Nebius, Truist Pivot to Agentic AI as Compute Costs Eclipse Payroll

算力成本超越薪资支出，Nebius与Truist转向智能体AI 随着基础设施支出超过员工薪资，企业必须绕开同质化的云工具，将智能体模型部署到低风险、高摩擦的后台工作流中，以确保即刻实现利润增长。 该专家小组探讨了企业从简单的对话式聊天机器人向高影响力智能体工作流、特定领域自动化及实体工业机器人的转变。发言人强调，组织不应因担心落后而止步不前，建议首先在文档审计、供应链预测和后台财务等低风险、高摩擦的运营“安全区”进行部署。为了应对高昂的算力成本以及大型云厂商带来的平台同质化，企业被敦促寻求联合创新伙伴关系，并构建专用的应用层解决方案。归根结底，成功落地的关键在于将严格的租户级数据治理与积极的员工赋能相结合，把 AI 视为运营倍增器，而非削减员工的工具。 企业级 AI 的演进重心已从初期的对话式 Chatbot 快速转向智能体工作流（Agentic Workflows）与具身物理机器人；同时不断攀升的算力基础设施成本有时甚至超过人力薪酬，推动企业加速产品上市与推理商业化变现。 (0:04) 当前企业采用 AI 的核心瓶颈主要在于组织内部的接纳度、企业文化犹豫与数据安全考量，而非底层技术能力；高管应以结构化实验和明确的风险框架取代“唯恐落后”的焦虑情绪。 (3:31) 智能体自动化的理想切入点是高业务量、规则明确且操作摩擦大，但在合规方面无重大负面风险的流程；受强监管行业应优先在呼叫中心实时 Copilot 等运营“安全区”落地，而非直接介入高风险自主决策。 (7:54) 将企业内部领域数据与垂直 AI 初创团队结合的联合创新模式，可将项目原型周期从数年压缩至几个月；相较于基础的内部效率提升，高管与董事会更看重能将核心营收引擎毛利率提升 25% 至 30% 的 AI 项目。 (11:34)

Navigating enterprise artificial intelligence requires moving past early conversational experiments into targeted agentic workflows and specialized application layers. Panelists across banking compliance, industrial manufacturing, analytics, and cloud infrastructure argue that sustainable returns come from deploying low-risk operational automations, establishing secure data guardrails, and augmenting existing revenue engines. Targeting Low-Risk, High-Friction Operational Safe Zones (6:21) Enterprise leaders often fear they are falling behind, yet Truist Chief Privacy Officer Ron Whitworth observes that peer benchmarking across top-10 banks reveals nearly every organization feels lagging. Rather than deploying AI to make consequential automated decisions—such as credit underwriting or lending approvals, which carry heavy regulatory exposure—institutions succeed by targeting operational safe zones like call center guidance, complaint handling, and policy retrieval. Deploying Agentic Systems in Back-Office and Supply Workflows (7:54) Tavant's Paul Jeyasingh recommends focusing on rules-heavy, manual processes with low existential downside, such as auditing loan servicing lifecycles and reconciling complex agricultural invoices against purchase orders. In one servicing audit deployment, autonomous agents expanded coverage beyond an offshore team of ten that previously reviewed only 10% of the loan portfolio. Consolidating Fragmented Tooling in Industrial Operations (35:59) In manufacturing and logistics, AI integration layers replace fragmented software stacks. International scaled from zero to roughly 50 AI initiatives in under a year, capturing returns from six-figure savings up to eight-figure cost reductions. By unifying five or six disparate platforms—including SAP and internal scheduling software—into a single interface, planners generated rolling three-month demand forecasts and aligned order schedules in 90 days rather than the traditional two-year cycle. Navigating Infrastructure Bottlenecks and Hyperscaler Dominance (2:00) Nebius VP Shane Zide notes that high-density AI infrastructure costs can exceed payroll, driven by physical constraints as legacy enterprise colocation facilities lack the electrical power and cooling required for modern GPU clusters. This capital squeeze forces organizations into specialized rented cloud capacity to shorten time-to-market. Concurrently, Paul Jeyasingh warns that major cloud hyperscalers rapidly commoditize horizontal tooling—such as basic retrieval-augmented generation chatbots—meaning builders must focus strictly on specialized, vertical enterprise applications to stay viable. Securing Tenants to Unlock Internal Workforce Productivity (18:44) Employee adoption requires establishing strict data tenant isolation and role permissions to eliminate fears of intellectual property leakage or accidental internal exposure. Once governance guardrails are settled, internal deployments—such as automated HR leave processing or legal policy synthesis—yield measurable efficiency, freeing up estimated weekly savings of approximately 25 hours per employee. The speakers stress framing adoption around eliminating administrative toil and accelerating delivery rather than cutting headcount. While AI tooling democratizes rapid prototyping and compresses planning cycles—such as reducing fleet corridor electrification studies from 18 months to under eight—long-term enterprise viability remains tethered to rigorous data validation, specialized domain applications, and cultural education across the workforce.

行业专家在Ai4研讨会上指出，企业无需因技术浪潮产生落后焦虑，而应从低合规风险与高操作摩擦的业务场景切入，依托协同创新和垂直应用避开平台层同质化竞争，并通过严格的数据治理释放员工的生产力红利。 破除落后焦虑与锚定高摩擦合规安全区 (6:21) 针对广泛存在的落后焦虑，Truist银行首席隐私官Ron Whitworth指出，即使是全美前十的大型银行在闭门交流中也普遍认为自身进度滞后。企业不应被行业炒作所瘫痪，而应将重心转向平衡业务机遇与成熟的风险框架。Tavant人工智能副总裁Paul Jeyasingh建议，企业初期的切入点应避开信贷审批、个性化定价等强监管、高后果的自动化决策“雷区”，转向运营摩擦高但合规风险可控的“安全区”。 以金融与供应链实际场景为例，某大型贷款服务银行此前由10人组成的离岸团队受限于人力只能手动抽检10%的贷款资产包，通过部署智能体（Agentic）审计流程，在低合规风险的前提下实现了审计覆盖面的倍增；类似地，大型跨国农业企业利用智能体自动化核对数千份PDF发票与ERP采购记录，极大消除了规则繁复场景下的人工匹配错误。云服务巨头提供的基础设施已大幅降低部署门槛，关键在于率先选定具有显著降本增效潜力的非核心业务试验田。 算力成本激增与规避云巨头同质化碾压 (2:00) Nebius美洲销售副总裁Shane Zide强调，当前AI算力支出规模极为庞大，在部分企业中计算基础设施的开支甚至已经超过了人员薪资总和。现代高密度GPU集群对电力供应与散热系统有着严苛限制，传统企业托管数据中心根本无力承载，迫使企业放弃造价高达2000万美元的自建方案，转而寻求专业GPU云托管服务，以加速从模型训练到推理变现的周期。 在应用层竞争方面，Paul Jeyasingh根据自身创业经历提出预警：云巨头（如AWS、GCP、Azure）拥有庞大的研发实力，两年前初创团队研发的通用型检索增强生成（RAG）聊天机器人，仅仅三四个月后便被各大云厂商转化为开箱即用的原生平台服务。他警告企业与创业者切勿重复构建基础通用中间件，而必须扎根于云厂商难以覆盖的垂直领域，专注研发具有特定行业壁垒的应用层产品。 协同创新加速与重塑传统工业工作流 (11:34) 大众集团旗下International新兴技术负责人Srini Gowda提出“合作即新领导力”模式。传统内部研发往往需要24至36个月，企业通过将自身业务数据与外部垂直AI初创企业的算法结合，能够在6个月内将可行性原型直接交付给百事、沃尔玛等关键客户进行实地测试。在车队电动化规划方面，过去依靠咨询机构耗资数百万美元、历时12至18个月的跨州电网与线路排布方案，借助智能体模型仅用6至8个月便能以更低成本完成交付。 在工业制造内部运营中，AI整合展现了显著的降本效应。Srini Gowda表示，International在一年内将AI项目从零扩展至近50个，带来数十万至数千万美元的财务效益。面对企业内部ERP系统（如SAP）与各类遗留平台割裂的“工具蔓延”痛点，团队构建了统一的AI集成层，将原本需要两年的集成周期压缩至90天，实现了三月期供需预测与装配线生产排程的实时协同。他同时指出，在此类高自动化转型中，建立全程严格的数据验证机制是确保系统可靠性的首要条件。 确立租户安全边界与释放全员生产力 (18:44) 企业推进全员采纳AI的前提是建立清晰的数据治理与安全边界。Ron Whitworth与Paul Jeyasingh指出，员工对数据外泄和知识产权受损的担忧是主要阻力，必须在企业内部租户内建立严格的访问控制，防止类似自定义GPT错误共享导致的数据越权。一旦安全边界稳固，组织就能迎来实质性的生产力红利——数据显示内部定制化工作流平均每周可为每位员工节省约25小时工时。 在职能部门的日常运作中，AI正在重塑业务流程。Shane Zide介绍，销售团队通过AI提取目标客户画像与采购意图，将原本需要6周的线索筛选缩短至数天，使销售团队摆脱了依赖超额招聘的传统粗放模式；Srini Gowda展示了耗时半年上线的HR智能体，员工单句指令即可联动休假额度扣除、日程冻结与团队通报；Ron Whitworth亦证实，法务合规部门能利用AI助手秒级检索历史会议记录与规章文档。随着辅助编程技术的普及，员工日常工作模式正逐步转向“在模型交付成果的基础上进行二次深加工”。 研讨会总结指出，企业AI落地的核心并不在于自行烧钱训练基础模型，而在于结合自身高摩擦业务场景，利用既有算力与云生态开展为期8至12周的概念验证。在此过程中，破除组织内部的职业替代恐惧、确立安全合规护栏以及保持对AI输出结果的持续校验，是企业从试验性探索顺利迈向规模化商业变现的关键保障。

## Background 本内容由 Ai4 于 2025 年 8 月 24 日发布在 YouTube 平台。根据页面简介，这是一场专为 AI 初学者打造的圆桌研讨会，汇集了来自金融合规、制造业、算力基础设施及技术咨询领域的资深专家，旨在提供一套涵盖基础概念、常见商业用例、潜在风险与价值落地的全方位企业 AI 入门指南（The AI Beginner’s Playbook）。 ## TL;DR 圆桌讨论围绕企业开启 AI 旅程的实操路径展开，指出行业焦点已从早期的问答机器人转向具备自主行动力的 AI 智能体（Agentic AI）。专家们建议初学者应优先从高摩擦、低试错风险的内部场景（如发票与合同核对、合规审计、请假审批）切入，再逐步延伸至供应链排产与核心业务增收。同时，嘉宾们强调必须在企业专属租户内筑牢数据安全护栏以缓解员工的职业焦虑，并通过与云厂商生态共创来大幅缩短研发周期，最终让 AI 充当跨部门知识整合的统一管理视窗。 ## Atmosphere & style 整场圆桌务实坦诚且充满启发性，彻底打破了行业内常见的虚浮炒作与“落后焦虑”。来自不同领域的嘉宾在轻松融洽的互动中频频抛出极具温度的洞见，既有大银行首席隐私官自嘲庆幸逃过律所底层卷工的幽默风趣，也有算力高管直面物理功耗瓶颈与成本账本的冷静犀利。观众能真切感受到一种从容不迫的赋能感——只要守住合规护栏、保持对未知的好奇心，“人类工作始于 LLM 结束之处”便不再是一句口号，而是一张人人皆可付诸实践的行动图纸。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=4s) 生成式 AI 正从单纯的问答聊天机器人（Chatbots）全面转向具备执行能力的自主智能体（Agentic AI），成为企业落地的核心方向。 - [01:31](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=91s) 随着基础模型研发与高密度推理需求的激增，AI 基础设施和 GPU 算力开销在某些企业中甚至超过了员工薪酬总额。 - [03:16](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=196s) 传统工业与制造业落地 AI 的最大阻力在于团队对数据外泄及知识产权流失的信任顾虑，企业需通过“运行、成长、转型”三阶段框架逐步推进。 - [05:49](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=349s) 整个行业目前普遍存在“落后焦虑”，即使是全美前十大银行也觉得自己发展滞后，初学者应摆脱焦虑并专注从先行者的经验教训中学习。 - [07:32](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=452s) Agentic AI 理想的切入点应具备“高业务摩擦、繁复重复劳动且即使出错也不会产生致命灾难”的特征，如贷款合规审计和跨文档发票校验。 - [11:12](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=672s) 生态协同合作已成为新领导力，企业通过共享数据与联合 IP 协同创新，能将原先 24-36 个月的内部研发周期压缩至 6 个月内交付可用原型。 - [14:12](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=852s) 要打动董事会与高管层，AI 战略不能仅停留在零散的内部提效，而应直击盈利核心，将成熟赚钱业务的利润率实质性提升 25% 至 30%。 - [15:23](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=923s) 强监管金融行业在涉及自动化贷款等敏感决策上保持审慎，现阶段落地聚焦于客服实时话术辅助、投诉政策匹配等“安全区”内的低垂果实。 - [18:44](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=1124s) 在专属租户（Tenant）内建立严格的数据隔离护栏是消除员工顾虑的关键，成熟落地后单名员工每周平均可节省约 25 小时工作时间。 - [19:58](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=1198s) 企业应当确立“人类工作始于 LLM 能力终点”的职场心智，引导员工将 AI 视为放大产出的杠杆而非裁员威胁。 - [21:14](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=1274s) AI 重塑了销售组织模式，背调周期从 6 周压缩至数天，促使销售团队摆脱粗放的“广撒网式过度招聘”，转向依靠高人效产出的精益模式。 - [29:32](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=1772s) 现代 GPU 服务器的极端功耗与散热限制迫使大型企业放弃新建本地机房，转而借助专业 GPU 云算力平台加速商业化交付。 - [32:14](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=1934s) 超大规模云厂商将 RAG 等通用能力快速平台化，迫使 AI 创业团队必须放弃通用基础设施竞争，深耕特定垂直行业的应用层场景壁垒。 - [34:25](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=2065s) AI 辅助编码工具（如 Claude Code）颠覆了原型研发速度，非技术背景人员也能在极短时间内独立交付可调 API 的可视化界面。 - [39:50](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=2390s) AI 在供应链排产、法律合规检索及财务关账中充当打破数据孤岛的“统一玻璃视窗”（Single Pane of Glass），实现企业全景业务的可视化穿透。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=4s) 圆桌嘉宾背景介绍与 AI 演化阶段 本场圆桌由 CB Insights 洞察主管 Jason Saltzman 主持，汇聚了来自技术咨询、金融合规、工业制造及算力基础设施领域的四位资深从业者，共同探讨企业开启 AI 之旅的实践路径： - **Paul Jeyasingh**（Tavant 人工智能与分析副总裁）：拥有超过 15 年机器学习经验，曾在 AWS、IBM 和 Oracle 任职。他指出过去三年来生成式 AI（GenAI）发展迅猛，前两年半全行业都在聚焦构建问答聊天机器人（Chatbots），而当下技术重心已全面转向 AI 智能体（Agents）。 - **Ron Whitworth**（Truist 首席隐私官）：来自总部位于北卡罗来纳州夏洛特的美国大型银行 Truist，负责领导全行的隐私合规、技术合规以及新设立的 AI 合规治理体系。 - **Srini Gowda**（International 新兴技术与自动驾驶副总裁）：所属的 International 为大众汽车集团（Volkswagen Group）旗下品牌。自 2017 年起全职深耕 AI 与自动驾驶技术。 - **Shane Zide**（Nebius 美洲区销售副总裁）：Nebius 是全球领先的 NVIDIA GPU 云算力托管服务商之一，线上投产运行的 GPU 超过 30,000 张，并正在部署下一代 Blackwell 架构芯片。Shane 形象地表示，无论企业的 AI 规划从何处起步，当进入最终的模型训练与推理环节时，都需要算力基础设施的支撑。 ```mermaid flowchart TD accTitle: AI 技术演进与落地生态分工 subgraph S1["技术与应用演进"] A["传统机器学习 (ML)"] --> B["生成式问答机器人 (Chatbots)"] B --> C["自主智能体 (Agentic AI)"] end subgraph S2["企业落地支持体系"] D["场景落地与实施 (Tavant)"] E["隐私安全与合规治理 (Truist)"] F["工业与供应链实操 (International)"] G["大规模 GPU 算力底座 (Nebius)"] end S1 -.-> S2 ``` --- ### [01:31](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=91s) 企业级 AI 采用现状与算力基础设施成本 Jason Saltzman 在开场时指出，继互联网热潮（Dot-com boom）和云计算热潮（Cloud boom）之后，全行业迎来了有史以来关注度最集中的时刻，所有企业高层都在紧盯“AI”这两个字母。Shane Zide 从算力供给侧分享了当前的落地格局，指出目前采用最活跃的两大群体分别是“AI 原生初创公司”以及具备深厚机器学习团队的“AI 成熟企业”（例如 Zoom、ServiceNow、Airbnb 等）。此外，市场上不断涌现由前 Meta 工程师创立、刚成立数分钟即斩获 5000 万美元融资的基础模型初创公司。 这种高密度的模型研发直接推高了基建成本。