Veritiv Deploys AI Agents Across ERPs to Automate 1.7 Million Orders

Veritiv跨多套ERP部署AI Agent，实现170万笔订单自动化 Veritiv利用集成至现有软件的计算机操作型AI Agent，在不更换核心基础设施的前提下，实现170万笔ERP手动订单录入自动化，每年可节省6万小时运营工时。 面对拥有十多套传统ERP系统的碎片化IT环境，Veritiv构建了名为Invaiet的定制化Agent框架，以实现Salesforce与Oracle之间手动订单录入工作流的自动化。通过将浏览器导航能力与精益流程优化相结合，Veritiv将复杂的业务工作流（如退款与重新开票）从36个步骤缩减至4个，目标每年节省多达6万小时的运营工时。演示文稿概述了战略部署原则，重点强调依托现有工具促进员工融入、持续进行自研与采购评估，以及追踪单次成功的总成本而非单Token成本。 Veritiv是一家年营收80亿美元的专业包装分销商，其IT环境高度碎片化，在120个分销中心运行着十多套各不相同的ERP系统。 (0:04) 事实证明，传统RPA过于僵化，无法在高度定制化的传统企业系统之间实现工作流自动化。 (0:04) Computer Use（计算机操作）功能的发布，为利用AI Agent充当贯通传统架构的“数字粘合剂”提供了技术基础。 (0:04) Veritiv成立了专门的AI转型实验室，利用内部运营与数据科学人才来设计Agent工作流。 (0:04) 客户体验被视为AI赋能杠杆率最高的领域，其目标是实现5到10倍的人工生产力提升，而非缩减员工编制。 (0:41)

Chris Bradley details how specialty packaging distributor Veritiv deployed computer-use AI agents across more than a dozen legacy ERP systems to achieve operational leverage rather than cutting headcount. Unifying Fragmented Legacy ERPs via Computer Use (0:04) Veritiv, an $8 billion specialty packaging distributor with approximately 7,500 employees and 120 distribution centers, accumulated more than 12 distinct ERP environments—including SAP, Oracle, Amtech, and homegrown software—through organic growth and acquisitions. Traditional Robotic Process Automation (RPA) struggled within this fragmented architecture because customizations implemented by past Global System Integrators confined automation to narrow edge cases, requiring pervasive manual hand-offs and duplicate data entry across legacy platforms. Following Google's late November release of computer-use capabilities, Bradley identified browser-navigating agents as a digital glue capable of bridging disconnected enterprise systems without complete architectural overhauls. After an executive briefing post-Christmas, Veritiv established an internal AI Transformation Lab in February, pairing Lean Six Sigma operational leadership with senior data science talent to design agentic workflows across order entry and financial operations. Deploying the Invaiet Harness for Order Entry (0:41) An internal assessment across 7,800 employee roles pinpointed Customer Experience as the company's highest-leverage automation opportunity, specifically targeting workloads managed by 420 customer service representatives and 500 financial operations personnel. Facing a lack of commercial off-the-shelf software for emerging computer-use models, Veritiv developed a proprietary agent harness named Invaiet. The system is designed to automate the keying of 1.7 million orders submitted to Oracle annually, targeting the recovery of up to 60,000 human labor hours each year. To minimize change management friction, Invaiet operates entirely within existing Salesforce interfaces, triggering workflows through status fields and case comments. Structurally, the agent executes logins and interface navigation using a secure Chromium browser, injects parsed order data directly into Oracle via API, and posts completion updates back to Salesforce. Veritiv transitioned the platform from pilot to production within four months, aiming to handle roughly 500,000 orders in the near term. Lean Optimization and Deployment Economics (11:05) Bradley emphasizes that organizations must streamline legacy workflows using Lean methodologies before applying AI agents. In a credit and rebill workflow optimization, a Kaizen event reduced a 36-step manual process down to 4 steps prior to automation, lowering turnaround cycle times from 7–15 days to under 24 hours. Bradley also maintains that technical leadership should actively engage as Forward Deployed Engineers to directly monitor evolving frontier model benchmarks. From an architectural perspective, Veritiv treats model selection as a model-agnostic continuum rather than a static procurement choice. Because multi-turn agents average 30 to 50 non-deterministic turns per order, the lab assesses model performance using total cost per success rather than basic token pricing. The company utilizes OpenRouter to run continuous evaluations against actual order histories, simultaneously compiling fine-tuning datasets to reduce long-term inference costs and retain leverage to upgrade or switch model backends. Veritiv frames computer-use agents as a practical strategy for elevating workforce productivity and margins on top of rigid legacy infrastructure, while ongoing implementation milestones continue to hinge on managing multi-turn execution reliability and optimizing end-to-end task costs.

