Skan AI Targets Corporate AI Waste With Digital Twin Action Models

Skan AI以数字孪生行动模型直击企业AI浪费 企业AI投资未能产生底线回报，因为企业资助的是通用任务机器人，而非去捕捉 发言者探讨了企业AI中普遍存在的“概念验证迷雾”，指出当前支出不成比例地浪费在对损益影响微乎其微的个人和平台机器人上。要获得真正的ROI，必须转向跨系统的“流程智能体”，瞄准占据绝大部分人力成本的核心运营价值链。通过持续的屏幕观察来构建人工工作流的数字孪生，Skan AI训练出大型行动模型（LAM），能够在人工监督下自主执行诸如Guidewire理赔裁决等复杂的跨系统工作流。 多数企业 AI 项目仍深陷“概念验证迷雾”（POC Fog），主因是企业只部署了通用的任务机器人（Task Bots），而未能针对端到端运营模型进行重构与跨职能协同。 (0:04) 企业大部分 AI 预算虽投入在低财务回报的个人与平台级 Agent，但能贯穿多系统的端到端流程 Agent（Process Agents）才能真正触达占支出 60% 至 70% 的人力成本，带来最高 ROI 与竞争壁垒。 (4:01) 通用 LLM 无法处理复杂的企业专有流程，必须依托大动作模型（LAMs）将人类操作员的隐性知识、不确定性下的推理及完整上下文转化为结构化动作。 (8:56) Skan AI 借鉴自动驾驶的遥测训练理念，通过连续捕获人类屏幕交互生成流程数字孪生（Digital Twin），并以此训练 LAM 实现跨 Guidewire 等核心系统的保险理赔全流程自动化，使单一操作员可监管 10 至 15 个 Process Agent。 (13:06)

Enterprises struggle to achieve agentic AI returns because investments remain concentrated on personal and platform bots rather than cross-system process agents, according to Dayton Semerjian and Vinay Mummigatti. They argue that realizing measurable bottom-line value requires capturing granular operational telemetry to train Large Action Models capable of executing end-to-end enterprise workflows. The Disconnect Between Agent Spending and Enterprise ROI (4:01) Enterprises face an agentic paradox: capital flows predominantly into personal productivity assistants (such as email drafters) and vendor-contained platform agents (such as Salesforce or ServiceNow), both of which yield negligible impact on corporate profit and loss statements. Mummigatti notes that while task-oriented bots may help workers complete shifts early, they do not register on enterprise balance sheets. Realizing structural cost reduction requires targeting core value chains—including inquiry-to-onboarding, transaction-to-fulfillment, and issue resolution—where 60% to 70% of enterprise operating budgets are tied to labor compensation. Semerjian reports that business cases routinely stall because executives underestimate the difficulty of transition or lack multi-year savings roadmaps. Furthermore, technology pilots fail when treated strictly as software upgrades rather than operational reorganizations. In one operational audit cited by Semerjian, Skan AI's telemetry revealed that an ostensibly overworked team spent 20% to 30% of their working hours on non-productive websites, illustrating that structural interventions require baseline process visibility before automation can succeed. Encoding Tacit Human Agency via Operational Telemetry (8:56) Standard commercial large language models offer limited differentiation because they operate like rigid menus unable to navigate complex, proprietary business rules. In contrast, enterprise process agents must mechanize human agency, which relies on three distinct faculties: abstract reasoning under incomplete information, tacit operational knowledge that cannot be transcribed into standard operating procedures, and accountability for business outcomes. To bridge this gap, agents require deep contextual training data rather than basic text prompts. Analogous to training autonomous vehicles on millions of hours of driving sensor data instead of hard-coded driving logic, process agents require continuous operational telemetry. This data foundation must capture discrete application actions, document interactions, process paths, environmental variables, and target systems across core transaction engines. Translating Digital Twin Observations into Large Action Models (13:06) Skan AI’s architecture converts screen-level operator activity into a digital twin of enterprise workflows, which trains a specialized Large Action Model (LAM) using action embeddings. Rather than following brittle scripted sequences, the LAM is engineered around three parameters: fulfilling defined business outcomes (such as claims adjudication), interpreting end-to-end historical context (including customer history and policy endorsements), and balancing multi-objective trade-offs between turnaround time, customer satisfaction scores, and regulatory compliance. In a demonstration of insurance claims adjudication, the model autonomously accesses Guidewire ClaimCenter, navigates the claims queue, retrieves documentation, evaluates coverage endorsements, and decides whether to pay, pend, or deny the claim. Real-time logging tracks the model's underlying reasoning step by step alongside cross-system transaction execution. Operating Model Restructuring and Human-in-the-Loop Supervision (18:03) Achieving returns from process agents requires active organizational redesign, combining cross-functional governance, Six Sigma process management, and workforce restructuring to redeploy or eliminate full-time equivalents. In production deployments, human adjusters shift into supervisory roles, with a single operator overseeing 10 to 15 autonomous process agents concurrently. When edge cases or unfamiliar anomalies emerge, the agent routes exceptions to human specialists. The resolution patterns chosen by human operators are ingested back into the system, allowing the model to adapt and handle identical scenarios autonomously in future cycles. Semerjian cautions that deploying agents over broken or unmapped workflows guarantees stalled implementations, making complete operational visibility a prerequisite for deployment. While Mummigatti acknowledges that scaling fully autonomous process agents across complex enterprises remains an evolving journey, both speakers emphasize that organizations cannot bypass baseline process discovery; attempting to deploy agentic technology without structured operational telemetry and operating model restructuring inevitably stalls measurable return on investment.

