馈“我不喜欢这个结果，因为某个角度完全错了”，企业该如何利用这类反馈与上下文数据来持续优化个人或企业级 Agent？Avinash Misra 指出，利用上下文与反馈来优化系统在客户部署中是非常核心的场景。当前企业流程分散在各类古老的 Mainframe 大型机、Web 系统或桌面软件中，模型在洞察员工如何在这些异构应用中穿梭后，能够直接给出“短路（Short-circuit）”优化建议——在流程起点或终点直接把所需数据呈现给用户，省去员工全凭个人记忆在多个系统间搬运数据的繁琐负担。 Avinash 进一步提出了深刻的洞察：传统的用户界面（UI）本质上是人类“上下文窗口（Context Window）”极其有限的产物与映射。人类大脑脆弱且容量有限，一次只能记住 3 到 5 件事，屏幕上最多只能处理 15 个要素而非 500 个，因此软件才被设计成当前的 UI 交互形态。随着企业从“人类执行”迈向“智能体执行（Agentic execution）”，正在发生一场彻底的“流程坍缩（Process collapse）”。这不仅是流程与技术架构的渐进式改良，而是整套人机交互界面与操作流程被直接折叠进 Agent 的自动化执行中。理解人类当前如何穿越复杂的执行图景，既是为了优化当下的人类操作，更是为了给未来的 AI Agent 提供最真实、无缝的行动蓝图。 - **人类上下文窗口的物理限制**：现有 UI 之所以复杂分散，是因为软件必须迁就人类只能同时关注少量信息的心智局限。 - **流程坍缩（Process Collapse）**：随着 AI Agent 接管任务，原本为了配合人类生理限制而设计的长链路、多系统 UI 交互将直接坍缩为智能体底层的端到端调用。 --- ### [20:28](https://www.youtube.com/watch?v=2kjXV50-iiA&t=1228s) 捕获决策的隐性知识，决定 Agent 成败的关键是“地图质量” Amit Patel 补充强调，Skan AI 之所以能在企业级生态网络中脱颖而出，核心就在于它能够捕获决策背后的细微差别与隐性历史——例如“某个看似可行的决策在 3 年前的某间会议室里因为 XYZ 原因曾被明确否决过”。大多数平台只给 Agent 提供标准场景的数据流，完全忽略了这类散落在真实组织活动中的决策变体与潜规则。 在对谈的尾声，Avinash Misra 留下了整场演讲最核心的总结观点：企业当下真正面临的挑战，“根本不是各家大模型底座在‘驾驶员智力水平’上的收敛与竞争，而在激烈的竞争中，真正起决定性作用的是驾驶员手中那张‘地图的质量’（What is at play is the quality of the map that the driver has）”。无论 AI 驾驶员有多聪明，如果没有高质量、实时反映企业真实工作现状的上下文图谱作为指引，它在实际落地中必然迷路。这对 CIO、CXO 以及所有主导企业 AI 战略落地的领导者而言，是构建长久技术壁垒的最关键法则。 - **核心金句**：“底座大模型就像智力趋同的驾驶员；最终决定成败的，是驾驶员手中所拥有的地图质量。” - **捕获组织隐性知识**：Agent 必须掌握企业过往决策中的细微偏好与历史教训，才能在严苛的企业生产环境中做出合规且精准的判断。 ## Notable quotes > [00:08](https://www.youtube.com/watch?v=2kjXV50-iiA&t=8s) 当全世界都沉迷于寻找马路上最聪明、最强悍的驾驶员时，我们却专注于为你需要行驶的区域绘制最详尽、最精准的高清地图。 — Avinash Misra > [01:36](https://www.youtube.com/watch?v=2kjXV50-iiA&t=96s) AI 的职责不是天马行空地重新发明路径，而是在经过验证的既有正确路径之间做出精准抉择。 — Avinash Misra > [03:08](https://www.youtube.com/watch?v=2kjXV50-iiA&t=188s) SOP 和规章制度并非毫无价值，但它们只是企业对工作应当如何开展的意图声明，根本不是工作实际发生的真实状态。 — Avinash Misra > [07:08](https://www.youtube.com/watch?v=2kjXV50-iiA&t=428s) Agent 在沙盒里表现亮眼，一进生产环境就彻底崩盘。 — Amit Patel > [07:08](https://www.youtube.com/watch?v=2kjXV50-iiA&t=428s) 沙盒中干净的数据与少量的变体，正是最终毒死生产落地的毒药。 — Avinash Misra > [13:44](https://www.youtube.com/watch?v=2kjXV50-iiA&t=824s) 你不仅可以租赁模型，你其实连智能本身也是租来的。企业真正需要拥有并誓死捍卫的，是你自己的竞争性上下文。 — Avinash Misra > [16:51](https://www.youtube.com/watch?v=2kjXV50-iiA&t=1011s) 将工作视为一种语言。在企业实际环境中，屏幕上的 UI 交互、操作步骤就相当于一个个词汇，而一个案件从最初录入到最终裁决的端到端生命周期，就构成了一个句子。 — Avinash Misra > [18:02](https://www.youtube.com/watch?v=2kjXV50-iiA&t=1082s) 传统的用户界面（UI）本质上是人类上下文窗口极其有限的产物与映射。 — Avinash Misra > [20:28](https://www.youtube.com/watch?v=2kjXV50-iiA&t=1228s) 根本不是各家大模型底座在驾驶员智力水平上的收敛与竞争，而在激烈的竞争中，真正起决定性作用的是驾驶员手中那张地图的质量。 — Avinash Misra ## Mentioned Avinash Misra, Amit Patel, Jaya Gupta, Satya Nadella, Alex Karp, Skan AI, Plug and Play Ventures, Ai4, Foundation Capital, OpenAI, Nike, Walmart, Mercedes-Benz, 企业工作上下文图谱 (Enterprise Context Graph of Work), 工作遥测 (Telemetry of Work), 流程坍缩 (Process Collapse), 检索增强生成 (RAG), AI Agent, 数字孪生 (Digital Twin), Transformer ## Video description Avinash Misra - CEO & Co-Founder, Skan AI Amit Patel - Partner, Plug and Play Ventures Frontier models are brilliant but blind. They know everything except your company: how work actually runs, where it breaks, and the institutional knowledge that lives in no training set. This layer, enterprise context, is the most underestimated and strategically important part of the AI stack, and the hardest for any model to replicate. Most teams try to fill it with stale documents and system logs, and agents grounded in fiction break in production. Avinash Misra, CEO of Skan AI, is ready to explore a different path: grounding AI in a living, observed model of real work. By observing how work runs across people and systems, enterprises can build an Enterprise Context Graph of Work, a continuously updated source of ground truth that closes the gap between what agents are told and what is true. The session covers why context, not model choice, is becoming the durable advantage, and how enterprises can start mapping it today. [Source](https://www.youtube.com/watch?v=2kjXV50-iiA)

[0:08] Avinash Misra: 当全世界都沉迷于寻找马路上最聪明、最强悍的驾驶员时，我们却专注于为你需要行驶的区域绘制最详尽、最精准的高清地图。 [1:36] Avinash Misra: AI 的职责不是天马行空地重新发明路径，而是在经过验证的既有正确路径之间做出精准抉择。 [3:08] Avinash Misra: SOP 和规章制度并非毫无价值，但它们只是企业对工作应当如何开展的意图声明，根本不是工作实际发生的真实状态。 [7:08] Amit Patel: Agent 在沙盒里表现亮眼，一进生产环境就彻底崩盘。 [7:08] Avinash Misra: 沙盒中干净的数据与少量的变体，正是最终毒死生产落地的毒药。 [13:44] Avinash Misra: 你不仅可以租赁模型，你其实连智能本身也是租来的。企业真正需要拥有并誓死捍卫的，是你自己的竞争性上下文。 [16:51] Avinash Misra: 将工作视为一种语言。在企业实际环境中，屏幕上的 UI 交互、操作步骤就相当于一个个词汇，而一个案件从最初录入到最终裁决的端到端生命周期，就构成了一个句子。 [18:02] Avinash Misra: 传统的用户界面（UI）本质上是人类上下文窗口极其有限的产物与映射。 [20:28] Avinash Misra: 根本不是各家大模型底座在驾驶员智力水平上的收敛与竞争，而在激烈的竞争中，真正起决定性作用的是驾驶员手中那张地图的质量。

[0:08] Avinash Misra: A foundation model is only as good as the context around it. [3:08] Avinash Misra: declarations of intent [13:36] Amit Patel: rent the model, but own the intelligence [20:28] Avinash Misra: The intelligence convergence of the driver is not at play. What is at play is the quality of the map that the driver has.