Shane 特别强调，在某些极端场景下，AI 基础设施的开销甚至已经超越了员工的薪酬总额（payroll）。因此，无论是跨国巨头还是初创团队，在启动项目前都必须明确自身目标，借助高性能 GPU 加速训练与推理，以远快于通用算力的效率推进产品商业化变现。 --- ### [03:16](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=196s) 制造业供应链落地痛点与“运行-成长-转型”推进模型 Srini Gowda 结合 International 的实践，阐述了传统工业制造企业面临的紧迫感与实际进展。在年初时，公司内部的 AI 项目数量为零，而目前已有近 50 个推进中的项目，其预期财务回报跨度极大，部分高价值项目有望带来高达七至八位数的千百万美元级效益，而基础效率型项目回报则在六位数左右。应用场景涵盖了供应链管理、智能制造、人力资源及全员通用生产力工具。 然而，将 AI 落地到具有 200 年历史的传统企业并非易事，最大的瓶颈在于团队的接纳度与信任度。Srini 举例说明了制造业零部件规划与库存管理的严峻现实： ```mermaid flowchart LR accTitle: 制造业零部件管理痛点与 AI 介入 subgraph PainPoints["传统流程痛点"] P1["供应商款项重复支付"] P2["仓库备件无法精准调度"] P3["设备维护停机延误"] end subgraph AIChecks["AI 落地顾虑"] C1["数据流向何处？"] C2["是否脱离租户 Tenant？"] C3["核心知识产权 IP 泄露？"] end subgraph Target["目标形态"] T1["业务效率提升"] T2["具身物理 AI 协作"] end PainPoints --> AIChecks --> Target ``` 面对员工对“数据外泄、丢失核心知识产权”的焦虑，企业必须依赖长期的内部培训与教育。Srini 提出了企业 AI 落地的“运行、成长、转型”（Run, Grow, Transform）三阶段框架，并指出公司目前正处于扎实的“运行”（Run）阶段。此外，物理 AI（Physical AI）与人形机器人的应用也已近在眼前，在制造车间引入 6 英尺高的机器人旨在辅助现有劳动力而非取而代之，如何引导一线工人和团队接纳这一变革是当下培训的核心议题。 --- ### [05:49](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=349s) 克服“落后焦虑”：初学者的组织心智与行业基准 针对处于“预运行阶段”（Pre-run phase）的初学者，Ron Whitworth 结合其连续三年参与 Ai4 大会的经历指出，会议规模逐年翻倍的现象印证了 AI 浪潮的爆发。两年前，行业探讨的焦点仍是“AI 旅程启动时到底该让哪些部门坐到会议桌前”，如今整个行业已经形成共识——AI 绝非投机炒作，而是正在眼前重塑世界。 谈及企业的心理状态，Ron 给出了极具共鸣的观察： > “我和所有人交流时，大家几乎都觉得自己已经落后了。但觉得落后是完全正常的。即便在全美前十大银行的同行基准对标会议上，大家坐在一起讨论用例和组织架构时，所有人依然觉得自己远远落后了。” Ron 强调，企业无需陷入焦虑，而是应当认清摆在面前的巨大机遇与潜在风险，积极从过去两三年中先行部署的大型企业身上汲取成功与失败的经验教训。 --- ### [07:32](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=452s) 智能体 AI（Agentic AI）的切入点：审计与发票匹配的快速落地 Paul Jeyasingh 从一线技术交付的视角分享了如何跨越“低垂的果实”（Low-hanging fruit）去挖掘高价值业务场景。他每周都会与两到三家企业客户深入沟通 Agentic AI 的应用方案，并给出了明确的初学者建议：**优先选择业务影响范围可控、风险相对隔离的非核心主干环节切入**。 | 行业领域 | 实际案例客户 | Agentic AI 落地场景与业务逻辑 | |:--- |:--- |:--- | | **金融与银行业** | 美国大型贷款服务银行（Loan Servicing Bank） | 部署 Agentic AI 自动执行贷款服务全流程的**合规审计**，在不直接触碰资金流的前提下建立合规防护网。 | | **农业与制造业** | 全球顶尖大型农业跨国企业 | 解决海量非结构化 **PDF 发票处理难题**，利用智能体自动提取数据，并与采购订单（PO）及合同条款进行智能比对，快速排查异常与漏洞。 | Paul 指出，这类涉及跨文档比对、核查与规则校验的高频重复场景，不仅试错风险较低，而且能够立竿见影地体现业务回报，是初涉 Agentic AI 的企业最理想的切入战场。 ### [09:23](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=563s) 寻找适合 Agent 与生成式 AI 的高摩擦业务场景 在切入人工智能应用时，明智的做法是选择那些即使出现小差错也不会对业务造成灾难性负面影响的场景作为起点。技术在当今时代往往反而是最容易解决的一环，只要能找到符合合规与隐私要求的正确商业用例，AWS、GCP 以及 Azure 等主流云厂商早已铺平了落地路径，甚至会主动指派解决方案架构师（Solutions Architect）提供实操协助。然而，核心挑战在于如何识别哪些业务流程最契合 Agentic AI（智能体 AI）、生成式 AI 或传统的智能自动化技术。 对于这一技术选型问题，破局的关键在于审视企业内部痛感最强、阻力最大的环节。以贷款服务审计（loan servicing audit）为例，某企业过去雇佣了 10 名离岸员工专职审核贷款流程，但在巨大的工作量之下，人工处理能力只能覆盖全部贷款子集的 10%。对于这种“高摩擦、工作繁重且极易出现人为失误”的业务节点，引入 Agentic 解决方案正是立竿见影的黄金切入点。 ```mermaid flowchart TD accTitle: Agentic AI 落地选型流程 subgraph S1["痛点识别与评估"] A["业务高摩擦与高耗时"] --> B["高失误率与覆盖率受限<br/>例如仅能覆盖10%贷款"] B --> C{"是否影响核心底线安全?"} end subgraph S2["技术选型与实现"] C -- "安全/低风险场景" --> D["引入 Agentic AI / 生成式 AI"] C -- "高风险场景" --> E["增加人工审核与合规护栏"] D --> F["借助云厂商架构师协同落地<br/>AWS / GCP / Azure"] end ``` - **云厂商生态支持**：主流超大规模云服务商（Hyperscalers）提供了高度集成的实施工具和架构师支持，大幅降低了底层工程门槛。 - **Agentic 适用特征**：高摩擦力（high friction）、重复繁琐劳动、人工覆盖率受限（如 10 人仅能覆盖 10% 业务量）、存在显著人为错误风险的业务场景。 --- ### [11:12](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=672s) 外部共创与内部供应链：生态合作成为新领导力 在快速演进的 AI 格局中，"合作即是当下的新领导力"（partnership is the new leadership）。单靠企业内部团队及其现有技能储备，能做的事情极其有限；而外部 AI 初创公司正呈现爆发式增长。企业需要建立一种以结果为导向的分类评估机制，考量项目能否在既定时间表与预算内交付正向投资回报率（ROI）。通过联合知识产权（joint IP）和数据共享的“协同创新模式”（co-innovation model），企业提供实际业务数据与应用场景，初创公司带来前沿技术与专业人才。 这种外部共创模式颠覆了传统的研发周期。如果完全依靠内部团队自研，往往需要 24 到 36 个月才能交付并实现正向 ROI；而借助外部协同，团队能在 6 个月内将可用原型交付到百事可乐（Pepsi）、沃尔玛（Walmart）、Penske、Ryder 等大型车队客户手中进行实地测试与反馈迭代。在内部应用上，AI 被深度嵌入到制造与供应链体系中：面对俄乌冲突对欧洲供应链基地造成的严重冲击，企业利用 AI 预测极端地缘事件对物流循环路线（milk run）的影响，并精细测算关税政策波动对物料清单（BOM）端到端成本的冲击，从而决定成本吸收策略或客户转嫁方案。 - **交付周期对比**：传统纯内部研发需 24 至 36 个月才能见效，协同创新模式在 6 个月即可产出供标杆客户试用的工作原型。 - **标杆客户试点**：涵盖 Pepsi、Walmart、Penske、Ryder 等大型车队与物流巨头。 - **供应链风险量化**：利用 AI 评估欧洲地缘战争冲击、优化零件循环取货路线，以及动态测算关税对整体物料清单成本（BOM cost）的影响。 --- ### [14:12](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=852s) 直击核心营收：为现有产品提效增收 如果想要真正打动企业高管与董事会，就不能仅仅局限于内部微小的效率优化。帮助内部员工提升 21% 的工作效率固然是“锦上添花”（nice to have），但远不如直接作用于盈利核心来得震撼。最具说服力的方式是找到当前已经在赚钱的产品或服务，通过 AI 技术将其利润率提升 25% 至 30%，帮助客户更快速地使用产品，从而为企业创造更多增量收入。 与会将技术赋能业务的实践案例进行了具体拆解：Cloudflare 在其 Worker ID 等产品底层接入高性能 GPU 计算资源，从而让远程办公人员能够更快、更高效地调度内部资源；Shopify 则在电子商务平台底层引入 GPU 算力，显著优化了购物车结账体验。这些原本运行在普通数据中心通用算力上的成熟产品，在 AI 算力加持后大幅提速，直接带动了商业变现能力的阶梯式跃升。 - **高管关注焦点**：相较于内部零散的生产力提升，董事会和 C-level 更看重对既有核心业务利润率提升 25%-30% 的实质贡献。 - **Cloudflare 案例**：利用 GPU 算力支撑 Worker ID 等产品，提升远程协作成效与响应速度。 - **Shopify 案例**：将底层算力迁移至 GPU，大幅提升电商购物车响应体验并直接拉动结算转化。 --- ### [15:23](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=923s) 金融合规护栏与低垂果实：在安全区内释放效能 在银行业等强监管行业，AI 落地的一切探讨都离不开“视情况而定”（it depends）的前提。银行面临着极其严格的监管合规审视，任何可能直接对个人产生重大决策影响的 AI 应用——例如自动化贷款审批或差别化利率厘定——都必须保持高度谨慎，并在极度严格的安全护栏建立完备后才能推进。因此，银行现阶段的落地重点明确集中在“安全区”内的低垂果实（low-hanging fruit）。 呼叫中心、分支网点助手、投诉管理和政策指引等场景是典型的低风险、高收益切入点。过去，网点或客服人员面对复杂客户诉求时，需要耗时翻找内部冗长的合规规程；而如今通过集成式 AI 实时指引（real-time guidance），系统在对话过程中无缝推送操作建议与辅助工具，既提升了客户体验与运营效率，又规避了直接替人类做决定的法律风险。随着超大规模云厂商推出开箱即用（out-of-the-box）及可定制的 Agent 护栏方案，合规门槛已被大幅降低。要推动内部团队深度采纳这些工具，关键在于自上而下清晰传达端到端战略，将 AI 应用与公司在短、中、长期的 ROI 目标进行精准对齐，并彻底厘清数据安全与权限归属。 - **强监管红线与安全区**：暂不涉足直接决定个人贷款审批或定价的敏感算法，聚焦于客服指引、分支机构实时话术支持与投诉政策匹配。 - **合规护栏实践**：主流云厂商提供集成于 Agent 方案内的现成护栏，企业亦可结合特定合规要求构建定制护栏。 - **内部采纳核心驱动力**：战略目标与短中长期 ROI 强挂钩，并在数据所有权与合规治理上给予员工清晰透明的指引。 ### [18:44](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=1124s) 数据安全护栏与 Custom GPTs 的内部落地收益 在企业内部推行生成式 AI 和智能体工作流（Agentic Flows）时，员工采用率的核心瓶颈在于数据安全感。讨论指出，必须在企业专属租户（Tenant）内建立严格的合规护栏，确保内部数据绝不会外流，员工才能放下戒备、放心地向系统共享敏感数据集。以 Custom GPTs（定制化 GPT）为例，其早期推广遇到了不少波折：员工常有“手滑”（butter fingers）误操作的情况，在发布定制应用时不小心将权限共享给了不该访问的同事，导致信息越权暴露。因此，在技术落地初期，围绕权限配置与数据隔离的充分沟通必不可少。 克服了这些安全初期的阵痛后，AI 带来的投资回报率（ROI）十分惊人。在实际统计中，每位员工每周平均能节省约 25 个小时的工作时间。当这种效率提升按月度和年度在整个组织内放大时，释放出的生产力呈几何级增长——员工可以将原本消耗在重复性琐事上的时间，投入到对前沿趋势的深度调研以及思考如何为客户创造更多商业价值上。 - **核心落地前提**：在专属租户内构筑安全护栏，杜绝数据外溢，建立员工对数据隔离的信任。 - **早期风险场景**：Custom GPTs 发布时的权限配置误操作（如误将受限数据暴露给非相关人员）。 - **量化生产力提升**：单名员工每周平均节省约 25 小时，极大释放了战略规划与高价值客户服务的时间。 ```mermaid flowchart TD accTitle: 建立企业内部 AI 采用的信任闭环 subgraph Security ["数据合规与护栏"] A["专属租户隔离"] --> B["防范权限误配"] B --> C["建立员工数据安全感"] end subgraph Value ["价值实现"] C --> D["Custom GPTs / 智能体落地"] D --> E["单人每周节省 ~25 小时"] E --> F["专注深度调研与客户增值"] end ``` --- ### [19:58](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=1198s) 克服被替代恐惧：“人类工作始于 LLM 结束之处” 针对“如何引导员工适应新角色”的提问，讨论聚焦于员工普遍存在的职业危机感。许多人觉得 AI 正在威胁自己的饭碗，但企业必须向员工明确传递一个理念：AI 的目的不是裁员替代，而是赋能提效，正如内部常说的那句话——“我们的工作始于大语言模型（LLM）能力的终点（our job starts where an LLM stops）”。在当前阶段，成功运用 AI 的企业将赢得市场，而行动迟缓者则会迅速衰退。 管理者需要通过提供安全可控的工具，引导员工思考如何用 AI 优化个人工作流。一旦员工不再将 AI 视为威胁，而是认识到它能成倍放大个人产出与客户服务响应速度，积极性就会被彻底激活，形成自发探索和技能提升的良性循环。 --- ### [21:14](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=1274s) 销售效能变革：从 6 周到数天的高精画像与精益团队 AI 正在彻底重塑现代销售组织的运作逻辑。在过去的“电话黄页”时代，销售线索挖掘和背景调研极其耗时；而如今，销售人员可以在数分钟内完成 500 个目标客户（Target Accounts）的深度分析，不仅能迅速锁定理想客户画像（ICP）的关键决策人，还能精准挖掘出潜在项目意向——例如识别哪些客户当前正有采购 GPU 服务器的活跃需求。以往长达 6 周的繁重背调工作，如今在几天之内就能全部拆解完成并直接进入触达阶段。 这种效率跃升深刻改变了销售团队的组织与招聘模式： - **传统销售模式**：依赖“广撒网”式的过度招聘（Overhiring），寄希望于大约 60% 的销售人员能完成业绩指标（Quota），资源配置犹如“在黑暗中盲目掷镖”。 - **AI 赋能下的精益模式**：单个销售人员的产出效能（Yield）大幅提升，企业无需庞大人海战术即可达成更高业绩，实现了销售规划与资源配置的精准化和精益化。 --- ### [23:20](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=1400s) 跨行业颠覆：智能体票务与智能客户细分 AI 智能体的应用正在跨越传统界限，重塑各大行业的商业模式与运营链路。现场分享了两个生动的企业案例： 1. **体育票务与消费体验重塑**：一家收购低级别棒球小联盟（Minor League Baseball）球队的私募股权机构，正在利用智能体技术打造专属交互式售票系统。该系统直接与球迷对话，理解其个性化需求并提供优惠激励，允许球迷直接购票并打包周边商品及餐饮，彻底绕开 Ticketmaster 等第三方售票平台高达 50% 的高昂便利服务费（Convenience Fees）。 2. **重工业与农业机械的精准营销**：某大型日本农机巨头（在美业务规模庞大）借助 AI 技术重构了客户细分（Customer Segmentation）与个性化营销流程。以往需要耗费数周的数据分析和邮件营销工作，现在只需选定客户子集，即可由 AI 自动生成定制营销内容并完成后续效果追踪，执行周期从数周压缩至极短时间。 --- ### [25:11](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=1511s) 个性化、主动性与重型车队电气化布局 在财富管理与专业咨询领域，AI 正在推动服务模式从“被动响应”向“极致个性化与主动出击（Proactivity）”转变。顾问不再需要耗费数天进行手动数据挖掘，AI 能够快速合成海量信息并输出 3 到 4 种高度精准的方案选项，让人类专家的精力完全聚焦于高价值的情感沟通与深度决策。销售和咨询团队在与客户进行第一次接触（First Touch）之前，就已经掌握了全量背景画像与定制建议。 在交通运输行业，AI 也展现出了强大的空间与基建规划能力。以中型和重型卡车运输业（Medium and Heavy Duty Trucking）的电气化转型为例，充电网络和基础设施建设一直是核心痛点。团队通过 AI 算法结合客户的车队运行足迹数据（Footprint Data）、行驶路线、停靠频次以及电网负荷影响（Grid Impact），精准推算出充电站的最佳选址布局，使企业在与客户接洽时能够直接提供具备高度实操价值的数据驱动方案。 ### [28:00](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=1680s) 借助智能体与多源数据整合重塑客户交付 在汽车与交通电气化领域，将海量分散的数据集整合进 AI 驱动的模型中所释放的效能是颠覆性的。传统模式下，规划一条走廊的电气化方案（例如评估得克萨斯州或全美其他区域的充电与运力走廊）极度依赖外部咨询公司，不仅需要投入数百万美元，整个交付周期往往长达 12 到 18 个月。然而借助当下的 AI 技术，这一过程被大幅缩减至 6 到 8 个月，同时成本也呈断崖式下降。 更关键的转变在于与客户的协同模式：企业不再处于“信息孤岛”中闭门造车，而是直接把客户及其业务数据拉到同一张桌子上进行实时推演。当客户提出特定的车辆配置构建规范（Build Spec）或个性化运营诉求时，传统流程中冗长繁琐的 Excel 表格流转、PPT 汇报和多层审批已被彻底淘汰。取而代之的是高效的智能体模型（Agentic Model）——业务人员直接输入原始参数，AI 即可在瞬间生成极其精准的最优解。客户对这种能够“秒级”解答疑问、迅速给出解决方案的敏捷度感到非常惊艳，这也正是企业当前全力推进的核心方向。 ```mermaid flowchart TD accTitle: 传统咨询与AI驱动电气化规划流程对比 subgraph Traditional["传统咨询模式"] A["高额咨询团队 (数百万美元)"] --> B["多重 Excel / PPT / 流程审批"] B --> C["漫长交付周期 (12-18 个月)"] end subgraph Agentic["现代 AI 智能体模式"] D["客户业务数据 + 多源数据集"] --> E["Agentic 智能体模型解析"] E --> F["敏捷方案输出 (6-8 个月 / 成本骤减)"] end ``` ### [29:32](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=1772s) GPU 算力租赁的爆发与 AI 落地多元生态 随着企业从 AI 初学阶段向复杂的高阶应用场景演进，计算与 GPU 算力资源的需求呈现爆发式增长。甚至连制造 AI 革命核心服务器的英伟达（NVIDIA）本身，内部也有庞大的研究团队在全天候运行自主训练与推理，利用 AI 优化自身产品，形成技术反哺。整个市场的客户画像呈现出极为宽广的光谱： - 一端是位于帕洛阿尔托（Palo Alto）斯坦福大学宿舍里的两位大二学生，刚刚拿到 1000 万美元种子轮融资，就已经在宿舍里租用算力训练自己的基础大模型； - 另一端则是体量庞大的《财富》500 强跨国集团，争分夺秒地抢占市场先机（Speed to Market）。 当前推动企业转向 GPU 算力租赁市场（如 Nebius 等云算力平台）的根本痛点在于**物理空间与功耗限制**。现代 GPU 服务器的功率与散热需求极其夸张，企业 IT 负责人根本不可能直接向高管申请“为了做 AI 先花 2000 万美元新建一个数据中心托管机房（Colocation）”。这迫使原本坚持本地部署（On-Premises）的大型企业改变策略，将已经验证能带来实际商业营收的场景剥离出来，寻找具备完整合规资质与严密数据隐私保护的专业云算力提供商。 在应用前沿，除了从 Perplexity 或 OpenAI 获取 API 调用的初阶用法外，更多前沿企业正在展现惊人的生产力： - 剑桥市（Cambridge, Massachusetts）的机器人初创公司通过 GPU 服务器输入一行指令，机器人便能即时执行复杂动作；更惊人的是，机器人甚至开始根据既往学习自主探索并执行此前未编排的动作。 - 医疗与生命科学领域的初创企业，医生正在借助大规模模型加速新药研发与临床数据洞察。 ### [32:14](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=1934s) 超大规模云厂商的生态挤压与创业公司的应用层生存空间 对于从未涉足过 AI 项目的企业而言，与专业系统集成及咨询伙伴（如 Tavant）合作是快速破局的高效路径。服务商通常会在 8 到 12 周内协助客户基于其选定的超大规模云厂商（Hyperscalers，包括 AWS、Google Cloud Platform、Microsoft Azure 等）完成概念验证（POC），帮助企业以极低试错成本快速验证商业价值。 然而，云计算巨头的快速演进也给初创企业带来了巨大的生态挤压。两年前行业刚兴起时，许多创业者试图开发基于检索增强生成（RAG）的问答聊天机器人；但仅仅三四个月后，各大超大规模云厂商就将 RAG Chatbot 变为了平台的标配原生功能——这种现象就像是“你刚开了一家定制 T 恤店，转眼亚马逊自营平台就上线了同款自营 T 恤”。 面对超大规模云厂商每年数十亿美元的底层研发投入，初创公司的生存法则必须高度聚焦于**垂直应用层（Application Layer）**： - 避免在通用基础设施或底层框架上与巨头正面竞争； - 专精于为极度细分（Niche）的人群或特定行业场景构建具备高行业壁垒的深层业务应用。 ```mermaid flowchart LR accTitle: 现代 AI 价值分层与创业定位 subgraph Hyperscalers["超大规模云厂商 (AWS / GCP / Azure)"] H1["底层算力与基础设施"] H2["通用平台组件 (如原生 RAG / 开箱即用工具)"] end subgraph Startups["初创公司与企业突破口"] S1["垂直细分应用层 (Application Layer)"] S2["深度融合特定业务场景与工作流"] end Hyperscalers -->|基础能力赋能 / 挤压通用工具| Startups ``` ### [34:25](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=2065s) AI 辅助编码与供应链多系统智能聚合的快速成效 在探讨如何从初学者视角借力云巨头并快速见效时，行业内正在发生两项显著变革： 1. **借助云原生现成方案与辅助编程（VIP / Cloud Coding）实现极速原型构建** 对于许多通用或特定内部需求，市场上已有成熟的开箱即用工具。例如在服务阿姆斯特丹的一家全球顶尖旅游巨头时，团队直接复用了现成的工具链并快速集成到其既有的云环境中。更进一步，借助 Claude Code 等前沿辅助编码工具，服务商甚至能够在短短几天内为客户交付一个可交互的 POC 原型。尽管将其投入生产环境（Productionalize）仍需 DevOps 和底层基础设施团队做严谨的工程化保障，但它彻底颠覆了产品原型的构思与业务价值展示周期。 2. **供应链领域的智能聚合：打通数据孤岛** 在实体制造与供应链场景中，AI 带来了极其直观的提效。过去企业需要同时在 SAP 以及其他 5 到 6 个互不连通的业务系统中反复切换查询。如今通过 AI 整合，所有数据被汇聚到一个统一的操作视图中： - 能够一眼纵览未来 3 个月的整体供应链全貌、实时需求波动与供给状况； - 自动拉通音频排程（Audio Schedules）与制造工厂排产，将碎片化的订单智能合并为单一高效批次，直接消除供应链堵点； - 传统需要反复经历 2 年设计迭代的复杂系统整合，在 AI 赋能下仅用 3 个月即完成上线交付。在此过程中，AI 方案的精准验证（Validation）构成了落地推行的核心基石。 ### [37:30](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=2250s) 从供应链验证到一键式请假：AI 落地日常业务的“小而美”胜利 在讨论自动化验证时，嘉宾指出验证环节已不再依赖纯人工操作，而是由 AI 全面驱动。这意味着企业能够进行 24/7 全天候测试，实时观察修改某一个参数对供应链产生的上千种连锁反应。然而除了复杂的供应链仿真，AI 落地最立竿见影的成果往往来自极其简单的日常场景，比如员工申请带薪休假（PTO）。过去，员工提交 PTO 申请需要跨越多个系统界面、来回截屏确认；而在团队仅花费大约 6 个月时间上线 AI 对话机器人后，整个流程简化为一次简短的聊天互动。 这种便捷性在应对突发紧急状况时尤为凸显。当员工家中发生私事或急事需要当天请假时，无需再逐一向直属上司发邮件报备或向 HR 部门申请，只需向 AI 机器人发送一条快捷消息即可全流程搞定。系统会自动更新员工的日程表、完成 PTO 扣减、核算剩余年假天数，甚至向团队同步请假反馈，并自动设置好“不在办公室”（Out of Office）状态。演讲者强调，尽管这些看起来是“微小的胜利”（small wins），但从未来拥抱 AI 的角度来看，它们正是组织迈出的“巨大飞跃”（giant steps）。 - **验证自动化**：实现 24/7 不间断运行，测试单个参数变动对供应链带来的数千种潜在影响。 - **请假（PTO）自动化**：开发周期仅用时约 6 个月，替代多页面人工操作。 - **全流程联动**：单次对话即可同步联动员工日历、HR 考勤系统、剩余假期间隔核算及邮件自动回复状态。 ```mermaid flowchart TD accTitle: AI驱动的日常休假与事务自动化流程 subgraph User["员工端 (Employee)"] A["输入单条请假消息<br/>'今日因突发情况请假'"] end subgraph Agent["AI 智能助手 (AI Chatbot)"] B["解析休假类型与时间"] C["核查剩余年假额度"] end subgraph Actions["后端自动联动 (Automated Actions)"] D["更新企业个人日历"] E["同步扣减 PTO 假期间隔"] F["向主管及团队发送通知"] G["自动开启 Out of Office 状态"] end A --> B B --> C C --> D C --> E C --> F C --> G ``` ### [38:38](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=2318s) 法律与合规的效率变革：从数日外部备忘录到秒级知识综合 Truist 首席隐私官 Ron Whitworth 从法律合规的角度分享了自己的切身体会。他坦言自己并非技术极客，可能无法拿出在场技术专家那样炫酷的案例，但 AI 已经彻底改变了他的日常工作模式。他打趣道，看到现在的 AI 工具，自己很庆幸没有身处律所第一年实习律师的岗位，否则初级工作恐怕早就不保了。以往需要大量初级人力完成的案情背景调查、特定裁决命令检索等基础工作，现在完全可以由 AI 代劳。 在应对银行内部的日常咨询时，AI 也展现出了惊人的记忆与整合能力。例如，当业务部门询问“罗恩，我们在 Truist 能不能开展某项业务？”时，以往需要耗费大量时间翻阅文档，甚至可能因时间不足而不得不外聘律师事务所，花费数天时间等待一份法律分析备忘录。如今，只需将问题输入企业内部的 AI 助手，它便能立刻抓取所有相关内部政策，甚至提醒道：“上个月你和首席信息安全官（CISO）开会讨论过该议题，当时的决议如下。”AI 将内部规章、会议决策与外部公开法规拼接综合并瞬间呈现在眼前，极大地唤醒了个人记忆并成倍加速了决策进程。 - **初级法律研究颠覆**：快速生成关于特定法规与命令的全面事实背景，压缩传统律所初级研究工作量。 - **内外部信息全维度整合**：AI 将公司历史会议决议（如与 CISO 的会谈记录）、内部合规制度与外部法规模板实时串联。 - **效率飞跃**：将过去需要外部律师耗时数天出具的备忘录检索周期缩短至数秒之内。 ### [39:50](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=2390s) 代理型 AI 进军财务领域与全景式企业全息视角 谈及企业落地 Agentic AI（智能体 AI）的热门阵地，嘉宾指出财务和首席财务官（CFO）内部团队正成为核心突破口。越来越多的企业正在财务应收账款（Accounts Receivable）管理和发票自动化处理（Invoice Processing）中部署 AI 智能体。这不仅使财务专员能够更加灵活敏捷地调用全公司的财务数据，还能直接大幅缩短财务月结、季结与年结的关账周期。由于财务场景完全发生在组织内部且业务量庞大，其带来的投资回报率（ROI）几乎是立竿见影的。 随后有嘉宾总结指出，无论是供应链管理、人力行政、法律合规还是财务结算，AI 正在扮演企业管理的“统一玻璃视窗”（Single Pane of Glass）。虽然这句行业老生常谈略显陈词滥调，但它极其形象地概括了 AI 的知识管理本质：AI 打破了各个孤岛系统之间的数据壁垒，为管理者和员工提供了一个穿透整个企业业务流程、看清全貌的全局可视化管理入口。 - **核心落地部门**：企业 CFO 团队及内部财务运营。 - **首选业务环节**：应收账款追踪、发票智能审校及加速财务关账（Faster Closings）。 - **核心价值**：纯内部业务闭环，数据安全可控且 ROI 呈现迅速。 - **角色定位**：充当企业知识管理的“单块玻璃视窗”（Single Pane of Glass），提供跨部门全景可视性。 ### [40:50](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=2450s) 善用 AI 重塑职场沟通与突破技术边界的终极心得 在谈到个体员工如何赋能时，嘉宾强调使用 AI 能直接让你成为更出色的职场人，不过前提是必须时刻注意数据安全，切忌随意在公共 OpenAI 请求中输入敏感数据。在日常商务协作中，许多员工提交的报告或沟通文字冗长且缺乏重点，如果先使用 ChatGPT 这类工具润色梳理，往往能以更简练有力、更具说服力的方式呈现。无论在哪个岗位，员工都可以借助 AI 提升沟通品质，进而更有技巧地在组织内进行向上管理与内部宣导（sell up in your organization）。 全场讨论在对初学者行动指南的提炼中推向高潮。与会者提出，现代职场人的工作逻辑应当转变为“从大语言模型（LLM）停下来的地方起步”——将模型生成的结果作为思考的起跳板，去探索如何将原有业务做得更快、更高效，甚至去涉足过去不敢尝试的全新领域。一位自认非技术背景的嘉宾分享了自己的实战经历：他仅用了 7 分钟时间，配合 Claude Code 编写代码，就成功上线了一个调用全部底层 API 并实现高效数据可视化的营销页面。在过去，这需要同时协调 UI 设计团队、工程开发团队甚至外部自由职业者才能完成。嘉宾总结道，在这份“AI 入门手册”中，最核心的原则就是保持创意并赋能员工；在完善的合规与治理前提下，永不熄灭的好奇心正是拥抱 AI 时代最好的起点。 - **职场沟通提升**：借助 LLM 精炼商务写作结构，增强说服力与感染力，实现高效的内部对齐与向上汇报。 - **提示词与安全边界**：强调在日常使用公共大模型时保持数据合规意识，规范敏感信息的输入。 - **非技术人员的生产力跨越**：非专业技术背景员工借助 Claude Code 等工具，在 7 分钟内独立完成原本需要多名工程师与设计师数天才能交付的 API 可视化网页开发。 - **新手入局核心心法**：在良好治理（Well-governed）的前提下，充分释放好奇心，让人类工作接棒 LLM 的输出起点。 ## Notable quotes > [01:31](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=91s) 在某些极端场景下，AI 基础设施的开销甚至已经超越了员工的薪酬总额（payroll）。 — Shane Zide > [05:49](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=349s) 我和所有人交流时，大家几乎都觉得自己已经落后了。但觉得落后是完全正常的。即便在全美前十大银行的同行基准对标会议上，大家坐在一起讨论用例和组织架构时，所有人依然觉得自己远远落后了。 — Ron Whitworth > [07:32](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=452s) 优先选择业务影响范围可控、风险相对隔离的非核心主干环节切入。 — Paul Jeyasingh > [11:12](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=672s) 合作即是当下的新领导力（partnership is the new leadership）。 — Srini Gowda > [14:12](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=852s) 帮助内部员工提升 21% 的工作效率固然是“锦上添花”（nice to have），但远不如直接作用于盈利核心来得震撼。 — Shane Zide > [19:58](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=1198s) 我们的工作始于大语言模型（LLM）能力的终点（our job starts where an LLM stops）。 — Srini Gowda > [32:14](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=1934s) 你刚开了一家定制 T 恤店，转眼亚马逊自营平台就上线了同款自营 T 恤。 — Paul Jeyasingh > [37:30](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=2250s) 尽管这些看起来是“微小的胜利”（small wins），但从未来拥抱 AI 的角度来看，它们正是组织迈出的“巨大飞跃”（giant steps）。 — Srini Gowda > [38:38](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=2318s) 上个月你和首席信息安全官（CISO）开会讨论过该议题，当时的决议如下。 — Ron Whitworth > [40:50](https://www.youtube.com/watch?v=ID_cwUqcXEU&t=2450s) 在良好治理（Well-governed）的前提下，永不熄灭的好奇心正是拥抱 AI 时代最好的起点。 — Shane Zide ## Mentioned Jason Saltzman, Paul Jeyasingh, Ron Whitworth, Srini Gowda, Shane Zide, Aaron Miller, Ai4, CB Insights, Truist, International, Nebius, Tavant, Volkswagen Group, NVIDIA, Blackwell, AWS, Google Cloud Platform (GCP), Microsoft Azure, OpenAI, Claude Code, Perplexity, Custom GPTs, Zoom, ServiceNow, Airbnb, Meta, Cloudflare, Shopify, Pepsi, Walmart, Penske, Ryder, Ticketmaster, SAP, IBM, Oracle ## Video description Ron Whitworth, Chief Privacy Officer, Truist Srini Gowda, Vice President - Emerging Tech & Autonomous Vehicles, International Aaron Miller, Senior Cloud Solutions Architect, Nebius Paul Jeyasingh, Associate Vice President, AI & Analytics, Tavant Shane Zide, Vice President of Sales - Americas, Nebius Jason Saltzman, Head of Insights, CB Insights Abstracts This panel is designed to provide newcomers to AI with a comprehensive playbook for getting started on their AI journey. Panelists will break down the foundational concepts of AI, explain common use cases, and offer practical advice on how to navigate the rapidly evolving AI landscape. Learn how to identify the right opportunities to integrate AI into your business, understand the challenges and risks, and discover the critical steps needed to derive value from AI. Whether you’re an executive, entrepreneur, or professional, this discussion will equip you with the essential knowledge and confidence to begin your AI journey. [Source](https://www.youtube.com/watch?v=ID_cwUqcXEU)

[1:31] Shane Zide: 在某些极端场景下，AI 基础设施的开销甚至已经超越了员工的薪酬总额（payroll）。 [5:49] Ron Whitworth: 我和所有人交流时，大家几乎都觉得自己已经落后了。但觉得落后是完全正常的。即便在全美前十大银行的同行基准对标会议上，大家坐在一起讨论用例和组织架构时，所有人依然觉得自己远远落后了。 [7:32] Paul Jeyasingh: 优先选择业务影响范围可控、风险相对隔离的非核心主干环节切入。 [11:12] Srini Gowda: 合作即是当下的新领导力（partnership is the new leadership）。 [14:12] Shane Zide: 帮助内部员工提升 21% 的工作效率固然是“锦上添花”（nice to have），但远不如直接作用于盈利核心来得震撼。 [19:58] Srini Gowda: 我们的工作始于大语言模型（LLM）能力的终点（our job starts where an LLM stops）。 [32:14] Paul Jeyasingh: 你刚开了一家定制 T 恤店，转眼亚马逊自营平台就上线了同款自营 T 恤。 [37:30] Srini Gowda: 尽管这些看起来是“微小的胜利”（small wins），但从未来拥抱 AI 的角度来看，它们正是组织迈出的“巨大飞跃”（giant steps）。 [38:38] Ron Whitworth: 上个月你和首席信息安全官（CISO）开会讨论过该议题，当时的决议如下。 [40:50] Shane Zide: 在良好治理（Well-governed）的前提下，永不熄灭的好奇心正是拥抱 AI 时代最好的起点。

[2:00] Shane Zide: the cost for AI infrastructure becomes even higher than payroll in some cases. [3:31] Srini Gowda: People would think that my data is not secure. I'm losing my IP. But then the education would come into play. That education takes time. The transformation takes time. So we're looking at like run, grow, transform phase, and right now [we are] in the run phase with AI adoption within the company. [6:21] Ron Whitworth: Everybody that I talked to thinks that they're behind. It's okay to think that you're behind. When I do benchmarking, even with the top 10 banks, we get together and talk about use cases... and everybody feels like they are way behind. So, it's okay to feel that way. [11:34] Srini Gowda: partnership is the new leadership. [20:22] starts where an LLM stops [21:36] Shane Zide: don't let the hair fool you [23:20] 50% of it in fees [38:38] Ron Whitworth: I'm glad that I'm not a first year associate in a law firm because I think my job would be out the door [40:50] start where an LLM drops us off.