Skan AI 与前 LPL Financial 高管 Dayton Semerjian 指出，企业在智能体 AI（Agentic AI）上的投资大多陷入概念验证与个人生产力工具的误区，难以产生实际财务回报；真正的投资回报率在于跨系统的端到端流程智能体，而这需要通过数字孪生捕获人类操作数据并构建大动作模型（LAM）。 概念验证迷局与资本错配 (0:04) 企业在推进智能体 AI 时面临四大摩擦点：概念验证难以落地为生产力、缺乏企业级训练数据、单平台供应商锁定，以及将 AI 误当成纯技术而非运营模式重组。Dayton Semerjian 指出，企业商业论证受挫往往是因为管理者低估了价值创造路径，或缺少为期多年的成本缩减规划；要真正实现回报，企业必须在技术部署后对工作架构进行重构，合理调整全职人力配置，并组建涵盖流程所有者、六西格玛专家与变革管理专家的跨职能团队。 当前市场上的智能体呈现为从个人智能体（处理邮件、文档）、平台智能体（如 Salesforce、ServiceNow 生态内的内嵌工具）到流程智能体（跨多系统核心业务流程）的光谱分布。Vinay Mummigatti 强调，目前绝大多数企业资金集中在对损益表几无影响的个人与平台工具上，能带来高回报的端到端流程智能体反而资金不足。例如，一家保险公司的首席 AI 官曾戏称，其内部自动化工具虽然让员工提早下班陪伴家人与宠物，但未能为公司创造任何可衡量的财务利润。 聚焦人力核心链条与结果自动化 (6:19) 由于企业 60% 至 70% 的运营预算集中在人力薪酬上，削减结构性成本必须切入人类劳动最密集的价值链，包括客户咨询到入职流程、交易到履约处理，以及问题排查与解决流程。要实现以结果为导向的自动化，流程智能体必须整合五大支柱：业务目标、流程路径、企业上下文、工具与动作执行，以及持续优化。 Dayton Semerjian 强调企业绝不能为了采用技术而购买技术，自动化必须解决实质的人力效率盲区。他分享案例称，某运营团队此前声称工作完全超载已达上限，但在部署 Skan AI 的流程智能平台后，系统遥测数据显示该团队约 20% 至 30% 的工作时间实际上花费在浏览社交媒体等非生产性网站上，这证明了流程可视化对人员管理与成本管控的重要性。 大动作模型与类自动驾驶的数据构建 (8:56) 通用大语言模型（LLM）如同千篇一律的快餐，无法为企业提供竞争壁垒，而真正的流程智能体则需要量身定制的大动作模型（Large Action Models, LAM）。人类在工作时具备在信息不完整下抽象推理、依靠隐性经验肌肉记忆以及承担决策后果这三项关键特质；要模拟人类的感知、推理、行动与学习循环，流程智能体需要在流程路径、具体任务、操作文档、界面动作、时间戳等多层次操作数据上进行训练，而此类上下文数据的匮乏构成了当前的主要瓶颈。 Skan AI 借鉴自动驾驶依靠真实路况采集训练的思路，提出从观察到智能体（Observation to Agent）的架构。系统通过在后台非侵入式观察员工屏幕与操作流程，将原始界面交互转化为数字孪生模型与流程遥测数据，随后输入基于 Transformer 架构的大动作模型训练动作嵌入向量，以此合成具备跨系统执行能力的流程智能体。 理赔实测验证与人机协作规模化 (16:09) 在针对保险理赔流程的演示中，Skan 展示了其 LAM 驱动的流程智能体运行机制：系统能自主登录 Guidewire ClaimCenter，在索赔队列中调取卷宗，核对保单批注与索赔人历史记录，在权衡客户满意度（CSAT）、周期时间以及法定保险监管规则的前提下，自主决定赔付、挂起或拒赔，并直接在底层系统完成交易。系统在做出每一步决策时均向操作者透明展示其思维过程与评估的数据源。 在企业生产实践中，该架构并非完全取代人工，而是转向监督式协作模式：目前一名人类核赔员可同时监督 10 至 15 个自主运行的流程智能体。当智能体遭遇未知的复杂异常场景时，会将案件转交人类专家；在专家解决问题的同时，系统会自动吸收处理逻辑，动态学习应对边缘案例的方法，从而逐步扩大自主处理的覆盖边界。 Dayton Semerjian 总结指出，若企业跳过数字孪生与基础流程梳理，直接将智能体部署在混乱未理顺的工作流上，项目必将停滞；依靠观察与数据驱动的全栈方案是落实投资回报的关键路径，而复杂流程从半自主演进到完全自主的规模化能力仍处于持续演进与验证之中。

## Background 本内容由 Ai4 于 2025 年 8 月 24 日在 YouTube 平台发布，由 Skan AI 策略与客户转型执行副总裁 Vinay Mummigatti 与前 LPL Financial 首席客户关怀官 Dayton Semerjian 共同呈现。视频简介指出，企业数百万美元的 AI 投资若无法变现便毫无价值，内容重点探讨了传统任务型 AI Agent 遭遇的瓶颈，以及企业如何通过 Skan AI 将顶尖员工的业务流程知识转化为上下文驱动的大型动作模型（LAM），以实现实质性的财务回报。 ## TL;DR 演讲深入剖析了企业落地 Agentic AI 时普遍陷入概念验证（POC）迷雾、缺乏私有流程数据与锁定供应商等核心困境，并指出个人级和平台级 Agent 无法带来显著损益影响，唯有跨系统的“流程智能体（Process Agents）”才能实现最高 ROI。Skan AI 借鉴特斯拉自动驾驶捕获人类驾驶行为的逻辑，通过“从观察到智能体（Observation to Agent）”构建数字孪生，捕捉员工日常操作与内隐知识，并利用大型动作模型（LAM）生成目标导向的流程智能体。现场以保险理赔裁决为例，展示了 1 名人类主管协同监控 10 至 15 个智能体的实战场景，强调唯有重组运营模型与实现结果自动化，企业才能真正兑现 AI 投资价值。 ## Atmosphere & style 整场演讲充满务实干练的企业级实战氛围与敏锐的商业洞察，两位讲者配合默契，将前沿的 AI 架构理论与残酷的管理实战经验（school of hard knocks）无缝融合。Vinay Mummigatti 善用生动的类比（如快餐菜单与私人主厨、特斯拉自动驾驶数据采集）将复杂的大型动作模型娓娓道来，而 Dayton Semerjian 则以极具说服力的客户视角不断拉回财务损益与组织设计的核心现实。观众不仅能感受到技术从单纯流程挖掘演进到自主智能体的震撼，更能体会到剥除 AI 炒作泡沫后回归硬核商业价值的冷静与从容。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=SryWySKqrQU&t=4s) 企业在落地 Agentic AI 时普遍面临四大瓶颈：困于 POC 概念验证迷雾、缺乏内部流程级训练数据、押注单一供应商导致锁定，以及仅将其视为纯技术工具而忽视了运营模式变革。 - [00:04](https://www.youtube.com/watch?v=SryWySKqrQU&t=4s) POC 难以落地的根源在于缺乏长远商业论证，企业必须配合组织架构重组（如人力 FTE 的移出或重新配置）才能将技术运作转化为企业损益表（P&L）上的硬性财务回报。 - [04:01](https://www.youtube.com/watch?v=SryWySKqrQU&t=241s) 智能体市场可划分为个人级、平台级与流程级三大类；当前大量投资聚焦于个人智能体但商业回报极低，跨越系统与复杂决策的流程智能体（Process Agents）才是最高 ROI 区域。 - [06:19](https://www.youtube.com/watch?v=SryWySKqrQU&t=379s) 企业需从“任务自动化”彻底走向“结果自动化”，将重点聚焦于占运营预算 60% 至 70% 的核心高耗能链路（如签约入职、交易履约、售后解决）。 - [09:23](https://www.youtube.com/watch?v=SryWySKqrQU&t=563s) 通用的商品化大模型无法构成竞争壁垒，企业需借助大型动作模型（LAM）将专有数据、业务动作与合规规则深度嵌入，打造如“私人主厨”般量身定制的流程智能体。 - [10:40](https://www.youtube.com/watch?v=SryWySKqrQU&t=640s) 构建流程智能体的核心挑战在于复刻人类能动性（不完备信息下的逻辑推理、无法言传的肌肉记忆与内隐经验、明确的决策问责），这需要依赖高价值的全维度业务遥测元数据。 - [13:06](https://www.youtube.com/watch?v=SryWySKqrQU&t=786s) Skan AI 提出类似特斯拉自动驾驶训练的“从观察到智能体”（Observation to Agent）理念，利用流程挖掘与数字孪生技术记录员工真实操作轨迹，转化为训练 LAM 的核心输入。 - [15:50](https://www.youtube.com/watch?v=SryWySKqrQU&t=950s) 在保险理赔裁决等实战场景中，流程智能体能在客户满意度、处理时效与合规治理三大约束下自主决策，并支持 1 名人类主管监管 10 至 15 个智能体的人机协同与持续在线学习模式。 - [19:13](https://www.youtube.com/watch?v=SryWySKqrQU&t=1153s) 企业向 Agentic 架构演进必须以深入理解真实业务流程为基础，技术架构中若缺乏完整的流程感知与上下文分层支持，落地极易陷入停滞。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=SryWySKqrQU&t=4s) 企业 Agentic AI 的现状与落地困境 在当前的科技趋势下，Agentic AI（智能体人工智能）引发了极高的行业关注度。然而，大部分企业在背负着利用智能体推动业务变革并展现实际商业价值的双重压力下，正身处一个关键的拐点——这个拐点正在对人类生产力、企业运营模式以及整个行业生态造成巨大冲击。目前，企业在迈向 Agentic AI 的过程中普遍面临四大核心瓶颈： 1. **困于 POC（概念验证）迷雾**：绝大多数企业都在进行大量的智能体实验，但极少数能真正走向生产环境并实现投资回报（ROI）。 2. **缺乏流程级训练数据**：市面上有数以百计的供应商提供千篇一律的标准智能体，但一旦深入到流程中心型（Process-centric）智能体，企业严重缺乏用于训练这些模型的内部专属数据。 3. **供应商锁定（Vendor Lock-in）限制想象力**：面对紧迫的落地压力，许多企业急于押注单一生态供应商，将资金全部倾注其中，从而错失了快速演进的技术可能性。 4. **定位误区（技术视角而非运营模型）**：许多组织仅仅将智能体视为纯技术工具，而忽视了它实质上是一场运营模式的变革，关键在于构建人与智能体如何协作的目标运营模型（Target Operating Model）。 来自实际企业管理一线的视角补充道，从“残酷的实战经验”（school of hard knocks）来看，POC 难以落地的根源在于缺乏明确的商业论证（Business Case）和对价值创造的长期（跨年度）节省预估。此外，技术本身运作正常并不等于能直接变现，企业必须进行组织架构重组（Organizational Design），将全职人力（FTE）移出或重新部署到其他岗位，才能兑现硬性财务回报。同时，清晰的数据指标、严密的数据治理以及跨职能团队协作（包括业务流程负责人、数据分析师、六西格玛专员及变革管理人员共同构成的完整战队）是让项目走出试点、迈入生产部署的基石。 --- ### [04:01](https://www.youtube.com/watch?v=SryWySKqrQU&t=241s) 智能体市场的三层分类与 ROI 错位 在快速膨胀的智能体供应商生态中，企业级智能体大致可以划分为三大类别，它们在适用范围和价值产出上有着鲜明的对比： ```mermaid flowchart LR accTitle: 智能体演进与价值层级 subgraph S1["个人级智能体 (Personal Agents)"] A["邮件/文档等常规任务自动化"] end subgraph S2["平台级智能体 (Platform Agents)"] B["绑定特定 SaaS 数据与流程<br/>(Salesforce / ServiceNow 等)"] end subgraph S3["流程级智能体 (Process Agents)"] C["跨系统/复杂决策/端到端流程<br/>(最高 ROI 区域)"] end A --> B B --> C ``` - **个人智能体（Personal Agents）**：专注于提升个人生产力，处理日常的邮件自动化、文档自动化和例行任务。目前大量企业投资集中于此，但整体商业回报极低。 - **平台智能体（Platform Agents）**：由 Salesforce、ServiceNow、NICE 等厂商推出，依赖其专属数据模型和特定流程运行。虽然满足了特定场景，但被严格局限在单一平台的生态围墙内。 - **流程智能体（Process Agents）**：深入企业最核心的端到端业务交易与客户旅程，跨越多种记录系统（Systems of Record）、主数据源、复杂决策及各类业务异常。这是当前企业最感吃力、投入最少、但潜在 ROI 最高的领域。 针对个人与流程智能体的 ROI 差距，一位保险公司的首席 AI 官（CAIO）曾幽默地指出当前落地的尴尬现实： > “我热爱智能体，因为我们开发的所有智能体，最大的受益者是我们的业务操作员——他们可以早点下班回家，花更多时间陪伴宠物和家人。但在企业损益表（P&L）层面，却没有任何实质影响。” --- ### [06:19](https://www.youtube.com/watch?v=SryWySKqrQU&t=379s) 从任务自动化到结果自动化的范式转移 要撬动企业实质性的财务回报，核心在于改变成本结构。在绝大多数企业中，约 60% 至 70% 的运营预算被锁定在人力薪酬成本上。企业价值链中最耗费人力的三大关键环节包括： 1. **咨询到入职/签约（Inquiry to Onboarding）**：客户旅程的前期阶段； 2. **交易到履约（Transaction to Fulfillment）**：业务流转与交付的核心链路； 3. **问题到解决（Trouble to Resolve / Issue to Resolution）**：客户服务与售后支持。 要让流程智能体在这些关键业务链中产生价值，必须整合五大核心要素： - **明确的目标（Goal）**：清晰界定业务成果； - **严格的流程（Process）**：按时序执行业务活动； - **丰富的上下文（Context）**：融合企业数据、主数据源及客户历史； - **多元的工具与行动（Tools & Actions）**：支撑执行操作的能力； - **持续的优化（Optimization）**：对业务流程进行动态调整与改进。 这种转变要求企业从“任务自动化”彻底走向“结果自动化”。企业不会为了技术而买单，一切投资必须以移出系统成本、交付确定成果为目的。在过往的实际落地案例中，流程分析工具甚至揭示出运营团队声称“人手不足、超负荷运转”的表象下，实际上有 20% 到 30% 的工时被消耗在社交媒体等非生产性活动中，这凸显了流程透明度对管理变革与商业论证的决定性意义。 随后，发言人以快餐连锁店（如汉堡、披萨店）的“预设菜单”打比方，指出个人智能体和平台智能体就如同只能提供通用选择的快餐，而真正的流程智能体则需要完全不同的架构来应对高度定制化与复杂的企业流程。 ### [09:23](https://www.youtube.com/watch?v=SryWySKqrQU&t=563s) 通用大模型与流程智能体：企业专属的“私人主厨” 商品化的通用大型语言模型（Commodity LLMs）缺乏差异化能力，因为它们并没有针对企业的专有数据进行深度训练。Vinay Mummigatti 用了一个生动的比喻来阐明这一点：这就像为你聘请一位专属的“私人主厨”——这位厨师在你的厨房里工作，使用你的烹饪器具，完全按照你的营养偏好和定制菜单提供服务，精准交付你所期望的一切。这就是流程智能体（Process Agents）的设计初衷。它们依托企业的数据模型、流程规范与合规决策规则，深度驱动业务流程运转，而这也正是企业构筑核心竞争优势的关键所在。 市面上充斥着成百上千家企业都能轻易接入的标准 Agent 工具，如果所有人都使用相同的商品化模型，长远来看这只是一场“零和博弈”（net neutral game），根本无法形成竞争壁垒。相反，企业唯有依靠流程智能体才能持续放大独特优势。这也催生了大型动作模型（Large Action Models, LAM）的兴起——作为语言模型的下一代演进形态，LAM 能够将企业实际运行中的动作与业务流程深度嵌入模型之中。 - **通用 LLM vs 流程智能体**：通用模型属于缺乏差异化竞争力的商品化技术；流程智能体则像“私人主厨”，深度绑定企业私有流程、数据模型与合规规则。 - **差异化竞争价值**：直接调用公开市场 Agent 无法建立竞争壁垒，基于自身流程构建的 Agent 才能实现长期盈利与护城河。 - **大型动作模型（LAM）**：语言模型的关键演进方向，将企业的实际操作动作与流程逻辑直接转化为可执行的行动嵌入（Action Embeddings）。 --- ### [10:40](https://www.youtube.com/watch?v=SryWySKqrQU&t=640s) 人类能动性与流程智能体的构建难题 构建流程智能体之所以极其困难，根源在于“Agentic”（智能体化）这个概念本身。智能体化并非单纯的技术术语，而是将技术赋予“人类能动性”（Human Agency）。Vinay 指出，定义人类能动性主要包含三个核心维度：首先是在信息不完备的情况下进行抽象与逻辑推理的能力（例如呼叫中心客服人员即便手头信息有限，依然能实时做出业务判断）；其次是内隐知识（Tacit Dimension）或所谓的“肌肉记忆”，即人类掌握大量隐性经验却往往无法言说“为什么这么做”；第三则是问责机制（Accountability），即人类清楚自己行为的后果并对决策结果负责。 人类员工在日常工作中遵循着“感知（Sense） $\rightarrow$ 思考与推理（Think & Reason） $\rightarrow$ 行动（Act） $\rightarrow$ 学习（Learn）”的完整闭环。要让流程智能体复刻这一系列机制，就必须依托庞大而复杂的训练元模型（Meta Model）。这套分类体系涵盖了流程要素、任务节点、接触的文件、执行的操作、时间戳、环境变量以及交互的系统等众多参数。目前从个人助手或平台级 Agent 跨越到流程智能体，最核心的“缺失拼图”正是高价值的企业级训练数据。 ```mermaid flowchart LR accTitle: 人类能动性到流程智能体的转化闭环 subgraph S1["人类能动性核心特征"] A["不完备信息推理"] --> D["人类操作闭环<br/>感知/推理/行动/学习"] B["隐性知识与肌肉记忆"] --> D C["行动后果问责机制"] --> D end subgraph S2["模型构建与训练"] D --> E["业务遥测元数据<br/>流程/动作/系统/文档"] E --> F["流程智能体<br/>(Process Agent)"] end ``` - **人类能动性三大支柱**：不完全信息下的推理能力、无法言传的隐性知识（肌肉记忆）、明确的决策问责意识。 - **认知操作闭环**：智能体需要完整模拟人类“感知 $\rightarrow$ 推理 $\rightarrow$ 执行 $\rightarrow$ 反思学习”的行为链条。 - **训练数据困境**：构建流程智能体不仅需要界面和文本，还需要时间戳、系统状态、关联文档与环境参数等全维度元数据。 --- ### [13:06](https://www.youtube.com/watch?v=SryWySKqrQU&t=786s) 从观察到智能体：Skan 的数字孪生与特斯拉类比 在探讨如何获取高质量训练数据时，Vinay 提出了一个“特斯拉自动驾驶”的比喻：特斯拉训练自动驾驶汽车时，工程师并没有一行行手写人类应该如何开车，而是在车辆上安装传感器和摄像头，全面记录人类驾驶时的每一次操作与环境反馈，再将这些海量数据下载用于训练模型。同样的逻辑完全适用于企业流程智能体——通过细致观察人类员工的操作行为，捕获真实工作流数据，进而训练大型动作模型。 这就是 Skan AI 所倡导的核心理念：“从观察到智能体”（Observation to Agent）。Skan AI 建立在流程挖掘与数字孪生（Digital Twin）技术之上。在 Agent 浪潮爆发之前，Skan 主要作为流程情报平台，通过屏幕录制与交互遥测生成流程模型与数据分析；而现在，这些持续捕获的人类操作数据成为了训练智能体最核心的“原材料”。借助基于 Transformer 架构的大型动作模型（Skan LAM），平台将数字孪生中沉淀的操作轨迹转化为行动嵌入，直接生成具备上下文理解能力的流程智能体。 - **特斯拉自动驾驶哲学**：不靠人工硬编码规则，而是通过传感器捕获人类驾驶行为来训练模型；Skan 亦通过捕获员工真实操作来训练 Agent。 - **Observation to Agent**：观察业务操作 $\rightarrow$ 构建流程数字孪生 $\rightarrow$ 提取工作流遥测数据 $\rightarrow$ 训练生成流程智能体。 - **Skan LAM**：基于 Transformer 架构的大型动作模型，以员工操作的动作嵌入作为核心输入。 --- ### [15:50](https://www.youtube.com/watch?v=SryWySKqrQU&t=950s) 实战落地：理赔裁决智能体与人机协同 在现场演示中，Vinay 展示了 Skan 大型动作模型（LAM）在保险理赔裁决（Claims Adjudication）场景中的真实运行情况。屏幕右侧实时显示着 Skan LAM 的“思考脉络”脚本，精确呈现模型在每一步调用的数据与决策逻辑；左侧则是智能体在 Guidewire ClaimCenter 系统中自主执行端到端理赔操作——从系统登录、调取理赔单，到审查批单背景、客户历史赔付与争议记录，再到自主做出“赔付（Pay）”、“拒赔（Deny）”或“挂起待审（Pend）”的终审决策。这与传统的 RPA 机器人有着本质区别，因为该智能体是完全目标导向（Goal-oriented）的，并在三大核心边界约束下自主寻优： 1. **客户满意度（CSAT）**：优先保证顺畅的客户理赔体验； 2. **处理时效（TAT）**：优化周转时间以间接拉升满意度； 3. **合规治理（Compliance）**：严格遵循企业内部决策规则与风控红线。 该方案已经在真实客户的生产环境中平稳运行。在当前的人机协同模式下，一名人类主管可以同时监管 10 到 15 个自动化的理赔流程智能体。遇到模型无法处理的复杂异常时，智能体会自动抛出错误由人类接管解决；更重要的是，Agent 会在人类修正的过程中持续学习，确保下次遇到同类异常时能够独立应对。前 LPL Financial 高管 Dayton Semerjian 在总结时感慨，面对企业数字孪生与流程自动化的演进，选择能够兼顾“当下产生价值”与“长远战略演进”的技术伙伴至关重要，而 Skan 在数字孪生与流程智能方面的实力令人叹为观止。 - **目标导向决策**：流程智能体不仅执行任务，更围绕业务目标（如赔付、拒赔、挂起）进行端到端逻辑判断与系统交互。 - **三大优化目标**：在客户满意度（CSAT）、处理周期时间（Turnaround Time）与合规治理之间寻求最优平衡。 - **1:15 人机协同监控比**：1 名人类专家可管理 10 至 15 个流程智能体，实现异常介入、在线纠偏与持续自进化。 ### [19:13](https://www.youtube.com/watch?v=SryWySKqrQU&t=1153s) 走向 Agentic 的自然延伸与致谢总结 Dayton Semerjian 在演讲尾声总结了与 Skan AI 合作的关键价值。他指出，在观察到企业实际业务运转的真实流程后，顺理成章的下一步就是自然过渡并向上延伸至智能体（Agentic）架构。这正是他们选择与 Skan AI 携手合作的原因，也是全球众多领先企业纷纷与 Skan AI 开展合作的核心动因。 对于所有准备启动相关项目的企业与团队，Dayton 提出了诚恳的建议：务必确保技术架构中具备所有必要的流程感知与上下文分层要素，否则项目极易在实际落地时受阻或陷入困境（"otherwise you're going to get yourself hung up"）。他结合自身的实战合作经历强调，Skan AI 确实具备这一完整的分层能力。最后，两位演讲者互相致谢，并在全场观众的热烈掌声中结束了本场分享。 - **迈向 Agentic AI**：深入理解业务流程运转是企业向自主智能体（Agentic AI）演进的必然路径。 - **架构完整性的重要性**：实施相关项目时，若缺乏底层完整的流程与上下文感知分层支持，企业很容易遭遇阻碍。 - **客户背书**：来自实际客户的落地经验证实，Skan AI 具备支持企业实现 Agentic 转型所需的完整技术分层。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=SryWySKqrQU&t=4s) 困于 POC（概念验证）迷雾：绝大多数企业都在进行大量的智能体实验，但极少数能真正走向生产环境并实现投资回报（ROI）。 — Vinay Mummigatti > [00:04](https://www.youtube.com/watch?v=SryWySKqrQU&t=4s) 许多组织仅仅将智能体视为纯技术工具，而忽视了它实质上是一场运营模式的变革，关键在于构建人与智能体如何协作的目标运营模型（Target Operating Model）。 — Dayton Semerjian > [04:01](https://www.youtube.com/watch?v=SryWySKqrQU&t=241s) 我热爱智能体，因为我们开发的所有智能体，最大的受益者是我们的业务操作员——他们可以早点下班回家，花更多时间陪伴宠物和家人。但在企业损益表（P&L）层面，却没有任何实质影响。 — 某保险公司首席 AI 官（CAIO） > [06:19](https://www.youtube.com/watch?v=SryWySKqrQU&t=379s) 企业不会为了技术而买单，一切投资必须以移出系统成本、交付确定成果为目的。 — Dayton Semerjian > [09:23](https://www.youtube.com/watch?v=SryWySKqrQU&t=563s) 这就像为你聘请一位专属的“私人主厨”——这位厨师在你的厨房里工作，使用你的烹饪器具，完全按照你的营养偏好和定制菜单提供服务，精准交付你所期望的一切。这就是流程智能体（Process Agents）的设计初衷。 — Vinay Mummigatti > [09:23](https://www.youtube.com/watch?v=SryWySKqrQU&t=563s) 如果所有人都使用相同的商品化模型，长远来看这只是一场“零和博弈”（net neutral game），根本无法形成竞争壁垒。相反，企业唯有依靠流程智能体才能持续放大独特优势。 — Vinay Mummigatti > [10:40](https://www.youtube.com/watch?v=SryWySKqrQU&t=640s) 智能体化并非单纯的技术术语，而是将技术赋予“人类能动性”（Human Agency）。 — Vinay Mummigatti > [13:06](https://www.youtube.com/watch?v=SryWySKqrQU&t=786s) 特斯拉训练自动驾驶汽车时，工程师并没有一行行手写人类应该如何开车，而是在车辆上安装传感器和摄像头，全面记录人类驾驶时的每一次操作与环境反馈，再将这些海量数据下载用于训练模型。同样的逻辑完全适用于企业流程智能体 — Vinay Mummigatti > [19:13](https://www.youtube.com/watch?v=SryWySKqrQU&t=1153s) 务必确保技术架构中具备所有必要的流程感知与上下文分层要素，否则项目极易在实际落地时受阻或陷入困境（otherwise you're going to get yourself hung up）。 — Dayton Semerjian ## Mentioned Vinay Mummigatti, Dayton Semerjian, Skan AI, LPL Financial, Ai4, Salesforce, ServiceNow, NICE, Guidewire ClaimCenter, Tesla, Large Action Models (LAM), Transformer, RPA ## Video description Vinay Mummigatti, EVP - Strategy & Customer Transformation, Skan AI Dayton Semerjian, Former Chief Customer Care Officer and Managing Director, Former LPL Financial Abstracts Your million-dollar AI investment is worthless if you can’t monetize it. Your best staff hold the key to success, but how will you transfer human knowledge to your AI agents? We'll explore why your current task-based AI agents are hitting a wall and show you the critical shift to intelligent process agents that understand your entire workflow just like your top employees. So, how do you bridge the knowledge gap? Discover how Skan AI enables enterprises to capture and scale this knowledge, translating it into a context-driven Large Action Model. You'll also hear from a Skan AI customer sharing real-world experience and how they're building the foundation for agentic AI success. Join us to learn the keys to monetize AI in your enterprise with process agents instead of task bots. [Source](https://www.youtube.com/watch?v=SryWySKqrQU)

[0:04] Vinay Mummigatti: 困于 POC（概念验证）迷雾：绝大多数企业都在进行大量的智能体实验，但极少数能真正走向生产环境并实现投资回报（ROI）。 [0:04] Dayton Semerjian: 许多组织仅仅将智能体视为纯技术工具，而忽视了它实质上是一场运营模式的变革，关键在于构建人与智能体如何协作的目标运营模型（Target Operating Model）。 [4:01] 某保险公司首席 AI 官（CAIO）: 我热爱智能体，因为我们开发的所有智能体，最大的受益者是我们的业务操作员——他们可以早点下班回家，花更多时间陪伴宠物和家人。但在企业损益表（P&L）层面，却没有任何实质影响。 [6:19] Dayton Semerjian: 企业不会为了技术而买单，一切投资必须以移出系统成本、交付确定成果为目的。 [9:23] Vinay Mummigatti: 这就像为你聘请一位专属的“私人主厨”——这位厨师在你的厨房里工作，使用你的烹饪器具，完全按照你的营养偏好和定制菜单提供服务，精准交付你所期望的一切。这就是流程智能体（Process Agents）的设计初衷。 [9:23] Vinay Mummigatti: 如果所有人都使用相同的商品化模型，长远来看这只是一场“零和博弈”（net neutral game），根本无法形成竞争壁垒。相反，企业唯有依靠流程智能体才能持续放大独特优势。 [10:40] Vinay Mummigatti: 智能体化并非单纯的技术术语，而是将技术赋予“人类能动性”（Human Agency）。 [13:06] Vinay Mummigatti: 特斯拉训练自动驾驶汽车时，工程师并没有一行行手写人类应该如何开车，而是在车辆上安装传感器和摄像头，全面记录人类驾驶时的每一次操作与环境反馈，再将这些海量数据下载用于训练模型。同样的逻辑完全适用于企业流程智能体 [19:13] Dayton Semerjian: 务必确保技术架构中具备所有必要的流程感知与上下文分层要素，否则项目极易在实际落地时受阻或陷入困境（otherwise you're going to get yourself hung up）。

[0:04] The Proof-of-Concept (POC) Fog [0:04] Dayton Semerjian: school of hard knocks [0:04] Dayton Semerjian: team sport [4:01] Oh, I love agentic [6:19] Inquiry to Onboarding [6:19] Transaction to Fulfillment [6:19] Trouble to Resolve / Issue to Resolution [13:06] Observation to Agent [16:09] Vinay Mummigatti: how the brain is working step by step [18:03] Dayton Semerjian: breathtaking to see... in action

[0:04] Welcome to the presentation. So we'll [0:06] quickly move on to the uh the first [0:08] slide which is let's start with uh what [0:11] is the state of agentic AI today. Uh if [0:14] you generally see there is a a lot of [0:16] buzz about agentic and uh in general we [0:19] see that uh there is a lot of pressure [0:22] on most enterprises to firstly bring [0:25] change using agentic and you are also [0:28] expected to show value and all this is [0:30] because we are at a inflection point in [0:32] agentic and the inflection point is [0:35] driving very large disruptions in the [0:38] human productivity in the enterprise [0:40] operating models and at industry. ries [0:43] at various industries at large. So with [0:46] these kind of pressures coming on all [0:48] levels of management, we are seeing [0:50] there are distinct trends in the agentic [0:53] space. The first is almost every [0:55] enterprise today is stuck in the proof [0:57] of concept fog uh where every company is [1:00] trying to do a lot of agentic [1:01] experimentations and very few of those [1:04] experiments are reaching production [1:05] stage or return on investment stage. [1:08] Secondly, there is also a gap in terms [1:10] of the kind of data you need for running [1:12] agents because many times there are [1:14] standard very cookie cutter agents. [1:16] Hundreds of vendors are offering them. [1:18] The moment you come to certain specific [1:20] processcentric agents which we will be [1:21] talking more, you need enterprise [1:24] specific training data for running these [1:26] models and that's a big gap. And [1:28] thirdly, it is also the vendor lockin [1:30] where because of the sense of urgency, [1:33] we are all focusing on picking up one [1:36] ecosystem vendor and trying to put all [1:38] our dollars into that particular vendor [1:40] that is limiting us [clears throat] from [1:41] the art of possible because this agentic [1:43] space is fast changing and fast [1:45] evolving. And the last is we are all [1:47] looking at agentic to be like more of a [1:49] technology play. It is more of an [1:51] operating model play. So we need to look [1:53] at how humans and agents are going to [1:55] interplay as a target state operating [1:57] model. So bringing all this together you [2:00] see there is a promise and there's a [2:02] reality and the gap between these two is [2:04] what we are trying to address at scan AI [2:07] with our journey in the process agent [2:09] space. Uh Jayton would like to add [2:11] anything on this. [2:11] >> Yeah I think from the I think from the [2:13] business side there's a couple of things [2:14] that from the kind of school of hard [2:16] knocks that I wanted to share with folks [2:18] out there trying to go into production. [2:20] for some of the challenges in that proof [2:22] of concept on the business side is some [2:23] people uh don't get a clear business [2:25] case or they kind of underestimate uh [2:27] the value that's going to be created. So [2:29] having a comprehensive kind of [2:31] multi-year view of what your savings are [2:33] going to look like I think is um is [2:35] absolutely uh absolutely important [2:37] there. Um the other thing that folks [2:39] don't understand sometimes and we didn't [2:41] understand it when we came out of the [2:42] gates is that sometimes the technology [2:44] works fine but to realize the value you [2:47] have to reorganize how work gets done to [2:50] move FTE out of the system or redeploy [2:53] them to some other part and so we stared [2:55] at it for a little while and once we [2:57] realized the work isn't done around [2:59] organizational design after deployment [3:01] we actually saw some of the hard [3:03] benefits. Scan does something very well [3:05] in this data metrics area which is [3:07] sometimes it's impossible to see your [3:09] own processes because there's so much [3:11] variability in there in being able to [3:13] have a clean data set being able to have [3:16] solid data governance and having a great [3:18] foundation is the prerequisite to being [3:21] able to pull this off. And I think the [3:23] last thing around the missing playbook I [3:25] think the missing playbook is at [3:27] sometimes the function of having a [3:28] missing team. And so this is a team [3:30] sport. So when you're deploying a team [3:32] to do this to build the playbook, it's a [3:34] business process owner. Um it's a it's a [3:37] data analyst, it's a six sigma uh [3:39] leader, it is a change management [3:41] person. You have to put the whole team [3:43] on the field in order to get things out [3:46] of pilot into deployment and then to [3:48] kind of capture and recognize the value. [3:51] >> Thanks Jayton. [3:52] >> Sure. [3:52] >> Can we switch to the slide please? Sorry [3:54] the clicker is not working. So there [3:56] will be some manual back and forth on [3:57] slide switching. Uh so this is a vast [4:01] opportunity as we all appreciate because [4:03] we are fundamentally driving an industry [4:05] level change and this brings a question [4:07] like is there one vendor or one tool [4:09] that will be enough for an enterprise [4:11] definitely no and that's why we see uh [4:14] explosion in terms of the number of [4:16] vendors. Just in this conference you'll [4:17] probably find a few hundred vendors and [4:19] then at an industry level you're [4:20] probably going to see a few thousand [4:22] vendors at play. But essentially we are [4:24] trying to classify these vendors because [4:26] we want to analyze where we fit in in [4:28] this ecosystem. So on the extreme left [4:30] you'll see vendors who are very focused [4:32] on personal agents. The focus is on [4:35] improving personal productivity and [4:37] these are typical email automations or [4:40] document automations and various kinds [4:42] of routine tasks that are getting [4:44] automated through agents. Now move to [4:46] the right you'll see platform agents [4:48] where the likes of Salesforce service [4:50] now and uh nice and few other companies [4:53] are launching agents which are trained [4:55] on their specific data models and their [4:58] specific processes. So you still live [5:00] within the ecosystem of those platforms [5:02] which is fine. You are serving a purpose [5:04] but there is a containment. As you move [5:06] to the right you will see the role for a [5:08] process agent. And what we mean by [5:10] process agents is when you look at an [5:12] enterprise, you're running very complex [5:14] processes which run your transactions [5:16] you which are basically entire customer [5:18] journey is being automated through these [5:20] processes. Now this is where the role of [5:22] process agents come into play where you [5:24] are cutting across your various systems [5:26] of records, transaction master sources [5:29] and uh complex decisioning involved [5:31] various kinds of exceptions. So this is [5:33] the most complex area where we see [5:35] enterprises struggling with today. And [5:37] if you go from left to right, you'll see [5:39] that most of investment today is locked [5:41] in on the left side with very little ROI [5:44] to show. Come to the right, you'll find [5:46] that very little investment because it's [5:48] a struggle area and this is the highest [5:50] ROI area for us. Uh we were talking to [5:52] one of the chief AI officers of an [5:54] insurance company and he basically made [5:56] a very nice statement. He