[0:08] Thank you all for being here today. I know it's  been a long day. My feet are definitely tired. [0:12] Uh lesson learned from last conference. Never  wear shoes. So I I started wearing these Toms [0:16] and flats. So feet feel a lot better. Uh but uh  great to see you all. Uh my name's Emit. Uh I'm a [0:23] partner at Plug-and-Play Ventures. We are a global  venture firm with the world's largest corporate [0:28] innovation network. So we work with firms like  Nike, Walmart, uh Mercedes uh to support their [0:33] growth, innovation and AI activities. Um and the  way in which we do that is by helping them connect [0:38] with emerging companies from our ecosystem and  network. And then we use the feedback from these [0:43] interactions to make investment decisions. So uh  we've been investing for over a decade with uh 38 [0:49] unicorns in our portfolio. Um and then a why don't  you introduce yourself as well? Yeah, we we're [0:55] lucky to have been invested by plug-and-play as  a start of our journey and many years ago. Uh my [1:01] name is Abin Mistra, CEO and co-founder of scan.  And of course, you can see a tagline of scan, [1:07] you know, a foundational model is only as good as  the context around it. I'd put it much simply. In [1:12] a world obsessed with finding the smartest and  the most intelligent driver on the road, we are [1:18] obsessed with creating the most detailed map of  the areas in which you need to drive. So it's that [1:24] difference between getting the agency and the  intelligence and the context that drives that [1:30] agency and that intelligence in the enterprise.  Yeah. I mean before we dig into the the enterprise [1:36] context or enterprise context of work, I think  we should kind of determine like what is the [1:42] difference between probabilistic and deterministic  AI and why is that distinction very important when [1:49] you're thinking about building enterprise contact  for work? Absolutely. You see in in any enterprise [1:54] take a bank or insurance company whatever  a claim needs to either be approved or not. [2:00] payments either needs to go through or not and  uh there is and this is all governed by either [2:06] policy or contract. uh we do not start from  scratch and we do not build out everything [2:13] that we are doing right from scratch right uh  and so in real life as well right when we say [2:19] that you to do a particular piece of arithmetic  you say 2 plus 2 equal to 4 you don't go back [2:24] to you know the first principles to do 2 plus 2  equal there's certain observed memory same thing [2:29] in organizations that the part of the behavior of  the organization is largely codified in policy and [2:35] contract and therefore one must not go outside  the realms of that uh policy and contract when [2:41] making decisions and that's the deterministic  part of the organization. uh are we saying that [2:47] uh uh probabilistic engines that we have in the  form of LLMs must go back to now reinventing the [2:53] whole wheel and come to a decision or what we  are saying is that ground these probabilistic [2:58] engines in the operational reality of this policy  and contract and the lived execution history of [3:04] the organization to choose between the right  paths as opposed to coming up with the right [3:08] path and so I mean in simple terms what you're  basically referring to is like understanding the [3:13] variations and the nuances effectively like how  decisions get made, right? And and I, you know, [3:19] I want to give credit where credit's due. you know  the whole notion of enterprise context or just [3:24] context graphs I think picked up probably in the  last like six months or so when foundation capital [3:29] like you know published that post or Jupa from  foundation capital but you know credit should be [3:35] given cuz you've been talking about without naming  it without before it was called contacts like [3:41] uh from like 2019 if I remember like when I first  met with you cuz your notion was and you called [3:45] it telemetry of work at that time I remember that  kind of statement that you made because your your [3:50] I remember your thesis boss like you know you know  for enterprises to truly understand and and I and [3:56] I see this in in my line of work you know we work  with several chief transformation officers CIO [4:02] COOs everyone wants to redesign workflows in the  age of AI right everyone wants to rethink their [4:07] processes but how do you do that without having a  real like a real understanding of your operating [4:12] model and I remember you said like you know there  by the way even in this conference there are tons [4:16] of process discovery tools I I mean I've seen it  um uh but where I where you stood out for me and [4:22] and we saw this across our ecosystem is that you  were able to capture that variation nuances that [4:27] I think a lot of the platform and capabilities  have missed and I can say this with with kind of [4:32] fact because we see