[0:04] I'm Jason. I'm head of insights at CB [0:06] Insights. I'm here to lead the panel uh [0:09] today. Right. We have Paul from Tavant. [0:11] I'll let you give your own intro. [0:14] >> My name is Paul Jing. I'm a VP of AI and [0:17] analytics at Tavant. Um have 15 plus [0:20] years in ML, worked at AWS, IBM, Oracle. [0:24] And uh the last 3 years it's been a [0:27] spate of Gen AI. The first two and a [0:30] half years it was all Ga chatbots, now [0:32] it's all agents. So really excited to be [0:35] in this panel. [0:36] >> Absolutely. And then we have Ron. Ron [0:38] from Truist. [0:39] >> Yeah, nice to meet you. I'm Ron [0:40] Whitworth. I'm the chief privacy officer [0:42] of Truist uh which is a bank based in [0:44] Charlotte, North Carolina. Um I am the [0:47] chief privacy officer. So I'm [0:48] responsible for the privacy program uh [0:50] technology compliance and now AI [0:52] compliance. [0:54] >> Then we have Shiny from international. [0:56] >> Yeah. Hi all. Uh great to be here. So [0:58] I'm Shiny. a head autonomy and emerging [0:59] tech international. We're part of the [1:01] Volkswagen group. I've been working on [1:03] AI autonomy since about 2017 full-time [1:07] and uh looking forward to this panel and [1:08] taking a few questions later on. [1:10] >> All right. And we have Shane from Nebus. [1:12] >> Sure. Uh Shane Z from Nebus. Uh Nebius [1:15] is one of the largest NVIDIA GPU hosting [1:17] providers globally with over 30,000 GPUs [1:20] in production and more black wells to [1:22] come. And so whatever you start with in [1:24] your playbook for getting to AI, we're [1:26] usually at the end piece when you're [1:27] going to train an inference. So we're [1:29] excited to contribute to the panel. [1:31] >> Awesome. And then right, you know, I [1:32] think the whole point of this panel is [1:34] really to recognize the fact that, you [1:35] know, for the first time in history [1:37] maybe or at least since the dot boom and [1:39] then maybe the cloud boom, we know what [1:41] everyone at the enterprise level is [1:42] focused on, right? Everyone has two [1:44] letters that they care most about. [1:45] That's why we're all here in Vegas. and [1:47] really starting to think about, you [1:49] know, from the very foundational level [1:51] from where you get started recognizing [1:53] that we're all or most of us are all, [1:56] you know, two 2 and 1/2 years into this [1:58] AI journey where we can all start. So, [2:00] you know, I think to start I'd open it [2:02] up uh and really just try to get a sense [2:04] of, you know, um where we are now, where [2:08] we see the most enterprise adoption, and [2:10] you know, sort of some what some of [2:12] those early use cases are for AI at [2:15] enterprise level. Sure. Um yeah, at the [2:19] moment uh because we see all these [2:21] projects at the tail end of their uh [2:23] their training and inference is that you [2:25] know mainly I would call digital AI [2:27] native startups into what I would call [2:30] AI enterprises are the ones when I say [2:32] AI enterprise I think Zoom service now [2:35] uh Airbnb uh large ML and AI teams that [2:38] work for them. Um and then we of course [2:40] have brand new X meta engineer $50 [2:43] million funded startups uh like five [2:46] minutes ago that are training the next [2:48] foundational model uh that's out there [2:50] and so u you know the the cost for AI [2:53] infrastructure becomes even higher than [2:55] payroll in some cases for so you really [2:57] have to be sure at what you're doing um [3:00] you know whether you're at the [3:01] enterprise or a small startup and so [3:03] that's uh the use cases we see the most [3:05] at the moment are you know around [3:07] training or taking a product to [3:09] inference to make it to faster monetize [3:12] it uh to the marketplace faster than [3:14] general compute could do. [3:16] >> And so one of the things that comes [3:17] through in that is you know there's sort [3:18] of a pressure for everyone to make the [3:20] most of AI right you know I'd love to [3:22] get you know Shiny maybe your sense on [3:23] where that pressure comes from and [3:24] whether we you sort of how justified you [3:27] feel that is you know across various use [3:29] cases. Yeah, absolutely. So, just to put [3:32] in perspective, beginning of the year we [3:34] had about zero AI projects with [3:35] international, but today we have about [3:37] close 50 projects that we're working on [3:39] trying to get them into fruition very [3:41] quickly. Some of them have multi-million [3:43] dollar impact like high seven eight [3:45] figures. Some of them are maybe in the [3:47] low six uh six figures. But the key is [3:50] there are so many use cases we can go [3:51] after with the supply chain, whether [3:52] it's manufacturing, whether it's human [3:54] resources, whether it's just general [3:55] productivity increase within the teams [3:57] per se. It's great but then the [4:00] adoptability of AI is the question mark. [4:03] How how quickly can teams actually adopt [4:04] AI? So we rolled out various AI tools [4:06] within the company and there's so many [4:08] to choose from as well in terms of [4:10] what's more effective, what's more what [4:12] could lead the team to more um fruition [4:14] activity by by end of this end of this [4:16] year for example. So just to put that [4:19] into perspective, if I take parts and [4:22] the inventory and the planning per se, [4:24] that's a big challenge today within [4:25] within our industry because sometimes [4:28] you suppliers are paid twice or [4:29] sometimes the the parts available in [4:31] stores, but you can't actually deliver [4:32] the part to the to the teams or put it [4:35] to maintenance. That itself is a big [4:37] challenge. But AI makes it simpler. But [4:39] adoptability of AI within the teams is a [4:41] question mark because where does the [4:42] data go? Where does it sit? Is the data [4:44] safe? Does it leave a tenant? If the [4:46] data does leave a tenant, where does it [4:48] go? Do we lose IP? Do we lose the kind [4:51] of a ground that we've done? So my [4:53] company's been around for about 200 [4:55] years now, right? That I work for. And [4:57] taking that transformational shift from [4:59] how we used to do things before to AI is [5:02] a considerable step. People would think [5:04] that my data is not secure. I'm losing [5:07] my IP. But then the education would come [5:09] into play. That education takes time. [5:11] The transformation takes time. So we're [5:13] looking at like run, grow, transform [5:15] phase and right now in the run phase [5:17] with AI within within adoption within [5:19] the company. It's great to see that and [5:21] and physical AI is just around the [5:22] corner. So looking at robots in in in in [5:25] a manufacturing plants um how quickly [5:27] they can adapt and how quickly they can [5:28] supplement workforce not that replace [5:30] workforce but supplement workforce [5:32] that's a great advantage as well. the [5:35] acceptability of a a fully, you know, a [5:38] six foot size robot in the manufacturing [5:40] space. Um, how, you know, how does the [5:42] manufacturing team view that? How do [5:44] they accept it? That's the educational [5:46] piece that we're going through today. [5:49] >> And then, right, you maybe taking even a [5:52] step further back than that of right, [5:53] you know, you we are this is the [5:55] beginner's panel, right? For everyone [5:57] who is almost even pre-run phase, as you [6:00] just put it, right? you know sort of [6:01] where do you see you know various [6:03] enterprises across sectors really [6:05] starting their AI journey right figuring [6:08] out whether it's an internal project or [6:09] an external project and maybe for a lot [6:12] of people here who are preun phase [6:14] thinking about you know just getting [6:17] started you know what are those hurdles [6:18] that people seem to have to overcome [6:21] >> yeah so looking at the growth of this [6:23] conference is a good indication of just [6:24] how much this has exploded over the last [6:26] couple years this is my third time here [6:28] and I mean just it's doubled last year [6:30] it doubled again this year. So it's [6:32] really a good indication of I think the [6:33] message is out. I think a couple years [6:35] ago everybody was a true beginner and [6:37] our conversations were more about who do [6:40] you involve in the room when you're [6:41] starting your AI journey. You know who [6:42] who gets involved? Who's in charge of [6:44] it? You know I think over time now that [6:46] it's shifted to where people are [6:48] accepting it's here. This is not this is [6:50] not speculation. This is changing the [6:52] world right before our eyes. But I would [6:55] say that you know everybody that I [6:57] talked to thinks that they're behind. [6:59] It's okay to think that you're behind. [7:01] Uh when I do benchmarking, even with the [7:03] top 10 banks, we we get together and [7:05] talk about use cases. How are you [7:06] thinking about this? How are you [7:07] organizing it? And everybody feels like [7:10] they are way behind. So, it's okay to [7:12] feel that way. Um but I think it's also [7:14] time that everybody, you know, everybody [7:16] has to accept what we have in front of [7:17] us, the incredible opportunities and the [7:20] incredible risks and and there are ways [7:22] that we can learn from some of the the [7:24] bigger companies that have successfully [7:26] deployed over the last couple of years. [7:28] and what's what's been good, what's been [7:29] bad, and how can we learn from those use [7:31] cases. [7:32] >> And then Paul maybe, right, as someone [7:34] who's been on the front lines of helping [7:35] those deployments, you know, what have [7:37] you seen as sort of the use cases or [7:39] quick wins, right? You know, I think the [7:41] last presenter had lowhanging fruit uh [7:43] up on the board, right? You know, how [7:45] much have we progressed even at the the [7:47] beginner's level past that lowhanging [7:49] fruit and where are we sort of in terms [7:52] of where the most value is being driven [7:53] for different businesses? [7:54] >> Awesome, Jason. See, I I get I'm in a [7:57] unique role at Tamant. I get to speak to [8:00] two or three clients every week on [8:03] agentic AI use cases, right? What we've [8:06] seen is like if you take a bank, you [8:08] know, he's a privacy compliance, chief [8:10] privacy officer, [8:12] the if you choose a use case that's low [8:17] impact from a impact perspective, like [8:20] for example, we we're working with the [8:22] bank um they're a US loan servicing [8:25] bank, one of the big names. what we're [8:28] doing for them, what we're going to do [8:29] for them is um put agentic AI for the [8:33] auditing of the loan servicing process. [8:37] You know, that's a perfect place for [8:39] entry, right? So, think of areas where [8:42] that's not going to be too too [8:44] impactful. That's a great place for [8:47] companies to get into agent DKI. Like [8:50] today, I was talking to two different [8:52] clients today. Uh one client, they [8:55] wanted to do invoice processing. They [8:56] have so much issues processing invoices [8:59] especially they get uh PDF it's agri [9:02] firm large agri firm one of the largest [9:04] agri firm in the world they're getting a [9:07] ton of invoices [9:09] and they have to do a match to their POS [9:12] their contracts and find issues right so [9:15] again that's another perfect place where [9:18] you could implement agent [9:20] um we see so many use cases like this [9:23] where it's not going to negatively [9:26] impact your business. Pick use cases [9:29] like that. That's a great place to [9:30] start. And then one good thing about [9:34] technology now is technology is the [9:36] easiest part. I would say once once you [9:39] find the right use case, the