包装分销巨头Veritiv通过自主研发的Agentic套件Invaiet与计算机操作（Computer-Use）技术，将分散在十多个遗留ERP系统中的业务流程进行数字化连接。公司副总裁Chris Bradley强调，该战略旨在通过精益流程重构与人机协同为一线员工提供5至10倍的工作杠杆，而非削减人员编制。 破解多系统割裂与RPA局限 (0:04) Veritiv是一家年营收达80亿美元的专业包装分销商，在美墨两国拥有约7,500名员工及120个配送中心。由于业务增长与历史并购，公司IT环境高度碎片化，运行着包括Oracle、SAP、Amtech及自研软件在内的12套以上不同ERP系统。这些遗留环境以往依赖大量人工交接，导致数据重复录入且可见性极低；而传统机器人流程自动化（RPA）由于系统历经系统集成商（GSI）深度定制，仅能覆盖狭窄的边缘场景，难以规模化扩展。 转机出现在Google发布具备Computer Use能力的模型之后。意识到能够模拟浏览器操作的计算机操作Agent可以作为打通遗留架构的“数字黏合剂”，Bradley在圣诞节后紧急召开CEO会议，并于2月正式组建AI Transformation Lab。为确保跨系统业务闭环，实验室吸纳了持续改进与精益六西格玛主管Joe Condron优化作业流程，并配置顶尖数据科学家挖掘高价值业务切入点。 Invaiet架构与订单录入自动化 (0:41) 通过与首席人力资源官梳理全公司7,800名员工的岗位职能，团队发现客户体验（CX）领域存在最大的赋能空间，其中包含420名客服代表及500名深陷繁琐流程的财务运营人员。针对每年需要手动录入Oracle系统的170万笔订单，Veritiv打破传统的“自研还是购买”界限，开发出名为Invaiet的自研Agentic工作流套件，首要目标是每年为业务部门挽回多达60,000个工作小时。 在人机交互设计上，团队放弃推出独立新软件，选择让Agent直接嵌入员工现有的Salesforce CRM中，通过变更工单状态（如“Via assigned”或“Via needs input”）与工单备注进行协作。Invaiet负责解析各类订单及手写采购凭单并转化为JSON载荷，再通过受控的Chromium浏览器模拟登录Oracle完成界面导航，最终以API直接注入行项目并更新回Salesforce。该项目在4个月内从试点推进至生产，并计划在后续3至4个月内实现承接约50万笔订单的规模。 精益优化先行与消除流程浪费 (11:05) Bradley总结指出，在引入AI之前，企业必须运用精益管理消除既有流程中的浪费，否则只会加速混乱。在近期的信用与重开票（credit and rebill）流程优化中，团队率先开展改善活动（Kaizen event），将原先冗长分散的36个审批与操作环节大幅精简为4个核心步骤。在流程重构完成后接入Agent，该业务的整体处理周期直接从原本的7至15天压缩至24小时以内。 工程落地与“单次成功成本”评估 (11:05) 为保持对技术前沿的敏锐度，Bradley亲自担任前线部署工程师（FDE）。在系统架构上，Veritiv将“自研与外购”视为连续可调的旋钮而非静态选择，保持基础设施对底层模型的解耦，以便随着部署规模扩大在前沿闭源模型与开源权重模型之间无缝切换。同时，团队利用OpenRouter并结合真实历史订单数据评估新模型，并保存交互数据用于微调，以持续压低长期运营成本。 在评估体系上，Veritiv放弃单纯关注每Token单价，转而采用“单次成功成本”（cost per success）衡量全流程真实支出。Bradley指出，由于多轮非确定性Agent处理单个订单往往需要执行30至50个步骤交互，仅计算Token成本无法反映端到端任务的实际经济性。该策略的核心在于将人力从机械录入中解脱出来，转型至更高价值的客户服务、销售提速与利润率改善场景。 Veritiv的实践展示了传统分销企业如何利用Agentic技术打通遗留ERP壁垒，但多轮非确定性交互的高复杂度对任务稳定性和最终完成成本提出了持续挑战，企业仍需在模型稳定性、精益流程改造与长期微调策略之间寻求平衡。

## Background 本文档源自 Ai4 于 2026 年 8 月 14 日在 YouTube 发布的演讲视频《The Agentic Enterprise: Building a Digital Workforce on Top of Legacy Systems With Veritiv》。视频主要由 Veritiv 的首席 AI 官（Chief AI Officer）Chris Bradley 讲解，展示了 Veritiv 如何利用 Agentic AI 和 Computer Use 模型，在不更换核心遗留系统的前提下推动企业数字化转型。 ## TL;DR 面对拥有十余套老旧 ERP 系统及复杂业务流程的挑战，Veritiv 成立了“AI 转型实验室”，并自研了名为 Invaiet 的 Agent 框架。通过将计算机使用（Computer Use）技术与精益六西格玛流程优化相融合，Veritiv 将 Agent 无缝嵌入客服已有的 Salesforce 工作流中，成功实现了每年百万级订单录入与退款流程的高度自动化。演讲分享了企业落地 Agent 的核心经验，包括领导者亲身实践、精益流程再造、模型中立架构以及以“单次成功成本”评估模型性能等关键策略。 ## Atmosphere & style 整场演讲充满务实、严谨且前瞻性的工业级干货气息，讲者以亲历者的视角的深入阐述，既有技术落地的冷静剖析，又展现出推动传统企业重塑的激情与魄力。语言风格干练清晰、逻辑缜密，将高深的 AI 技术与接地气的企业实际管理经验紧密结合，带给听众强烈的实操启发感与变革信心。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=b24cDV2E8I4&t=4s) Veritiv 作为大型特种包装分销商，面临年营业额 80 亿美元、7,500 名员工及十余套老旧 ERP 系统带来的极高系统复杂度与数据透明度问题。 - [02:06](https://www.youtube.com/watch?v=b24cDV2E8I4&t=126s) Google 发布的 Agent 电脑使用（Computer Use）功能打通了无 API 接口的遗留系统，促使 Veritiv 成立“AI 转型实验室”推进自动化。 - [06:17](https://www.youtube.com/watch?v=b24cDV2E8I4&t=377s) Veritiv 打破了 30 年“只买不自研”的传统，针对老旧系统开发了自研 Agent 框架“Invaiet”。 - [06:17](https://www.youtube.com/watch?v=b24cDV2E8I4&t=377s) 为减少员工学习成本，Agent“Via”被无缝嵌入 Salesforce 工作流中，并实现了争议场景下的人机协作（Human-in-the-Loop）。 - [06:17](https://www.youtube.com/watch?v=b24cDV2E8I4&t=377s) Invaiet 平台通过前端 GUI 登录与 API 注入的混合模式处理订单，试点 4 个月内即准备承担每月约 50 万笔订单量。 - [11:05](https://www.youtube.com/watch?v=b24cDV2E8I4&t=665s) AI 领导者必须亲自投入实践（做首席前线工程师），深刻理解技术边界以适应每周都在迭代的 AI 技术。 - [11:05](https://www.youtube.com/watch?v=b24cDV2E8I4&t=665s) 在引入 AI 自动化之前，必须先通过精益 Kaizen 消除流程浪费；Veritiv 将退款流程从 36 步压缩至 4 步，处理周期从 7-15 天缩短至 24 小时内。 - [13:59](https://www.youtube.com/watch?v=b24cDV2E8I4&t=839s) 将 AI 融入员工既有的工作工具（如 Salesforce、Teams）中，可有效降低变更管理、培训与运维成本。 - [13:59](https://www.youtube.com/watch?v=b24cDV2E8I4&t=839s) Agent 架构应保持“模型中立”，初期使用前沿商业模型，后续可根据规模切换或微调开源模型以控制成本。 - [13:59](https://www.youtube.com/watch?v=b24cDV2E8I4&t=839s) 评估 Agent 性能应关注“单次成功成本”（Cost per success）而非单一 Token 成本，并针对不同子任务搭配专用模型。 - [13:59](https://www.youtube.com/watch?v=b24cDV2E8I4&t=839s) 企业引入 AI 的核心目的是为人赋能与提升杠杆率，而非单纯替代人力，从而将员工解放至高价值业务中。 ## Detailed walkthrough 以下是该演讲内容的逐段详细记录与梳理： ### [00:04](https://www.youtube.com/watch?v=b24cDV2E8I4&t=4s) Veritiv 业务概况与复杂遗留系统的挑战 Veritiv 是北美最大的特种包装分销商，年营业额约 80 亿美元，拥有 7,500 名员工以及遍布美国和墨西哥的 120 个分销中心，为数万家客户提供包装与设施用品。随着多年的有机增长与并购（M&A），公司内部积累了极高的系统复杂度。目前企业内部同时运行着十几套 ERP 系统，包括 Oracle、SAP、Amtech 以及各类自研系统。 这些遗留系统严重依赖人工对接，数据重复且透明度低。尽管公司很早就在使用传统 RPA 进行自动化，但由于许多系统（即使像 Oracle 这样的大型系统）都经过了外部系统集成商（GSI）深度的定制与配置，导致传统自动化工具和早期 Agent 只能覆盖极其边缘的场景，无法在企业核心业务中大规模推广。 - **公司规模**：年收入约 80 亿美元，员工约 7,500 人，拥有 120 个分销中心。 - **系统瓶颈**：拥有十余套 ERP 系统（Oracle、SAP、Amtech 及自研系统），遗留代码长达四五十载，大量依赖人工跨系统操作。 ### [02:06](https://www.youtube.com/watch?v=b24cDV2E8I4&t=126s) 突破时刻与 AI 转型实验室的设立 转折点发生在上一年底（11 月底），谷歌发布了具备 Computer Use（电脑使用）能力的 Agent 技术（如 Google Anti-Gravity）。这项技术能够像人类一样操作浏览器与桌面软件，彻底打通了过去无法用 API 连接的复杂系统。Chris Bradley 敏锐地意识到这一革命性变化，在圣诞假期结束后立即与 CEO 召开会议，强调企业面临“不进则退”的临界点。 公司于次年 2 月正式成立了“AI 转型实验室”（AI Transformation Lab），旨在为员工赋能，将订单录入、财务运营等原本只能由人工完成的繁琐流程实现自动化。为了确保落地效果，Chris 跨部门抽调了顶级人才：包括精益六西格玛/持续改进负责人 Joe Condron（负责重构流程以最大化 AI 效益）以及顶级数据科学家（负责分析业务数据，精准筛选高价值用例）。 - **团队架构**：由 CMO兼 AI 转型负责人带队，联合精益六西格玛主管与首席数据科学家。 - **目标设想**：从 7,800 名员工中按岗位梳理，发现客户体验（CX，涉及 420 名客服）与财务运营（500 人）是最大的效率杠杆空间。 ### [06:17](https://www.youtube.com/watch?v=b24cDV2E8I4&t=377s) 自研 Agent 平台“Invaiet”与订单录入落地 在评估市场现成方案时，由于技术迭代极快且缺少能处理几十年老旧系统的平台，Veritiv 决定打破过去 30 年“能买绝不自研”的惯例，自研了一套定制化的 Agent 框架——**Invaiet**（Generative AI Agent）。Invaiet 的首个核心场景是解决每年 1.7 百万笔手动录入 Oracle 的订单问题，该场景蕴含着每年高达 60,000 小时的人工节省潜力。 在产品设计上，团队避免增加额外的学习成本，将 Agent 完美无缝地嵌入到客服人员日常使用的 Salesforce 中。客服人员只需将 Case 状态修改为“Via assigned”或“Via needs input”，Agent“Via”就会自动接单并完成复杂操作；若遇到争议，Via 会直接在 Salesforce 的评论区发起人机协作（Human-in-the-Loop）。 ``` +-------------------------------------------------------------------+ | Salesforce Case 界面 | | [更改 Case 状态: "Via assigned"] -> 触发 Agent "Via" | +-------------------------------------------------------------------+ | v +-------------------------------------------------------------------+ | Invaiet Agent 框架 | | - 解析手写订单 / 复杂 PO PDF -> 转化为结构化 JSON 数据 | | - 前端模拟登录 Oracle GUI -> 启动安全 Chromium 浏览器 | | - 通过 API 批量注入订单行与数据 (混合模式) | +-------------------------------------------------------------------+ | v +-------------------------------------------------------------------+ | Salesforce Case 界面 | | - 自动回写处理结果与 Case Comments | +-------------------------------------------------------------------+ ``` - **部署成效**：项目从 Pilot 到正式上线仅用时 4 个月，预计在接下来的 3 至 4 个月内将承担每月约 50 万笔订单的处理量。 - **技术路径**：通过 GUI 进行前端身份认证与登录，进入系统后结合 Chromium 浏览器与 API 注入的方式高效处理订单。 ### [11:05](https://www.youtube.com/watch?v=b24cDV2E8I4&t=665s) 企业领导者经验（一）：亲身实践与精益流程再造 基于过去半年的实践，Chris Bradley 总结了几条核心的 AI 领导力原则： 第一，**领导者必须亲身实践**。作为 CMO 兼 AI 转型负责人，Chris 强调自己也是公司首席“前线工程师”（FDE）。因为 AI 技术每周都在变化，领导者如果不亲自写代码或深耕底层技术，就无法真正理解技术的前沿与边界。 第二，**自动化前先消除浪费（精益 Kaizen）**。企业不应盲目用 AI 去加速本就冗余的旧流程。Veritiv 规定所有拟 Agent 化的流程必须先经过 Kaizen（改善）研讨。例如在处理“客户退款与重开账单”（Credit and Rebills）流程时，团队先将原本复杂的 36 个步骤精简至 4 个步骤，再引入 Agent，成功将原本需要 7 到 15 天的处理周期缩短至 24 小时以内。 - **精益原则**：“在让 AI 执行工作之前，先问问这项工作是否应该存在。” - **实践案例**：退款流程从 36 步压缩至 4 步，处理时长从 7-15 天降至 24 小时以内。 ### [13:59](https://www.youtube.com/watch?v=b24cDV2E8I4&t=839s) 企业领导者经验（二）：无感集成与模型选择策略 第三，**融入既有工作流（Meet people where they work）**。避免为了 AI 而开发全新的独立 App，因为新软件意味着巨大的变更管理（Change Management）、培训成本和 24/7 的运维压力。将 Agent 无缝贴合在 Salesforce、Microsoft Teams 或 Slack 等现有工具中是成功落地的关键。 第四，**自研与采购动态平衡，保持模型中立**。建构框架时要保持“模型中立”（Model-Agnostic）。初期可使用前沿的商业模型（Frontier Models），但底层必须支持接入开源权重模型（Open-weights）。随着业务规模扩展到百万级订单，模型成本控制将变得至关重要。 第五，**以“单次成功成本”（Cost per success）而非“Token 成本”评估性能**。