said, "Oh, I I [5:58] love agentic." Uh because all the agents [6:01] we are developing are the biggest [6:03] beneficiaries of those agents are our [6:05] operators because they can go home early [6:07] and uh spend more time with their pets [6:09] and families. But at an enterprise [6:11] level, there is no P&L impact and that [6:14] is the reality of the uh process agents. [6:18] Can we go to the next one please? Yes. [6:19] Thank you. Let's go a little deeper into [6:22] uh what drives the ROI for these process [6:25] agents. Now when you look at uh most [6:27] companies today almost 60 to 70% of our [6:30] operating budgets are locked in labor [6:33] compensation costs. If you have to make [6:35] any dent in the productivity or the cost [6:38] structure of a company, you have to [6:40] touch into the areas where humans are [6:43] actively involved. And that is where you [6:45] see that most value chains whether it is [6:49] inquiry to onboarding which is early [6:51] part of customer journey or transaction [6:53] to fulfillment [6:55] this is a most of the customer journey [6:57] is locked into this space and then there [6:59] is trouble to resolve or issue to [7:01] resolution which is customer service [7:02] area. So traditionally humans are logged [7:05] into these areas and that is where we [7:07] see there's a big opportunity for [7:09] process agents but these process agents [7:11] are essentially meant to do five things [7:14] which is there has to be a goal which is [7:17] definition of what is a business [7:18] outcome. Then there is a strong element [7:20] of process where it is a sequential [7:22] execution of certain activities. There's [7:25] a context which brings enterprise data, [7:27] master data sources, customer history [7:29] and so many other things. Then there are [7:31] various tools and actions to be [7:33] required. And then last but not the [7:35] least, you want to continuously optimize [7:37] your processes. So when you bring all [7:39] these five elements together, this is [7:41] where the process agents get very [7:43] complex. I mean they are complex but [7:46] they are at the heart of generating ROI [7:48] from agentic and this requires a [7:51] paradigm shift in our outlook from task [7:53] automation to outcome automation. So [7:55] that's where we are going. You want to [7:56] add anything? The only the only thing [7:58] I'll add is and you you said it nicely [7:59] which was outcome automation there you [8:01] know nobody's buying technology for the [8:03] sake of technology and being on I had [8:05] been previously a vendor as well as a [8:07] customer using technology to get the [8:09] cases approved you have to be delivering [8:11] an outcome you have to be uh moving cost [8:14] out of the system it's absolutely uh [8:16] doable and a lot of people make the [8:18] mistake of focusing only on productivity [8:20] productivity is nice uh but the purpose [8:23] of management is to make sure that the [8:24] productivity is there I will tell you a [8:26] Funny anecdote though, using the scan [8:28] technology, I we'd been talking to an [8:31] operations team who had told me how hard [8:32] they've been working and they had no [8:34] time, they had no capacity. And after we [8:36] deployed that technology, uh we took a [8:38] peek and found that 20 to 30% of the [8:41] time was used on social media sites and [8:43] other areas that weren't necessarily [8:45] driving productivity and production. But [8:47] I leave that not to the business case. I [8:48] leave that to management uh to go get [8:50] sorted out. But the technology is there. [8:54] >> Thanks. [8:55] >> Yeah. [8:56] So what makes process agents different [8:58] right I mean there are many types of [9:00] agents u I'd like to start with an [9:03] analogy if you're looking at a fast food [9:05] chain like you are going to a burger [9:07] chain shop or pizza like what do we get [9:10] there is a cookie cutter menu available [9:12] you just have to pick and choose what is [9:14] offered by the restaurant or the chain [9:17] now this is exactly what is happening in [9:19] the personal and platform agents where [9:21] they are trained on generic models which [9:23] are commodity there's no differentiation [9:26] not specifically trained on your [9:27] enterprise data and this is where most [9:29] of the commodity large language models [9:31] are at play. Switch to the right. The [9:36] just imagine you are hiring a personal [9:37] chef who works in your kitchen with your [9:40] gadgets for your nutritional preferences [9:42] for your menu and basically they are at [9:45] your service trained to deliver exactly [9:48] what you expect from your chef. That is [9:51] exactly what process agents are meant to [9:53] do. They are supposed to be driving your [9:55] enterprise processes in the enterprise [9:58] data models with your enterprise [10:00] compliance decision rules. At the heart [10:02] of this whole process agents lies your [10:04] competitive advantage. Now if everybody [10:07] is using the same agents like there are [10:09] hundreds of companies who can access a [10:11] agents which are available in the [10:12] marketplace today. How are you going to [10:14] make any difference to your competitive [10:16] differentiators or advantage? In the [10:18] long run it's a net neutral game. But [10:20] the process agents are going to be the [10:21] ones where you will continue to thrive [10:23] on your competitive advantage. And that [10:25] is where the large action models come [10:27] into play which are the next evolution [10:29] of language models specifically [10:31] embedding the actions and the processes [10:34] that your organization runs. So maybe [10:37] let's go to the next one. And the big [10:40] question why are process agents so [10:42] difficult to build? Now at the heart of [10:45] this whole um process agents is the word [10:48] agentic. Now agentic stems from the [10:51] keyword human agency like it's agentic [10:54] is not a a technology term. It's [10:56] basically agentifying or applying [10:59] technology to the human agency. That's [11:01] where this whole word is or the word [11:02] comes from. Now what defines human [11:04] agency? So there are three terms that [11:07] come to my mind when I say human agency. [11:10] The first is our ability to abstract and [11:13] reason with incomplete information. [11:15] That's what defines us. You put an [11:16] operator on a call center, they don't [11:18] have all the information. They still [11:19] make decisions day in and day out in [11:21] real time. Then there is a tacit [11:23] dimension where humans know a lot of [11:26] things but we cannot tell why we did [11:28] certain things. And that is an embedded [11:30] knowledge or muscle memory you can call [11:32] it. Then the third element is the [11:33] accountability. Okay, I did lot of [11:36] things but then I'm also accountable for [11:37] it. I know that what is the consequence [11:39] of my actions. Now when you combine all [11:41] these three that's what makes the [11:44] process agents very unique because now I [11:47] have to emulate these three concepts in [11:49] an automation tool or an agent and for [11:52] that we also need to look at the human [11:55] operators how they work is we have [11:56] ability to sense ability to think and [11:59] reason then act and then learn. So if [12:03] humans are able to do this, why can't [12:05] process agents be trained to do this? [12:07] And that is at the heart of today's [12:08] discussion. Now to train these agents [12:11] it's not easy because you need to learn [12:13] uh you need to train them on many inputs [12:16] and we have identified a few inputs [12:18] which we have tried to capture as the [12:20] meta model of training these agents or [12:22] call it as a taxonomy of this agentic u [12:26] training model and it includes the [12:28] process elements task elements the [12:30] documents we touch the actions we [12:32] perform the date and time stamps [12:34] environment variables the systems we [12:36] touch this is not comprehensive by the [12:37] way I'm just capturing [12:39] key elements but it goes much beyond [12:41] this because we have to touch many [12:42] parameters. Now when you look at this [12:44] complexity [12:46] the missing link here from the personal [12:49] and the platform agents to the process [12:51] agents is the training data. Where does [12:54] this data come from? If you have to [12:55] train agents that are specific to your [12:57] enterprise performing your specific [12:59] enterprise processes with your [13:01] competitive advantages and that's what [13:03] we'll go next. Yeah. [13:06] So um just imagine you you're like Tesla [13:10] and then you are building your um [13:12] self-driving cars. Um how do you train [13:15] the self-driving cars? You don't sit and [13:17] program how cars have to be driven. You [13:19] have basically installed sensors, [13:21] cameras and various other tools which [13:23] are capturing every event that humans [13:25] are performing when they are driving and [13:27] then this data is downloaded to train [13:29] your models for your self-driving car. [13:32] Now if process agents have to behave [13:34] like human agents shouldn't we be doing [13:36] the same observe what humans do capture [13:39] the data and model the um you know the [13:41] large action models that's what is the [13:44] baseline of how we are approaching from [13:46] scan can you go to the next slide please [13:48] yeah so our philosophy we we sum it up [13:52] as observation to agent essentially what [13:55] scan does is scan is a process [13:56] intelligence platform built on the [13:59] digital tw technology and AI is at the [14:01] heart of [14:02] But until the advent of agentic [14:04] approach, we were basically a good [14:06] process intelligence platform generating [14:08] various types of analytics, the data, [14:10] models or to show you how your work is [14:13] performed. But now we have found that [14:15] the data we are collecting is [14:17] essentially the raw material or the um [14:20] it's it's basically the key ingredient [14:22] to train your agentic models. And this [14:24] is what we learned working with our [14:26] customers that the data we have been [14:27] generating by observing how operators [14:30] perform is basically able to predict [14:33] every action and using that data we are [14:36] now able to create agents and this is [14:37] what we have been doing as a next step [14:39] we'll explain how we are doing it and um [14:42] I'll show you two demos quickly to uh [14:44] validate these concepts the first is the [14:47] concept of digital twin now you are [14:49] seeing scan in action here where scan is [14:52] observing what an operator is doing. [14:54] It's looking at all the screens [14:56] recordings. It also converts those [14:58] recordings into understandable data [15:00] models and then it generates these kind [15:02] of process models, data analytics, [15:04] predictive models etc. So this is the [15:07] basic [15:08] foundation of what scan does. Now using [15:11] these models and