right [9:41] business case that meets compliance and [9:43] privacy needs [9:45] technology with uh I'm X AWS so I love [9:49] AWS. uh AWS, GCP, Azure, they made it [9:54] very easy to implement it and then the [9:56] best part is if you want to implement [9:58] it, you can reach out to them and they [10:00] provide you a solution architect to help [10:01] you too. Y [10:03] >> and then maybe right going back to why [10:06] some of those invoicing use cases or [10:08] that specific project was so right for [10:12] agentic, you know, application, right? [10:14] you know what what is it about certain [10:15] use cases that make them more suited to [10:17] something like Genai, something like [10:19] Agentic AI, something like even you know [10:22] we go one step further back you know [10:24] just intelligent automation right you [10:26] know how do how do you choose between [10:29] the different sort of technologies and [10:30] medium we have at our disposal [10:32] >> see one of the areas um that we ask [10:35] clients to look at is what's the area [10:37] that's most painful for you like for [10:40] I'll go back to my um auditing for the [10:43] loan servicing right? They had like 10 [10:46] people offshore doing this loan [10:50] servicing process [10:53] and then they could do only 10%age of [10:56] the entire subset of loans. So pick an [11:00] area that's high friction, [11:02] it's a lot of work and you see a lot of [11:05] potential for errors. I think that would [11:08] be a perfect sweet spot to get into [11:10] agent. [11:12] >> Awesome. And then right you know I know [11:14] the four of us at least you know coming [11:16] into this call talked a lot about the [11:18] difference between internal and external [11:19] use cases right a lot of what Paul was [11:21] just talking about is sort of the [11:23] automation or identification of an [11:25] internal process. I'd love to understand [11:27] what you guys are seeing you know in [11:29] terms of where teams that are starting [11:32] should focus and why. [11:34] >> Yeah maybe I can go. So I think I think [11:36] in in this space partnership is the new [11:38] leadership right now. There is only so [11:40] much that you can do internally in terms [11:41] of your own team, the skill sets that [11:43] you have. There is this enormous [11:46] ocean out there in terms of AI startups [11:48] that are doing some amazing jobs. It's [11:50] about selecting which ones actually make [11:52] an impact to your use case, your [11:53] business case. So when we classify [11:55] internal and external projects, for me [11:56] it's very much in terms of can you [11:58] deliver the project in that particular [12:00] timeline in a particular budget and can [12:02] we get a positive return on investment. [12:05] We've been working with a few startups [12:06] and we we've spearheading a co-inovation [12:09] model where we look at joint IP, joint [12:10] sharing, we give the data, we we get the [12:13] skill sets on board and we get those [12:14] going and we're seeing some remarkable [12:16] results already. Traditionally, if we [12:18] had done this in house completely in as [12:20] an internal project, it would have taken [12:22] me and my team probably about anywhere [12:24] 24 months to 36 months to deliver and [12:25] see the positive return on on [12:27] investment. But right now with the [12:29] co-inovation model, we actually able to [12:31] get through in about 6 months time and [12:33] we can actually get to the hands of [12:34] customers for a trial. I'm not saying [12:36] full production in 6 months, but we can [12:38] get to the hands of a customer, whether [12:39] it's a Pepsi, whether it's a Walmart, [12:41] whether it's one of the big fleets that [12:42] we have that we work with. Maybe it's [12:44] Pensky Rider. You can you can name all [12:46] of them. We can get a working prototype [12:49] into the hands and look at the feedback [12:51] and figure out is that something that we [12:53] want to do. So that's kind of the [12:54] external covation model that we're [12:55] working towards. more internal co- [12:57] innovation model is towards the [12:59] manufacturing basis that I mentioned [13:00] earlier on whether it's supply chain [13:02] manufacturing whether it's looking at [13:03] past logistics whether it's risk [13:05] quantification risk mitigation in supply [13:07] chain think about the war that's [13:09] happening in Ukraine we have a [13:11] tremendous supply base across Europe one [13:14] war can disrupt that that that our whole [13:16] entire supply chain so what we're trying [13:19] to do is using AI to look at future [13:22] impacts of that particular uh [13:24] catastrophe and what that means scan and [13:26] AI is is doing a great job for us in [13:29] terms of predicting a a a milk route of [13:31] supply parts. Not just that, but it's [13:33] also giving us an opportunity to look at [13:35] the cost crossorder parts. What does [13:37] that impact? And also the whole [13:39] conversation about tariffs that's going [13:40] on. What does that imply to us? Because [13:43] we import parts. We obviously we we [13:45] assemble a lot in the US, in Mexico, so [13:47] on and so forth. But AI is doing a great [13:49] job in terms of determining what does [13:51] that mean to us from an end billow [13:53] material cost perspective a bomb cost [13:55] and how can we absorb that cost moving [13:57] forward how do we pass that cost on to a [13:59] customer base these these are some of [14:01] the internal use case that we're working [14:03] towards but the the key point that I [14:05] want to make is co-inovation where [14:07] partnership leads a lot of the strategy [14:09] and getting those projects out the door [14:11] quickly [14:12] >> yeah I would I would also say uh don't [14:14] be afraid to find revenue in your [14:16] business if you really want to impress [14:18] your boss boss and your boss's boss find [14:20] a product or service that's making money [14:21] today and then figure out how to [14:22] increase the margin on it by 25 30%. [14:25] Because you used AI to be able to uh [14:28] make that thing faster in the [14:30] marketplace, help your your clients and [14:31] customers use it faster to make more [14:34] money for your business. Yes, anyone can [14:36] go internally and help Susie and Bob be [14:38] more productive 21%. That's kind of a [14:40] nice to have. uh but the board and the [14:43] the sea levels are not going to be [14:44] impressed like they could be if you [14:47] increase things that are already there [14:49] that they think is great but could be [14:50] even better. So those are the use cases [14:52] that we try and find. We have um for [14:54] example cloudflare that uses our GPUs [14:57] behind one of their products called [14:59] worker ID which ends up helping those [15:02] people in the uh remote workers be able [15:04] to be more productive and faster with [15:05] internal resources. And then we have [15:07] Shopify that uses our GPUs underneath [15:10] their e-commerce platform to help the [15:12] shopping cart experience be faster. [15:14] Those were existing products that now [15:16] are faster and making more money uh [15:18] versus when they were on general compute [15:20] uh in a data center somewhere. So it's [15:22] just a thought. [15:23] >> Absolutely. And I think one of the [15:25] things that we see for you know [15:27] ourselves is that you know the fast that [15:29] you know just helping people do things [15:30] faster delivering value for customers [15:33] faster right whether that's through [15:35] automating you know your internal team [15:37] members to provide that value right Rob [15:40] I'm sure you can speak to this in terms [15:42] of you know sort of how this works at a [15:44] bank uh where those relationships are [15:46] much more important and end up being [15:49] revenue generating revenue saving [15:50] whatever they might be. Yeah, a lot of [15:52] the answers to these questions are it [15:53] depends. I there's so many it depends [15:55] out there in terms of the size of your [15:56] company, the nature of your company, [15:58] what type of business are you in? So for [16:00] a bank, for example, uh you know, we [16:02] have to be very very concerned about any [16:04] usage of AI that would lead to decisions [16:07] about people. You know, our regulators, [16:09] our board, our executives, they want to [16:10] they want to make sure if and when we do [16:13] that as an industry, we have to have all [16:15] the guard rails in place. It has to be [16:17] very safe and secure because the the you [16:19] know the consequences are are pretty [16:21] severe. Um but to your point there are [16:24] certain things that are lowhanging fruit [16:26] like call centers uh or complaints [16:28] management policy management where uh [16:31] you can gain significant efficiencies [16:33] and it is sort of like lowhanging fruit. [16:35] So as an example from our our world, so [16:38] instead of a of a employee in a bank uh [16:42] you know a branch or in a call center [16:44] when somebody has a question or a [16:45] problem or an issue instead of having to [16:47] go and look for the right procedure and [16:49] what do I tell this client and what what [16:51] do I say next and where what do I do? If [16:53] you can seamlessly provide them real [16:55] time guidance and the job aids become [16:58] more of an integrated conversation with [17:00] that client, it significantly benefits [17:02] the client. It increases your own [17:04] efficiency. you may not have to hire as [17:06] many people and you're not diving into [17:08] that risky area because you're not [17:10] making decisions that impact you know [17:12] are am I going to loan to this person or [17:14] do I give this person a particular rate [17:16] which is coming but right now we are [17:18] more focused on on that more safe zone. [17:20] >> Yeah. [17:21] >> So adding adding to that Ron um one of [17:24] the key things is whenever we meet with [17:25] new clients right we talk to Ron Spear. [17:28] Yeah, [17:29] >> because compliance is the main thing and [17:31] I think to your point having the guard [17:33] rails in in place all of these [17:36] hyperscalers have made the guard rails [17:40] very simple, very easy to implement. [17:42] It's part of the agentic solution. [17:46] Obviously, there's a cost to it. In some [17:48] cases, it's