由于 Agent 的执行具有非确定性（处理同一订单可能需要 30 步到 50 步不等），评价指标应看单笔业务的综合完成成本。Veritiv 接入了 OpenRouter，通过在真实历史数据上跑 Evaluation 来测试新模型的性价比，并积极积累微调（Fine-tuning）数据集。 - **成本优化**：区分场景使用不同模型（如 OCR 使用专用模型，Computer Vision 交互使用另一类模型），着眼于综合任务单价。 - **以人为本**：AI 的定位是增效工具（Leverage）而非取代人力，旨在将员工从机械劳动中解放出来，转向提升客户服务、促进销售与增加利润等高价值工作。 ### [18:05](https://www.youtube.com/watch?v=b24cDV2E8I4&t=1085s) 总结与展望 Chris 表示，团队正在通过名为 *The AI Transformation Lab Podcast* 的播客记录并分享这趟转型的全过程。过去 6 个月的实践证明，新型 Agent 技术正在开辟全新的可能，彻底重塑企业运营模式。 ## Notable quotes > [11:05](https://www.youtube.com/watch?v=b24cDV2E8I4&t=665s) 在让 AI 执行工作之前，先问问这项工作是否应该存在。 — Chris Bradley ## Mentioned Veritiv, Chris Bradley, Ai4, Oracle, SAP, Amtech, RPA, Google, Google Anti-Gravity, AI Transformation Lab, Joe Condron, Invaiet, Via, Salesforce, Chromium, Microsoft Teams, Slack, OpenRouter, The AI Transformation Lab Podcast ## Video description Chris Bradley - Chief AI Officer, Veritiv Most enterprise AI initiatives still rely on copilots and chat interfaces while critical operational work remains trapped inside legacy systems. This session explores how Veritiv is using agentic AI and computer-use models to drive enterprise transformation by enabling AI agents to navigate enterprise software, process ERP order entry workflows, collaborate with employees in real time, and operate through existing systems like Salesforce, email, and shared inboxes. Chris Bradley will share lessons learned moving from prototypes to production, including governance, human-in-the-loop orchestration, model optimization, and scaling AI without replacing core infrastructure. Attendees will gain a practical view into the emerging enterprise operating model where humans and AI agents work together as coordinated digital teams. [Source](https://www.youtube.com/watch?v=b24cDV2E8I4)

[11:05] Chris Bradley: 在让 AI 执行工作之前，先问问这项工作是否应该存在。

[0:04] Chris Bradley: Tying Together Legacy Enterprise Systems with Computer-Use Agents [0:04] Chris Bradley: code red [0:41] Chris Bradley: Identifying High-Leverage Opportunities and Building "Invaiet" [0:41] Chris Bradley: build vs. buy [0:41] Chris Bradley: Via assigned [0:41] Chris Bradley: Via needs input [11:05] Chris Bradley: Strategic Leadership Principles for Agentic Enterprise Deployment [11:05] Chris Bradley: cost per success

[0:04] So, I wanted to talk a little bit about the legacy  systems we all have and how we've been deploying [0:11] agents as the glue to kind of tie a lot of the  complexity that we have together. Um but first [0:18] just a little bit about Veritiv. All right. So, um  Veritiv, we're a specialty packaging distributor, [0:27] uh about $8 billion, about 7,500 um employees, and  we have 120 distribution centers located all over [0:36] um the United States and Mexico. And so, we  provide packaging facility supplies to thousands [0:43] of customers. And uh we have grown over the past  years to really become this this large powerful [0:54] organization. We're the the largest of our kind  in North America. Um but we've grown through M&A. [1:00] We've grown organically. And we find ourselves in  this situation where many of you may find it where [1:06] we have like over a dozen ERPs. We have  Oracle. We got SAP. We got Amtech. We have [1:14] homegrown systems. And all of those are kind  of part of our daily life and work. And there's [1:20] a lot of manual hand-offs, duplicate  data, and really limited visibility. [1:26] And so, we've been doing automation for a  very long time. RPA using automation where we [1:36] But a lot of our systems, um  again, even when I say Oracle, [1:41] heavily configured, customized by  GSIs that are long gone. And