the data we generate [15:13] here, we are now able to train the [15:16] process agents. So let's click. [15:20] So this is the heart of what scan does [15:23] basically generating all the telemetry [15:25] of work data and now we have created [15:27] this scan large action model which [15:30] basically is a transformer model using [15:34] the action embeddings that we have [15:36] captured by observation. It's a concept [15:38] is very simple. We observe we create a [15:41] digital twin. We use the data from [15:42] digital twin using the transformer [15:44] model. We generate the agents. Now this [15:47] is the core concept and the next slide [15:50] is a little bit tricky where I'm going [15:52] to show you how our process agent is [15:54] going to run in a claims adjudication [15:56] scenario. Uh just pause just okay maybe [16:01] yes thank you uh so here I'm showing you [16:04] claims adjudication agent in action [16:07] basically this has been trained using [16:09] the scan digital twin data and what we [16:11] have done is on the right side you will [16:13] see a script appearing that is the scan [16:16] large action model in action it is [16:19] exactly telling how the brain is working [16:21] step by step what decisions are being [16:23] taken what data is being considered [16:25] everything is appearing here and on the [16:27] left side you will see the actual claims [16:29] agent in action and this agent is going [16:32] to perform a complete claims [16:34] adjudication uh action all the way from [16:36] logging into the guidewire claim center [16:39] looking at the list of claims picking up [16:41] a claim looking at the context [16:44] policy endorsements various undertaking [16:46] various decisions in terms of whether to [16:48] pay the policy bend it or deny it and [16:51] then taking it all the way through [16:53] completing the transaction. This is not [16:55] what a traditional automation tool like [16:57] an RPA bot or other kinds of automations [16:59] would do because firstly this is goal [17:02] oriented. You can click it please. Yeah. [17:05] So firstly there is a goal aspect. So [17:07] the goal we have defined in the slam or [17:09] large action model is either pay the [17:11] claim, deny the claim or pend the claim [17:14] or keep it pending. Second is we have [17:16] given a complete context to it where the [17:19] model knows exactly what is the policy [17:21] endorsement the history of this customer [17:23] past payment histories past disputes [17:26] everything it has ability to analyze. It [17:28] also knows the complete uh actions that [17:31] it has to perform and then it also knows [17:33] the optimization. These models are [17:35] trained for optimization where there are [17:38] three parameters given to this model [17:39] here. One is optimize for seesat [17:42] customer satis satisfaction. Second is [17:44] to optimize for the time because if you [17:46] have faster turnaround time your [17:48] customer sees that automatically goes [17:49] up. Third is the compliance. You need to [17:51] be highly compliant to the enterprise [17:53] rules and uh governance here. So we have [17:56] given these three as a constraining [17:57] factors and then you will see how this [18:00] agent is performing within those [18:01] boundaries. So essentially the lift we [18:03] are going to see here is and this has [18:05] been in production for one of our [18:06] customers right now where one human [18:09] agent is able to operate or look at 10 [18:11] or 15 uh automated claims process agents [18:15] and every time there is an exception [18:17] which the agent or the process agent is [18:19] not able to solve it throws an error and [18:21] then human goes and fixes it. It learns [18:23] from it and next time if the same error [18:25] happens it is aware of how to solve the [18:27] problems. So this is what we have been [18:30] achieving so far in terms of our journey [18:32] on a on a process agent uh technology [18:35] and we're not there yet. It is a long [18:37] journey because we have to evolve this [18:39] into so many other factors as we go to [18:41] more complex organizations. But then I [18:44] just wanted to share the whole journey [18:45] and the background and the concepts [18:47] here. Data you would like to summarize [18:49] let me just sure let me just wrap up [18:51] because I can see the shot clock is [18:52] about to uh the shot clock is about to [18:54] go off. you know, from the from the [18:56] enterprise perspective, what we care [18:57] about is making sure you get value [18:59] today, you're going to get value [19:00] tomorrow. And it's about partner [19:02] selection. So, we look at like the left [19:03] side of that slide and the building the [19:06] digital twin, process optimization, [19:08] process automation. They have it in [19:10] spades today. I've used it in spades. [19:12] It's breathtaking to see that technology [19:13] in action and it's a natural extension [19:15] to go up into uh go up into a genenic. [19:18] That's why we worked with you. I know [19:19] that's why uh enterprises around the [19:21] world are working with you and I would [19:23] uh I would encourage people especially [19:25] anybody who's going to be taking on any [19:27] projects to make sure that you have all [19:29] of those different elements of the layer [19:31] in there otherwise you're going to get [19:32] yourself hung up and I know for a fact [19:34] having worked with you that scan has [19:36] that. So thank you very much appreciate [19:37] it. [19:39] Thank you very much. [applause]