pretty high. So we've seen [17:50] clients do custom guardrails and in some [17:54] cases certain guardrails come out of the [17:56] box and that's that that's easy to [17:59] implement and to help in the compliance [18:01] aspect. [18:02] And then I mean almost right when I hear [18:04] you talk about guardrails, right? One of [18:06] the first things is how do you set up [18:07] good guard rails to your point earlier [18:10] on adoption of how do you get people [18:13] internal to your team to start to use [18:15] something to keep using it and to keep [18:17] using it almost in the direction or way [18:19] that you intend them to? [18:21] >> Yeah, great great question. One of the [18:23] key things is communicating the strategy [18:26] in terms of what the end to end goal is [18:28] and looking at how the strategy would [18:30] map with the company's ROI both in [18:32] short-term, medium-term and also the [18:33] long term. Once that's clear, it's very [18:36] easy to get employees to use and and [18:38] adopt AI. But there are a few question [18:40] marks in terms of what am I going to do [18:42] with this in terms of my data piece? [18:44] Where's the data stored? How do how [18:46] robust is it? So providing guardrails in [18:48] terms of your own tenant and making sure [18:50] that the employees understand that the [18:52] data is secure whether you want to [18:54] create custom chat GPS whether you want [18:55] to create custom um agentic flows for [18:58] particular use case or task the key for [19:00] adoption is figuring out is the data [19:02] safe and it's not going to leave a [19:04] tenant and and employees feel more open [19:07] to sharing those data sets. Custom GPTs [19:10] for example took some time for adoption [19:12] within the company because you know if [19:14] if a person publishes a custom GPD by [19:17] mistake include somebody else and [19:18] becomes accessible to that to that [19:20] person as well. Uh everybody has butter [19:22] fingers right and and you might include [19:24] some people where you don't want to be [19:26] on that list and and these are the kind [19:28] of conversations for early adoption that [19:30] needs to be had. Uh but the return on [19:33] investment is phenomenal. It's [19:35] brilliant. we're looking at about 25 [19:37] hours per week uh in terms in terms of [19:40] um hours saved per employee. And if you [19:42] scale that up across the month, across [19:44] the year, it's it's phenomenal. Um the [19:47] amount of time that that people have to [19:48] go back and do more research into what's [19:51] going to what's coming up next, how do [19:53] we add u value to our customer base, [19:55] it's it's it's great. The sky is the [19:57] limit. [19:58] And then maybe one of the questions [20:00] right ends up being how do you get [20:01] people to [20:03] take an you know take geni take an [20:05] agentic workflow you know sort of to the [20:08] point where they then start what the new [20:10] version of their role may look like um [20:12] right you one of the things that we've [20:13] talked about internally a ton is right [20:15] we want to get people to right our job [20:17] starts where an LLM stop stops right how [20:20] do you think about that [20:22] >> yeah it's a challenge I mean everybody [20:24] seems threatened by by AI it's like oh [20:26] it's coming for our jobs, etc. But, you [20:28] know, it's very clear that the companies [20:30] that use AI successfully are going to be [20:32] the ones that win at the end of the day. [20:34] And if they're if you're not using it [20:36] well, it you're you're you know, your [20:38] decline is going to be pretty fast, too. [20:40] It's just that we reached that point. [20:41] So, I think part of it is just making [20:44] sure that your employees understand and [20:46] you're not looking to replace them. It's [20:48] how do you become more efficient? How do [20:50] we use your time in better ways? Let's [20:53] leverage these tools in a safe and [20:55] secure manner. and the impact of what we [20:57] can do and how fast we can do it and how [20:58] we can help our clients is magnified [21:01] significantly. So I think once people [21:02] kind of buy into that and they don't [21:04] feel as threatened and they see the [21:05] opportunity, they can explore and figure [21:08] out for their world, you know, how can I [21:10] get better using these tools and it just [21:12] sort of uh spirals from there. [21:14] >> And then Shane, that maybe gets to one [21:15] of the things you said of like don't be [21:16] afraid to go get some revenue from using [21:18] all of this, right? you know what are [21:19] the use cases right from where you sit [21:21] on a sales team that where you've seen [21:23] you know sort of truly that ability to [21:25] go and whether it's prospect more [21:27] whether it's do better sort of account [21:29] diligence whether it's to go find [21:30] insights or to write better emails [21:32] whatever it might be right you know [21:34] where where are those use cases in a [21:35] sales or right now [21:36] >> well it used to be back in the olden [21:38] days you had like a telephone book that [21:39] you would sell in and the whole thing uh [21:41] don't let the hair fool you but uh now [21:44] with um the advent of AI you know my [21:46] sellers can take 500 target accounts [21:49] within minutes and then be able to have [21:51] a full breakdown of who the ICP buyers [21:53] are within those uh companies and do [21:55] they have active projects within 10 at [21:57] the moment for of course we care about [21:58] GPU servers uh that are out there buying [22:01] at the moment and so that process [22:03] usually would have taken six weeks in a [22:06] 10 years ago uh and now they they can [22:09] have we just assigned target accounts [22:10] last week and I think they already have [22:12] them broken out and done within days and [22:15] they're already contacting these new you [22:17] know prospects and target. So, it's [22:19] amazing to see uh how it's made their [22:21] job more efficient. It wasn't a threat. [22:23] You know, it helped them actually sell [22:25] and be better. [22:26] >> And then some what what are some of the [22:27] new considerations, you know, whether [22:29] it's data sources, whether it's, you [22:31] know, sort of the ways in which you do [22:32] territory planning, the ways in which [22:34] you can sort of allocate resources, you [22:37] know, from a both a human side and sort [22:39] of a technical side, right? You know, [22:40] what are some of those considerations [22:42] that you have to look at? [22:43] >> Well, I mean, now we we can actually be [22:45] more efficient. you used to be in the um [22:48] sales field whether you're selling tires [22:50] or technology you would overhire to then [22:53] end up getting to the quota that you [22:55] have for the business and able to do it [22:56] and then 60% of your sellers would end [22:58] up making the number that you needed uh [23:00] to be successful now with the advent of [23:02] AI you can you know actually be more [23:04] lean in terms of what you actually get [23:07] out of your yield you get out of your [23:09] your sellers that you do which wasn't [23:12] the case a number of years ago um where [23:15] you were kind of flying being blind and [23:16] kind of throwing a a dart in the dark. [23:18] Yeah, it's easier. Yeah, absolutely. [23:20] >> Added to that, right? And and the market [23:22] disruption that AI brings in is [23:25] phenomenal. Right. For example, I was [23:27] talking to a minor league baseball team [23:31] is a private equity. They're buying [23:33] minor league baseball teams, right? [23:35] They're starting with one. What they're [23:38] looking at is, you know, I hate when I [23:40] go to ticketmaster.com and buy a ticket [23:43] and then I see is like 50%age of it and [23:46] fees, right? So we talking to them to [23:50] see how we can leverage agenti [23:52] to enable the minor league baseball team [23:56] to sell tickets to have a conversation [24:00] with the buyer understand what they're [24:03] looking for provide them incentives [24:05] discounts and the ability for them to [24:08] buy tickets right there minus the fees [24:11] minus the convenience fees right and [24:13] also packaging merchandise you know food [24:17] and beverage age it is a guy is changing [24:21] this space now any industry it's [24:24] disrupting the space like for example [24:26] you you talked about sales right we [24:28] working with the um large one of the [24:31] largest agree machinery companies in [24:35] Japan they have a huge presence here in [24:37] US we are working with them to get sales [24:41] customer segmentation done and also [24:45] personalized marketing for their [24:47] customers and their dealers. What would [24:49] have taken them hours and weeks you [24:53] start and it's takes weeks to do that. [24:56] Whereas now with a DPI they could just [24:58] choose a subset of customers, do the [25:01] customer segmentation, get a personal [25:03] marketing setup for them and then get [25:05] the email out and uh do the actual [25:08] followup to see what the output was. [25:11] >> Absolutely. And then right Ron maybe to [25:13] what you were saying right this you know [25:14] one of the things that comes out of this [25:15] is the personalization or the ability to [25:16] ingest a ton of information you know [25:19] when you think about I guess two sides [25:20] of the coin which is one being the [25:22] privacy side of it and then two being [25:25] the sort of end use case of how do you [25:29] sort of get people to adopt that [25:31] technology right [25:32] >> uh what do you tend to look at [25:35] >> yeah I mean so the personalization is I [25:37] think where where we're all headed I [25:38] mean the the the ability to super serve [25:41] client clients and give them what they [25:42] need as fast as they want it and uh you [25:45] know the accuracy in which you can give [25:47] it to them. When you think about you [25:48] know wealth management ad advisory [25:50] services for example rather than the [25:53] adviser just sort of you know doing some [25:55] research and figuring out you know what [25:56] what makes sense for this particular [25:58] person based on data like you know that [26:00] can be automated in a very fast manner [26:03] where you know you can hone in on