so, [1:46] the limit of what we could really use traditional  automation tools and even what was called Agent [1:57] was limited to some edge cases and it  wasn't something that was really available [2:02] for the larger part of our enterprise. [2:06] So, last Christmas like many of you, that's when  the big aha came. For me, it was actually when [2:15] Google released Google Anti-Gravity at the end  of November. They had computer use agents that [2:22] could run browsers and they basically built this  capability that really wasn't available before. [2:29] And so, we started experimenting and looking at  it and realized these computer use agents could [2:34] patch together this complexity that we are  so desperate trying to to work through. [2:42] And so, coming back after the Christmas holiday,  I set up a meeting with our CEO and had a code [2:51] red moment that basically said the whole world  is about to change. And the things that are now [2:57] possible with Agility AI to do things that only  humans could do before was now here and we were [3:04] either going to be left behind or we were going  to be leaders in this next phase. So, launched [3:10] the AI transformation lab in February. And with  a single goal, how do we really find leverage for [3:21] our people and how do we take operational  processes like order entry, financial operations, [3:28] and how [clears throat] can we use these computer  use agents that just weren't available before [3:32] to do the things that could really provide  leverage to our company and our people. [3:38] And so, we recruited First thing I did is went out  and recruited some top talent in the organization. [3:45] Joe Condron sitting right over there was our  head of continuous improvement Lean Six Sigma. [3:52] Um cuz I if we were going to do this, we had to  go change the processes that we were going to do [3:58] to be able to get the real reap the real benefits  from it. And we and I tapped a top data scientist [4:04] so that we could look at all of our data  and figure out where are the top use cases, [4:08] how can we get the biggest bang for our buck? [4:13] So, there's a lot of things that are called agents  right now. Most of what I see are gen AI agents. [4:22] These are analyst agents or in marketing they're  like producing content or whatever. They're [4:29] but they very much aren't doing a lot of the  things that the people that I was interested [4:34] were doing. I'm looking at people that are have  three or four monitors and they're moving stuff [4:39] from one screen to another to another to another  to deal with this patchwork of complexity. [4:45] That's the type of agents that we were after. How  do we get them to navigate our legacy systems um [4:52] that don't have MCP servers connected to them uh  that aren't you know, digital first or AI first. [5:01] These have been around some of  these systems are 40, 50 years old. [5:05] So, what we looked at is I actually went  to our CHRO and I got a list of all of our [5:11] employees in the company, all 7,800 and  I sorted them by job type and job family. [5:18] And so, you come up with customer service,  we had 420 users, customer and finance, [5:24] financial operations 500. And I and the  jobs were you had a people the same type [5:34] or the same doing the same types of  things. And we really found that CX [5:39] was the big opportunity space. [5:45] And what we were after is how could we get 5 to  10x leverage on our people. So, we weren't looking [5:53] you know, to take people out, the job  apocalypse everybody is talking about, [5:57] or how can you get efficiencies. But our  our employees were actually