thousands we see hundreds of  thousands of companies in our ecosystem actually [4:37] the reason why we even invested in scan at the  time was because you know you pick any function [4:42] today any department there are dozens of AI tools  tools and platforms it's how do you separate the [4:47] signal from the noise I mean you work with the  top five banks, the top five insurance companies, [4:52] like you name it, you work with and I you know  you so you were able to you you you thought about [4:57] like the importance of capturing that kind of  the decision making the variations very early [5:02] on and one of the things I really liked what  you said earlier it's like you know you're not [5:06] going to get these kind of visibility into the  work into your operating metal through through [5:10] documentations or SOPs which what a lot of you  know CEOs might even point to you believe that [5:16] it comes you can only build that context through  observed model of work. So how does that work? [5:22] Why why is that important? I think first of all u  it's not true that uh we have to completely give [5:28] away what organization what the organization knows  in terms of its standard operating procedures or [5:33] its policies that it's written down and so on so  forth. The fact is that that those artifacts that [5:40] the organization has are declarations of intent of  how work should be done. They are not how work is [5:46] done. and in the lived history of the organization  in the kinetic work that it does that is what [5:51] hides all the nuances exceptions permutations  and everything that makes work real. So in the [5:57] ideal world and which is what we are now forced  to reckon with as we seed work to AI that AI must [6:03] be aware of this lived history and therefore it  is not enough that you give a rag based uh you [6:09] know pull into an AI model from the documentation  you also must ex must expose this AI model to the [6:16] execution plane to the human decisions that are  being made in the kinetic form as work happens in [6:21] this organization. So our technology actually  observes work on the screens of hundreds of [6:26] people simultaneously. It featurizes the screen  and understands everything that's happening on [6:31] the screen not just here at this screen but  hundreds of other screens simultaneously and [6:35] pulls that into the model. To your point AIT when  you remembered and you we don't say that anymore [6:40] because of course you know I think context has  become much more pregnant. That's the catchphrase. [6:44] But yes, this telemetry, this uh uh kinetic work  that is being done, this hides the competitive [6:50] advantage for an organization, especially in an  age when uh AI is making all organizations equal, [6:57] right? And so where do where does the competitive  advantage shift? Yeah. Shifts to the lived history [7:02] and the thing that distinguishes you allows you  to compete, which is in your execution. And I I'm [7:08] actually going to go kind of you know something  that a question that you know I wasn't planning [7:12] to ask you but I just thought about it and please  humor me. So yesterday I was talking to one of the [7:17] head of AI of a fortune for Fortune50 and he was  actually uh talking about like you know like he's [7:23] right now as he's thinking about his AI processes  and workflows. First thing that they're basically [7:28] doing is like we need to figure out map out our  or basically he's his words I'm trying to figure [7:33] out my digital optin of my operating model and  that's what I'm focused on right now and if you've [7:37] looked at if if I've seen any success from any of  across our enterprise ecosystem the first thing [7:42] foremost thing is is having that data but he said  there's two ways of looking at it you've got the [7:46] institutional memory right you've got the digital  operating model of your twin and that's your in at [7:52] the institutional level but what about at the  individual level right the decision decision [7:56] making that goes of the individual what it's in  their memory right how they make nuances and how [8:02] do you capture that what like that becomes even  more important than just capturing like you know [8:07] at the aggregate level because if you're talking  if you want autonomous agents you need individual [8:13] level data you [clears throat] you don't just  the institutional level is not going to be enough [8:17] and as I said I think the institutional level is  more from an intent perspective uh whereas uh in [8:23] the execution form and the expressed work the  instantiation of that intent is at the ground [8:30] level which is the millions of things that people  do in the organization on a day-to-day basis. [8:34] Now of course this if you try and capture all  of those signals that is extremely noisy right [8:39] you know think about your own work today even if  you were engaged in a task which involved let's [8:44] say n steps in between those