this [26:05] is exactly these are this is you know [26:07] this is the the three or four options [26:08] you have here's what I you know this [26:10] then the human element comes in now [26:12] let's talk about these options but you [26:14] know all the research that goes into it [26:16] is just expedited in a way that shows up [26:18] for that client where you're you're [26:20] providing them optionality, you know, [26:23] and and and accuracy that that we didn't [26:25] have before. You know, the ability to [26:26] take a large amount of data and uh and [26:29] and synthesize it and then provide very [26:32] targeted advice for individuals is [26:35] really uh exciting. [26:36] >> Yeah, absolutely. And then right, you [26:38] know, maybe the word that comes out of [26:40] that is almost proactivity, right? You [26:42] know, it allows you to take something [26:43] that otherwise would have taken, you [26:45] know, some reactionary thing, right? [26:46] Okay, I got this account. Now I have to [26:48] go and find all of the information to be [26:51] able to make that re relevant [26:52] recommendation or to have that relevant [26:54] conversation. Now you can go out and [26:56] basically have all of the information [26:58] coming into the conversation or coming [27:00] into the recommendation as the first [27:02] touch. So or as whatever touch it is in [27:05] that life cycle. Um I think maybe [27:09] thinking shiny one of the things that we [27:10] talked about uh prior to this was [27:13] thinking about you know sort of where we [27:15] get different stages of companies [27:18] thinking about you know what they're [27:20] building and for who right I would love [27:22] to think to hear your thoughts on sort [27:24] of how how different companies can sort [27:26] of segment where they can find that [27:28] customer value. Yeah, absolutely. I [27:30] think if you look at electrification in [27:32] the transportation sector, that's a [27:34] great use case. Um, medium and heavy [27:36] duty trucking has its challenges in [27:38] terms of electrification, charging [27:39] network, charging infrastructure is an [27:41] issue. And we've been using AI to pretty [27:43] much determine where would a charging [27:45] network make more sense based on a [27:47] customer footprint data. So, when we [27:49] actually go and speak with customers, we [27:50] have this data in terms of okay, we we [27:52] know your truck footprint, we know where [27:54] you go, we know how how often your your [27:56] trucks want to stop. We also know what [27:58] the impact of the grid is going to be [28:00] and getting all of these data sets into [28:02] an AIdriven model is phenomenal. So just [28:05] to put things into perspective like some [28:07] of the esteem panelists that that said [28:08] in terms of time requirement right [28:11] traditionally this would take us a [28:12] couple of consultants it would take us a [28:15] few million dollars with consultants. It [28:16] would take us about any a year to 18 [28:18] months worth of engagement to come up [28:20] with what would a corridor to electrify [28:23] a corridor look like. But with AI today, [28:25] we can do it in six to eight months and [28:27] we don't have to spend that much money [28:28] either. That's one use case where we can [28:30] go with customers and say this is what [28:32] we have for you from from day one. And [28:33] it's not about working in silos. It's [28:36] about working with the customer with [28:37] their data piece bringing them to the to [28:39] the table as well and showing them and [28:41] mapping that out that entire piece [28:43] whether it's in Texas, whether it's any [28:45] any any other part of of the US for [28:47] example. um it's been phenomenally [28:50] different in terms of how we engage with [28:51] customer bases today. How we take their [28:53] requirements in and how we digest them [28:55] and how we give out what they actually [28:58] want. So if a customer comes in and says [29:00] I want XYZ of this buildsp spec or I [29:03] want to do this with with my particular [29:05] vehicle, it's no longer going through [29:07] the loopholes of Excel or PowerPoints or [29:10] or various processes, but we have a [29:12] great agentic model where you input the [29:15] data and comes up with some some great [29:17] parame parameters that never did so with [29:19] with an Excel model and customers [29:22] generally impressed with that piece. how [29:23] quickly we can come back with to the [29:25] questions how quickly we can come back [29:26] with solutions to the problems and that [29:28] is something we're driving progressively [29:30] towards [29:32] >> and then right Shane right one of the [29:34] things that you know Shiny just [29:35] mentioned ended up being sort of this [29:37] adaptability and speed at which you can [29:40] adapt I imagine that in the compute and [29:42] GPU space right a ton of your customers [29:44] and a ton of your accounts end up you [29:47] sort of rapidly moving through different [29:49] stages of demand for compute right how [29:52] are to you guys sort of both seeing how [29:56] that changes and how that is changing [29:58] right as people move from you sort of a [30:00] beginner stage into more complex uh [30:03] higher level arenas and then you know in [30:05] terms of how you guys are using AI to [30:07] sort of meet that change in demand. [30:09] >> Yeah. Um it's ironic even even Nvidia [30:13] who makes the servers today that are [30:15] they're powering the AI revolution [30:16] they're they have researchers that [30:18] actually work for them that actually do [30:19] their own training and inference. So [30:21] they're actually using AI themselves to [30:23] actually have better outcomes for [30:25] clients. And so we have everyone on the [30:27] spectrum of two Stamford students that [30:29] are sophomores that just raised $10 [30:31] million in seed capital and are already [30:33] training and doing inance on their [30:35] foundational model uh from dorms [30:38] uh on campus in um in Palato. And then [30:41] we've got you know Fortune 500 [30:43] corporations that are that are doing [30:44] this I'd say. Um and so they're all [30:47] looking for kind of um you know speed to [30:49] market. It's everything for them in [30:51] terms of what they do and the the and [30:54] what's really driving it to the the [30:55] rental market for GPUs is that people [30:57] don't have space in their coloss for the [30:59] modernday GPU server. You know, they you [31:03] know, it's like uh you see the power [31:04] draws that these things do and and you [31:06] just can't go to your boss's boss and [31:08] say, "We want to do AI, but it's going [31:09] to be $20 million to get a new colo." Uh [31:12] and so that forces those companies who [31:16] predominantly would have done on prem in [31:17] the past to take the use case they've [31:20] actually know will make the money in [31:21] revenue and then find a provider who has [31:24] all the compliancies and uh that are [31:27] needed to protect client data and all [31:29] those things to allow them to uh to do [31:32] AI and go to market that's there because [31:33] they're starting out with API calls with [31:35] perplexity and and um and open AI which [31:38] is not a bad thing place to start. [31:40] you're you're accessing information for [31:42] clients at a faster clip than maybe you [31:43] could do yourself traditionally on on [31:46] what you would do. But um you know we're [31:49] incredibly excited to see what our [31:50] clients do every single day. These are [31:52] you know doctors who are life science [31:54] startups. These are robotics companies [31:56] in Cambridge, Massachusetts that are [31:58] you're seeing rob you're seeing a [31:59] command stroke with a GPU server and [32:01] you're seeing the robot do instantly [32:04] what has been doing and then and what's [32:05] even crazier sometimes they're doing [32:07] things on their own uh because it's [32:09] learned from what they're doing. So it's [32:12] just fascinating with our use cases. [32:14] >> Absolutely. Absolutely. And then one of [32:14] the things you brought up there was sort [32:16] of, you know, APIs, the usage of [32:18] different models through different uh [32:20] medium, right, Paul? Something you [32:21] brought up earlier was, you know, [32:23] partnering with an appropriate cloud [32:25] provider to have access to a ton of [32:27] those frameworks and sort of what do you [32:28] what do you do beyond an API call um you [32:31] know, with pre-baked, pre-anned or sort [32:34] of better frameworks, right? Where are [32:37] you seeing most people choose their [32:38] partners and how are they choosing those [32:40] partners? So um being from Tavant [32:44] obviously uh we do a lot of we work with [32:47] a lot of companies who've never done air [32:50] a air air project before right and it's [32:52] usually it's 8 to 12 weeks P and we help [32:55] them to work with the hyperscaler of [32:57] their choice and the best part about [33:00] working with the partner is obviously um [33:03] we we've done it so many times before so [33:06] we come come to you with that knowledge [33:07] we can do a quick P help you to quickly [33:10] gain value get business insight, [33:12] business value out of it and also um you [33:15] know to Shane's point right working with [33:18] hyperscalers like AWS GCP and Azour [33:21] helps you to quickly ramp up in fact uh [33:24] you know interesting thing was uh two [33:26] years ago when I started this I almost [33:29] launched a startup on rag chat bots and [33:32] then when I was seriously about to do it [33:35] about 3 4 months later all the [33:37] hyperscalers came up with rag chatbots [33:40] as part of their regular offering. [33:43] >> Kind of like they're having a t-shirt [33:45] business and all of a sudden they have [33:47] their own t-shirts on Amazon. [33:48] >> And that is something that you see [33:50] constantly now because the hypers [33:52] scalers are putting pumping in billions [33:54] of dollars in R&D coming up with a lot [33:57] of new capabilities. So I think startups [34:01] can only play in the application [34:04] layer component. I think that's the key. [34:06] So if you are a if you are a startup or [34:09] if you're thinking through that pro [34:10] process right you have to think of [34:13] what's an application that I can