swimming in [6:03] manual work. And that manual work  could only be done by them. But I [6:08] saw the potential that these computer use agents  could actually take that over and do it. And so, [6:17] we then looked at how do we do it? The tech  was so new, there was no platform that did [6:22] what I wanted it to do. And so, and we had  a lot of orchestration platforms in already, [6:30] but not to use these new type of models and these  new types of agents that honestly were changing [6:36] every day and every week. And I think at the  same time period, it's really where this whole [6:43] build versus buy default was destroyed. So, for  30 years, you always bought hands down, unless it [6:52] simply didn't exist and it was so strategic that  you could afford to do it, but um but we decided [6:58] to build our own custom agentic harness that  would basically pull all of these systems together [7:05] um and harden it in a way so that we could  deploy this as a mission-critical um application. [7:14] And so, [clears throat] what we um  built an application we call Invaiet, [7:19] the generative AI agent. And initially  it was built to do order entry. [7:25] We have our people manually key 1.7 million orders  a year into Oracle. And so, the prize in that [7:35] is if we can take a bunch of time out  of that, there's as much as 60,000 hours [7:41] a year of time that we could take back and  deploy on serving our customers better. And so, [7:49] one of the the key attributes was when we first  built it, we thought, "Okay, are we going to train [7:55] um our people how to use like a new system?"  That would take so much change management. [8:02] So, we looked at where they were already living.  All of our order cases were in Salesforce. [8:08] So, basically, we built an agent that  just interacted in Salesforce with um [8:14] with the customer service, and they couldn't  really tell if it was an agent or a person, [8:18] and it would pass over the order, and  do all of the things that used to take [8:25] sometimes as much as an hour to process a single  order for very large, complex ones. So, I'll show [8:32] you a quick view. So, within Salesforce, um all  we did is we basically we created a new status [8:42] for a case. And so, now it's Via assigned or  Via needs input. So, you assign the case to Via. [8:50] Um she'll know to come pick it up. If she has  a human in a loop question, she'll ask it. [8:55] It all comes back into the the case comments.  And so, the all of our users never see any of our [9:04] agents or interact with them other than right here  when they see the actual order. So, back in the [9:11] in the actual harness that we developed, it's  basically taking all the cases, pulling it in. [9:17] It could be a handwritten order like this. It  could be something that comes over clean in a um [9:24] in a in a purchase order, and it's basically all  parsed into a JSON, and then it basically gets um [9:33] sent into a we basically built a full history. So,  here I can just play any past order. So, this is [9:46] agent basically navigating through our  Oracle system through all of the steps. Now, [9:51] we couldn't use an API to do this for several  reasons because of the way it was built. So, [9:56] it uses the GUI at the beginning to get in,  and then once it's in, we basically use that [10:02] secure kind of um Chromium browser, and then  we can inject orders and lines with an API. So, [10:09] it's kind of a combination of both.  