end steps you would  deviate so many times so many distractions so many [8:49] other things that you would do so work in itself  is extremely noisy it's part of our technology to [8:54] a be able to observe distill context out of that  ground that ground models in that context collapse [9:01] the process that was while running with human  intervention now runs with an AI agent. Therefore, [9:06] that has larger context. Collapse that process  and then give this to a execution layer to [9:11] execute. I think that's the that's the golden way  in which some of our smartest customers are going [9:17] about bringing AI into into the enterprise  or operational and I think that distinction [9:22] is important because while everyone is talking  about context graphs and enterprise context like [9:27] they're all talking about the institutional level  and I think we need to get to at the individual [9:32] decision making level and I think that that is  extremely critical like I mean if you don't want [9:36] your agents to hallucinate if you want that full  autonomy I mean that is the granular level of data [9:41] that you need to get get at. But let's let's shift  gears a little bit because we we see this struggle [9:45] within our enterprise ecosystem a lot like we the  amount of times I've seen CEOs and CTOs like hey [9:50] you know agents I mean I mean it's all agents  everywhere. I mean every company's now building [9:56] agent something but the reality is that um what  I've seen from my ecosystem is while they work [10:02] great in a sandbox environment they tend to fall  apart when it gets into kind of the production [10:07] environment. I mean what you know actually when  I was talking to I was talking with the chief AI [10:14] officer of of a top five bank just a couple days  ago and he was telling me you know like when we're [10:18] bringing products and solutions in um it needs to  be truly be scalable like I can't even think about [10:23] this like the the the the scope of my mandate is  institutional level and so if if a company cannot [10:29] scale across multiple workflows that has millions  of variations right millions and that's why it [10:34] fails to in production like how do you how do you  account for that you've working with you work with [10:39] I mean from my I mean I've I've seen great success  stories from you working with some tier one brands [10:44] probably names you can't name on stage but how  have you kind of crossed that from from sandbox [10:50] to production I think uh first of all I think  there is this realization that uh in in sandbox [10:57] you will have a very clean process you will have  lesser variations so on so forth and so that [11:03] becomes the poison so whereas whereas you might  be able to ex you may be able to show something in [11:08] pilot very easily and you might get the starting  point of what needs to happen in production but [11:13] that itself becomes the poison to be getting into  production because it assumes so little variation [11:19] and so little scale right and so the trick lies in  as I said earlier trick lies in organizations that [11:25] understand that yes as a technology this works but  to bring it to production you we must enrich this [11:30] model with much more detailed execution pathways  and observed realities of that particular process [11:38] And if you feed that then of course the model  becomes much more sentient in situations that [11:42] it has it has seen historically the organization  handle and handle the central ways and if it [11:48] cannot then of course it leads to a human. Yeah.  I mean I one thing I I I will also add is like [11:55] some of the things that we've we've we've kind of  engaged with within the kind of the enterprise AI [11:59] leaders we've told them like I mean besides what  you said you need to design for production from [12:05] day one. You cannot be working with companies and  platforms and and looking at the scenario like [12:11] just what let's see if this works in a sandbox  environment because it will it will work as a [12:16] it will work. Exactly. It it's it it will work in  a sandbox environment. I mean all the things like [12:20] you may have heard about like token maxing right  tokconomics like all you don't you don't see that [12:25] in in in in in a sandbox environment like for  you to like truly understand the scale of what [12:30] a platform can do that design of what you're  thinking of when you work with the company that [12:35] production level design needs to happen on day  one and in fact I would say companies who do this [12:40] really well in our ecosystem are companies  like Walmart and a few others uh but they [12:44] that's what they've done they've they've they've  they've figured out a process where like they're [12:48] designing everything as if they're going to scale  it enterprisewide. Uh in fact, you know, we've I [12:53] few of the companies in our ecosystem, the amount  of rigorous diligence that they're going