provide [34:15] to a niche set of people I think that's [34:18] the that's that's sustainable that will [34:21] survive otherwise it's very tough to [34:23] compete against hyperscalers [34:25] >> and maybe thinking about it more from [34:27] you know the beginners standpoint right [34:30] not someone trying to build for you know [34:32] where what might be cannibalized by the [34:35] hyperscalers but by someone trying to [34:37] sort of draft in the wake of the [34:38] hyperscalers what do you see there [34:41] >> um I see a lot of if you go to these [34:43] hyperscalers there a lot of out of the [34:45] box solutions available in fact uh we're [34:48] working with another company one of the [34:50] biggest travel company in the world [34:53] based in Amsterdam [34:55] uh they want to do a very specific use [34:57] case for the internal use case we found [35:00] a similar tool available out there and [35:02] we we incorporating that in their [35:05] hyperscaler of choice [35:07] >> and that is something that you would see [35:10] a lot because there's a lot of folks [35:12] who've already done this journey so you [35:14] can build on top of that. So um and then [35:18] the other piece that I would like to say [35:20] is you know we have VIP coding now using [35:23] cloud code right um so in fact from [35:26] Tavan's perspective if you guys come to [35:29] us for a specific use case we'll [35:32] actually do a P for you within couple of [35:34] days and show it to you [35:36] >> v coding with cloud code has just [35:39] changed the entire spectrum obviously [35:42] you can't productionalize that that's [35:43] where your teams will have to do that [35:46] your DevOps infrastructure all of them [35:48] will have to do it. But that's just [35:50] changed the paradigm to actually see [35:53] what the actual product will look like [35:55] and what the potential business outcomes [35:57] would be. [35:59] >> That's another interesting sort of [36:00] thread, right? Maybe, you know, we have [36:02] four people who sit in very or five [36:04] people who sit in very different roles. [36:05] understanding where you're already [36:08] finding those super quick wins that you [36:11] know to Paul's point you know that [36:13] probably has some certain impact in your [36:15] ability to win any given contract any [36:17] given deal right you know what are those [36:19] things that each one of us is using AI [36:21] for in our daily task that might be a [36:23] good example for people you know on [36:26] where they can get started where they [36:27] can I mean maybe just throw some darts [36:29] at the dart board and see how curious [36:31] they can get and where they can start to [36:33] start their own adoption path [36:37] Yeah, I think I think in in the world of [36:38] supply chain we've seen quick adoption. [36:40] There are various tools today out there [36:42] that for example in my company we use [36:44] about five or six different tools that [36:45] we need to go through whether it's SAP, [36:47] whether it's various other I can't name [36:49] them for confidentiality reasons but [36:51] using AI consult them into one tool. So [36:54] we don't have to go through six [36:55] different screen or six different tools. [36:57] We can actually use one one tool to look [36:59] at what the supply chain is for the next [37:01] about 3 months, what the demand is, what [37:03] the supply looks like. We can also [37:05] configure in terms of the audio [37:06] schedules and match that with the [37:08] manufacturing teams to see the building [37:10] to actually [37:12] one of the orders not multiple orders [37:14] and that removes constraints in the [37:16] supply chain systems and we've been [37:17] we've been able to do this in a matter [37:18] of maybe of 3 months. Traditionally it [37:20] would taken us maybe about 2 years to do [37:23] something like this to go through all [37:24] the design loops and and make that [37:26] happen. What's key with AI adoption [37:28] especially is the validation piece. The [37:30] validation piece is no longer human [37:32] driven. AIdriven which means you can [37:35] test it 24 by7 you can look at what the [37:37] results might might be if you change one [37:39] of the parameters that has about a th000 [37:41] impact in the supply chain piece that's [37:43] just one of the examples but then if you [37:45] take for examp um a very simple thing to [37:48] book PTO vacation periods you'd have to [37:51] go through multiple screens multiple [37:53] snapshots to go and actually book your [37:54] PTO but now we can just do a chat [37:56] message with with an AI bot it does it [37:58] for you and we implemented that in about [38:01] 6 months time. Um, and also if you want [38:04] to take, for example, um, you have an [38:06] emergency at home, a personal life, and [38:08] and you want to take leave that day, [38:10] it's a chatbot. Again, you don't have to [38:11] email your your boss. You don't have to [38:13] email uh HR or anything like that. It's [38:16] just a quick chatbot message. It's all [38:17] done. It's all taken care of, including [38:19] from your calendar to your PTO off time [38:21] booking to how many days you have left [38:23] and what the feedback is from your teams [38:25] as well. And and you have an auto out of [38:27] office status set automatically. So [38:29] these things are great wins, small wins, [38:32] but great steps, giant steps in terms of [38:34] adopting AI for the future. [38:38] >> Yeah, I'm I'm not a super user, so I I [38:40] probably don't have as cool use cases as [38:42] many of you guys in the room, I'm sure, [38:43] but I think it's really changed my role [38:46] in terms of the legal and compliance [38:47] aspect of, you know, I am a lawyer and a [38:50] lot of the things that that I uh do with [38:52] AI. I it makes me, you know, uh re I'm [38:56] glad that I'm not a first year associate [38:57] in a law firm because I think my job [38:58] would be out the door, you know, in [39:00] terms of the initial research, you know, [39:02] tell me everything there is to know [39:04] about this particular order. Or maybe [39:06] somebody has a question like, hey, Ron, [39:08] are we allowed to do this here at [39:09] Truist? And you know, I'm like, okay, [39:11] well, let's let's let's put that into [39:13] our, you know, the AI assistant and [39:16] it'll pull up here all the relevant [39:17] policies. remember you had a meeting [39:19] about this last month with the chief [39:21] information security officer and this is [39:22] what was decided and he you know it it [39:24] get pieces together internal external [39:27] and it synthesizes it and puts it right [39:28] in front of you where it just refreshes [39:30] your memory and gives you all the [39:31] relevant information. If I had to do [39:34] that three or four years ago I may not [39:35] have had time to do it. So I probably [39:37] would have called outside counsel and [39:38] said hey can you do some research for me [39:40] and give me a memo in a few days. I get [39:43] it now within few moments and it just [39:46] expedites the ability to to to uh to do [39:48] my job. [39:50] I think one one common place where we [39:53] see enterprises jumping on to agenti is [39:57] on the finance side um for the internal [40:00] CFO team right we working with a lot of [40:03] companies who are doing exactly that [40:05] primarily on the accounts receivable [40:07] side invoice invoice processing and for [40:11] your finance associates to get the [40:14] enterprise data information regarding [40:17] financial all of that even closing doing [40:20] f even doing closings faster. So I think [40:23] that's a great place to look at because [40:26] it's internal and it gives you ROI [40:29] immediately [40:31] >> and maybe before you jump in right I [40:34] think the first three examples we've [40:35] heard as much as I hate that I'm going [40:36] to use this cliche right you know it [40:38] sounds like AI becomes this single pane [40:39] of glass through which you can view much [40:41] more of your business and have much more [40:43] visibility sort of across the business [40:45] into what is going on you know sort of [40:47] from a knowledge management perspective [40:50] >> I also think it makes you a better [40:51] employee uh by using it. You have to be [40:53] careful kind of what you're what you're [40:54] sharing in an open AI request and [40:56] everything. But I think that there are [40:58] so many times now when I see things that [41:00] other employees will present in business [41:02] writing and I think to myself if you [41:04] would have ran that through chat GBT you [41:06] would have had maybe a more concise way [41:08] to lay it out in a more you know [41:11] motivating uh way that's out there. So [41:13] there's ways that all of us can be able [41:14] to use AI even if it's not you could use [41:17] it to kind of get what you need [41:19] internally for motivation purposes to [41:21] sell up in your organization by using AI [41:23] to communicate better. [41:25] >> Y I mean right I think again to [41:27] something I said earlier, right? [41:28] all of our jobs can now sort of start [41:31] where an LLM stops or at least start [41:33] where an LLM drops us off and right [41:35] being able to use that as sort of the [41:37] jumping off point thinking about you [41:39] know okay where where can I do my what I [41:41] was already doing better faster more [41:43] efficiently where can I go and do new [41:45] things right one of my favorite use [41:47] cases is I am not particularly technical [41:50] but I was able to spin up you know a new [41:52] page on for our marketing website that [41:54] called all of our APIs and visualized [41:56] our data very efficiently [41:58] you know, in a way that took that would [42:00] have taken a design team, an engineering [42:01] team, maybe a freelancer, and all of it [42:04] was seven minutes of my time and some [42:06] clog code. So, right, you know, there's [42:08] a ton of things where if you're willing [42:10] to get creative, right, and willing to [42:12] empower people to along that adoption [42:13] journey, I think you get a ton of [42:15] immediate value. And I think we're still [42:17] in the especially in this beginner's [42:19] playbook, right? You know, curiosity, so [42:22] long as it's well governed, uh, you [42:24] know, is a great place to start.