It basically process all of the orders [10:15] and then write all the comments back in in  Salesforce. And so, we took this into production [10:23] um about 4 months after the pilot  began, and uh we're on track [10:30] to be doing about a half a million orders  um here in the next uh three or four months. [10:37] And so, this is an example, I'd say, of like a  pilot to production um and we're being successful [10:44] with it, and we're driving it forward, and  now we've added in additional use cases. [10:50] Um but again, this the computer use  vision was a was a absolute game changer. [10:56] It allowed us to do things that were just  weren't feasible um through any other method. [11:05] So, I also want to share some  leadership principles, I would say, [11:11] that we've learned over this last 6 months  when we launched the lab, and we've gone from [11:16] pilot to production, successful roll out, and  now expansion. And these were kind of hard-won [11:24] rules or lessons that I think are very very  valuable for the way that we approached it. [11:31] Um the first is you cannot design a capability  into a process if you don't know the capability [11:37] exists. So, I am I am the chief marketing officer.  I also head up our AI transformation, but I'm the [11:45] lead FDE at our company. And it's because if I  don't know what's possible, and it changes every [11:53] day and every week, I can't effectively lead it.  So, I challenge everybody who is a CIO, a CAIO, [12:01] um, or leading AI transformation, if you're  not actually actively coding and using these [12:07] technologies, you're not going to actually know  exactly what the potential is. And that's been a [12:14] key component, really understanding what the  frontier looks like and how we can adopt it. [12:21] The other big thing is, and we we're definitely  a a lean focused company, which is really we [12:28] should ask if the work should exist before AI  should do it. And it's not the obvious, um, [12:36] first question because people are always saying,  "Hey, I have this process that's painful. Can we [12:41] turn AI on to it?" And the answer is always  yes, you could, and it'll make it faster, [12:50] but we really need to remove the waste before  you automate it. So, we have a rule that [12:57] we're identifying all the use cases where we think  we can get the most value and the most leverage [13:01] out of AI, but we do a Kaizen on that process  before we agentify it. And so, for example, [13:10] we just did one not too long ago for credit and  rebills. This is where you have to get issue a [13:17] credit to a customer. So, it's part finance, part,  um, order entry. Um, and we identified 36 steps [13:26] that it took to basically issue. And by the time  we were done, we were down to four. And now we [13:31] can have an AI agents basically go from an input  form and do the process. It take it down from 7 to [13:38] 15 days in some cases down to um 24 hours or  less. And so but if we would have just taken [13:46] the original process the way that it existed for  years and agentified it, we could have done it, [13:51] but we would have just sped up all of that waste.  And you really need to remove the waste first. [13:59] This is probably the key one. I would say if you  remember one thing, remember this. Meet people [14:05] where they already work. Um using when I showed  you how we use Salesforce, we use Teams or Slack. [14:14] Use the tools where people you don't have to  train people. Cuz if you go create a brand new app [14:20] and it's so easy to develop software today. But  now you're going to have change management. You're [14:25] going to have to support that app and you're going  to have train people on how to use it. And now you [14:29] need 24/7 support for that app. If you can just  have your tools