through [12:57] and and it's for that is to ensure success when it  gets into that production. I I'll shift gears. Uh [13:04] your your token point reminded me of even on the  question that you raised on probabilistic versus [13:09] deterministic that has implications on energy  and that has implication on token economics as [13:15] well. I mean if you constrain the output of an LLM  to choose between certain deterministic outcomes [13:21] you are consuming much less tokens and much less  energy than if you were to say okay come up with [13:25] with an outcome right because I mean if you if  you if you have that full understanding like the [13:30] caching and whatnot like if you if you prepare for  that from day zero I want to say correct you avoid [13:36] all the the blind spots uh uh effectively um um  you know okay let me ask you this u um you know in [13:44] the last I would say month or so you know there's  big notion coming around like you know rent the [13:49] model but own the intelligence yes right you've  heard you've probably read the kind of the post by [13:54] setella Alex carp like the whole idea I mean the  the I mean it's well at this point is like models [14:01] are a commodity right models are a commodity uh  in fact but but how do you make sure that the the [14:09] underlying context that you build around it or  the operating that you build around it keeps you [14:14] gives you that edge because remember two I you  mentioned very earlier on everyone gets access [14:19] to the same thing so like where do you then now  build your differentiation and and your advantage [14:23] first of all I take um I would like to correct uh  something there it's not that um you know rent the [14:28] model own the intelligence you can rent the model  you can rent the intelligence also what you need [14:33] to own is your own competitive context uh and I  think the future is all about renting intelligence [14:40] uh but what will differentiate what organization  will compete on is is the way they organize that [14:45] intelligence based on their own competitive  stance, right? 100%. Internally, how their [14:50] work is organized in much the same way. You know,  the fact of the matter is that even 20 years ago, [14:54] we had access to large diag models just that they  came in 20 watt size inside the cranium and you [15:01] brought them in and then you gave them a process,  you trained them, you told them so much nuance [15:06] about your business, then they became productive.  Now if you scale that to a much larger model that [15:10] does not sit inside the crania, you still have to  do the same things again. So the same onboarding [15:15] has to happen again. Yeah. I mean the one answer  I always get from our partners is oh but it's in [15:19] our own environment. The the the model's not  going to get access to the data. But I think [15:23] the enterprises don't recognize that the model  frontier labs are not just capturing the data. [15:27] It's not just about the data. It's every decision  every exceptions. This was a right answer. This [15:32] was a wrong answer. that that is insights in  itself that is being captured by the frontier labs [15:37] and I'm willing to bet that that's going to become  more part of the the procurement conversation [15:43] of like who's going to own that like level of  intelligence and the in that context I have no [15:48] doubt in my mind I think that has to be owned by  the enterprise responsibility of vendors like us [15:52] and platforms like us to allow our customers to  own the context uh in our case we just don't we [15:58] just don't allow them to own the context we help  them to build it and then own it and link it to [16:04] uh execution layers. Yeah. Um I want to open it  up for questions being mindful of time. Um does [16:09] anyone have any questions or I can keep riffing  with him but uh yeah I'll repeat the question. How [16:14] do you capture the reasoning and the constraints  behind the decision-m? Yeah. So the it's done by [16:20] a a process of uh first of all exposing the  model to all execution that is happening over [16:25] a period of time. Number two is featurizing  the screen. Every time the screen is screen, [16:29] the screen is featurized with respect to whatever  attributes may be on the screen. For example, [16:34] it may be a customer that is being onboarded. Say  what's the state, what's the address, you know, [16:38] what's the income or a claim it may have. For  example, what's the kind of loss and what's [16:43] the state in which the claim is being filed.  