kind of interact where the people [14:38] that are going to interact with these apps already  work, you can eliminate and avoid all of that. [14:46] And then build versus buy, we say is a dial not  a switch. So we're still here looking for what [14:54] we will buy. And so we will buy applications  where they absolutely make sense. Um AI agents [15:01] and and capability within software that we already  own or where we have a somewhat vanilla process, [15:08] but we also know that we're going to build  um where it absolutely makes sense and where [15:14] we can get um incredible value out of it.  But we're going to stay model agnostic. [15:20] And so one of the other we built um the harness  that we did, is we didn't want to be locked in [15:27] to any one particular [clears throat]  platform or model. And so, yes, we are using [15:33] Frontier models at the beginning, um but  we've already plugged in open weights [15:39] and other things because I don't care about cost  right now, but once we're doing a million orders, [15:45] I certainly do care. And so, making sure that  you build in that optionality for the future. [15:53] And then, the way that we measure  performance is basically cost per success. [16:01] So, when I showed you one of those orders, as we  do hundreds and hundreds of those orders, we're [16:07] actually keeping all of that data. And now, um I  connected to open router, so anytime there's a new [16:14] model that comes out, we can actually run an eval  on our own data and see exactly what the cost is [16:22] to use that model instead of the one. And  that way, again, we keep this optionality [16:27] built in and we're also saving fine-tuning  data that we'll be able to use in the future [16:32] to basically fine-tune our own models to drive  the cost even lower. But we look at the cost of [16:41] of an order. And because it is non-deterministic,  sometimes the same type of order could take 40 [16:47] turns, sometimes it'll take 50 turns or 30 turns.  Um but we're looking at what is the total cost [16:53] for each one of those orders. And we started also  breaking down all of the different components. So, [16:59] the OCR step was one one type of model. When we're  doing the computer vision, it's a different type. [17:06] And so, but we're really focused  on that, not the cost per token. [17:12] And then, we really believe that AI should  be leveraged for people, not a replacement. [17:19] There was a lot of talk about  efficiency AI in the beginning. [17:22] We see it as we're really opportunistic.  We want to repurpose the talent that we [17:29] have and the people that we have to really  focus on serving our customers better or [17:36] getting more sales, increasing velocity  of of sales, or increasing margin. [17:42] And so we're going to try to find those components  of the do this manual effort that we can have [17:48] agents support them. Instead of trying to find  agents to basically take the cream, the best jobs, [17:55] and give that to agents, we're going to try to  lift our people up and give them AI leverage. [18:05] And so um we are kind of keeping track  of our journey and we do have a podcast [18:10] called the AI Transformation Lab podcast. If  you'd like to um come along on our journey, [18:16] see more of these lessons as we continue  to drive forward. But it's been a [18:21] it's been a fun 6-month ride and we have some  really amazing things on the on the you know, [18:28] coming up and I'm just excited about this next  chapter of AI and what we can do now that we [18:35] couldn't do 6 months ago that is literally going  to change the world. So. Thank you. [applause]