All of this is being featurized. And just like [16:47] you were to look at 10 different examples of 10  different things happening, then you would make [16:51] a sense that well, every time I see the state of  California, it goes around this path. Every time [16:56] I say the state of Texas, it goes around certain  other paths. So it build a mental model of how [17:00] this is happening. Now at scale, the the model is  then able to infer different paths. If I were to [17:06] give you an analogy of the kind of work that we do  and the the underlying thesis of our technology, [17:10] you see what did OpenAI do? OpenAI took the  totality of the world's text uh you know words [17:15] sentences paragraphs engrams and exposed it to a  trans to the transformer architecture and sentient [17:21] by trying to predict uh project the next token or  predict the next token you sentient understanding [17:27] emerged. We treat work as language. In our world,  work on the screen, the various aspects of the [17:34] screen, the the user interface, the steps that  were taken, the back and forth between people, [17:39] this would constitute the words of a sentence and  the sentence may be construed as a case that went [17:45] from uh intake to final adjudication. Now if you  have many of these sentences and you constrain the [17:52] model and ask it to predict the next token, you  also have sentient understanding emerge in this [17:56] model about what this language of work is for  that process for that organization. Hope that [18:02] answers your question. I'm thinking every time  when the user was using an agent or application [18:08] uh it's actually testing right the the user was  testing if that reach his or her goals. I was [18:14] wondering how can how have you think about how you  use that in kind of information like user feedback [18:20] says I don't like this result because it's it's  totally wrong in one angle or something. How can [18:24] we use that information to improve the context  like you mentioned like a personal or something [18:29] or like in general enterprise level of the context  to make the application do better next time. Have [18:35] you thought about that? Have any things about  that? Thank you. So if I understand correctly, [18:40] you asked um how do you use the user feedback  um uh to improve like the the system. [18:47] Uh in fact um that is a very dominant use case  uh in our customer deployments uh which is that [18:54] um of the many forms of technologies that is  being used in a given business process various [19:01] applications and they may be mainframe systems  or web-based systems or desktops and as people [19:06] navigate through various parts of the application  we can actually give suggestions and their model [19:12] is able to give suggestions as to what parts of  the applications can be shortcircuited because [19:17] this data can be made available more readily  to a user right at the start or right at the [19:21] end instead of the user trying to use their  own memory and their own context in moving [19:25] things around. So that is something extremely  natural and in fact I would I would argue that [19:30] the user interface was a reflection of the size  of the human context human context window. So the [19:38] the fact that we built user interfaces that the  way we have is reflects our own frail as we can [19:44] remember five things or three things or we can  see 15 different things on a screen not 500 and [19:49] that reflects us and as this process moves from  and what I call as process collapse happens from [19:56] human execution to agentic execution all of these  things come into play even more not just from [20:00] process improvement and technology landscape  improvement but fundamentally a collapse of [20:05] that technology and the collapse of that process  into agentic execution. So this is a fundamental [20:10] input into how the execution landscape is being  traversed today by the human a to optimize what [20:17] the human does if the human were to to to continue  to operate it but more importantly how does the AI [20:23] agent now work on the same landscape. Yeah. And  I think the most important distinction is is is [20:29] to your point is is capturing that unique  like an like you know every agent needs to [20:35] understand like you can give it a feed of like  different scenarios but you know they need to [20:39] understand that 3 years ago this exact decision  was rejected for XYZ reason in a conference room. [20:45] uh it's it's capturing that information I think  and I think this is where your I think I why I [20:51] probably think that your platform has has done  really well in our ecosystem across enterprise [20:55] network is that this importance of capturing  that variation in decision-m and the nuances [21:00] in in in decision-m that a lot of platforms in my  opinion miss right now um all right I think we're [21:06] over time but any closing thought uh before we  get off there is there is only one thought it's [21:10] not closing thought it is the only thought which  is that that the the intellig convergence of the [21:15] driver is not at play. What is at play is the  quality of the map that the driver has. Yeah, [21:20] I think that if you remember that then uh you  will you as CIOS, CXOs and you know owners of [21:27] the AI industries of your organizations will  do well. Agreed. And thank you. Thank you team.