Logistics Execs Deploy Agentic AI to Overhaul Fragmented TMS Systems

物流高管部署智能体AI重塑碎片化TMS系统 供应链自动化的瓶颈源于脏数据和遗留TMS孤岛，而非算法限制，企业需在既有记录系统之上构建智能体智能层。 物流运营商与技术专家探讨了在复杂货运和履约网络中务实应用AI、智能体工作流及量子计算的方案。发言嘉宾指出，现代企业面临的主要运营瓶颈并非仅在于模型架构，而是混乱的数据集成、因果推理以及遗留的EDI协议。他们重点介绍了实际部署案例，包括基于状态的仓库监控、动态定价、高端定制履约以及量子组合优化。在所有议题中，专家组均强调，智能工具旨在增强人类的领域专业知识并实现可衡量的ROI，而非取代既有的行业合作关系。 物流领域的企业软件采购方正迅速从传统SaaS订阅转向对嵌入式智能和自主系统的直接投资。 (0:04) 跨越10至15个第三方接触点的跨系统数据摩擦在多式联运供应链中造成了运营盲区，亟需智能体编排层来防止瓶颈和货损。 (2:41) 轻资产货运经纪商依赖综合考虑历史运费率、宏观经济指标和局部天气干扰的预测性定价引擎，以守住微薄的利润空间。 (4:43) 物流AI面临的核心瓶颈并非模型架构，而是数据质量、海运在途可视性不足，以及在不同遗留系统间缺乏因果建模。 (7:08) 量子计算通过扩展数学解空间对经典AI形成补充，无需历史训练先例即可模拟前所未有的黑天鹅突发事件。 (9:38)

Logistics enterprises are increasingly bypassing standard SaaS models to deploy intelligent and agentic AI architectures, yet operational breakthroughs remain heavily bottlenecked by legacy data fragmentation and multi-system integration challenges. Industry leaders emphasize that advanced technologies—ranging from computer vision to quantum computing—deliver the highest return when designed to augment human domain expertise rather than replace it. Data Fragmentation, TMS Silos, and Legacy Protocols (7:08) Karthikeyan Ilangovan of MODE Global asserts that machine learning algorithms are rarely the limiting factor in logistics; the true bottleneck lies in poor data quality, cross-system blind spots, and a lack of causal modeling. In multimodal transit, a single shipment must navigate 10 to 15 third-party vendor systems—including carriers, port authorities, and insurers—frequently leaving containers stranded at destination ports for weeks waiting on customs or drayage. Arjun Srinivasan of ShipStation Global explains that enterprise conglomerates often manage dozens of fragmented Transportation Management Systems (TMS), such as McLeod, Aljex, or proprietary tools. To turn these isolated systems of record into actionable intelligence, operators are migrating toward centralized lakehouse architectures on Snowflake or Databricks. Compounding this fragmentation is the industry's continued reliance on Electronic Data Interchange (EDI), which Ilangovan characterizes as "another COBOL" in the modern enterprise stack. While EDI is unlikely to vanish abruptly, both Ilangovan and panel moderator Brendan Baker predict a gradual displacement driven by modular, custom AI integrations and emerging standards like the Model Context Protocol (MCP), which standardize agent access and context across disparate legacy platforms. Quantum Computing for Combinatorial Optimization and Tail Risks (9:38) While classical machine learning relies strictly on historical datasets to generate predictive insights, Pouya Dianat of Quantum Computing Inc. argues this backward-looking dependency fails during unprecedented black-swan disruptions, such as sudden canal closures or regional warfare. Quantum computing expands the mathematical solution space, allowing supply chain networks to evaluate massive permutations, prepare for extreme tail-risk scenarios without historical precedent, and solve combinatorial routing challenges like the traveling salesperson problem far more efficiently than classical architectures. Dianat also highlights that quantum optimization provides an explainability advantage over deep learning, which often operates as an opaque black box susceptible to hallucination. Because quantum systems provide deterministic parameters that engineers can directly audit and tune, they present lower regulatory risk for mission-critical deployments. Concrete enterprise exploration is already underway, including AT&T testing quantum algorithms for real-time field crew dispatching and institutions applying quantum classification to fraud detection. Warehouse Computer Vision and Condition-Based Maintenance (13:40) Inside distribution centers, high downtime costs have turned computer vision and multimodal vision-language models into critical operational tools. Srinivasan notes that a single Amazon fulfillment facility houses 15 to 20 miles of continuous conveyor belts, where an unexpected breakdown can cost the enterprise roughly $32 million per hour in unrealized sales. Consequently, warehouse maintenance has shifted from reactive manual spot-checks to convolutional neural networks and autonomous aerial inspection drones capable of predicting component failures before belts snap. Computer vision is also transforming operations where standardized barcode automation falls short. Ilangovan recounts that at luxury retailer Neiman Marcus, managing high-value, bespoke merchandise—such as $50,000 bridal gowns and $20,000 monogrammed Birkin bags—demanded proprietary software and specialized packing stations because off-the-shelf warehouse systems could not accommodate non-standard items. Applying AI-driven visual inspection accelerated authentication and grading for luxury resale, while compressing product catalog preparation cycles for extravagant marketing showcases from six months down to immediate digital onboarding. Enterprise Adoption, Software Economics, and Human-in-the-Loop (22:55) At the enterprise rollout level, Srinivasan identifies "last mile adoption" as the primary barrier to digital transformation. End-to-end operational workflows, such as converting a freight quote into a booked shipment and generating a bill of lading, involve complex, interdependent steps that cannot yet run reliably without human supervision. To avoid user rejection, engineering teams must embed frontline operational staff into the transformation process as subject matter experts and beta testers during User Acceptance Testing. Simultaneously, AI is reshaping software economics. Baker notes that venture-backed intelligent platforms increasingly price against multi-million-dollar consulting engagements rather than modest SaaS subscription tiers. In response, firms like Accenture are deploying quantum hardware and generative AI interfaces directly in client showrooms to explore software-driven delivery models. In this fast-evolving environment, freight operators often adopt a "fast follower" strategy, monitoring large-scale experiments by first movers like Amazon and Uber Freight to adopt proven technical architectures without absorbing early exploratory losses. The panel concludes that emerging AI and quantum tools will not displace logistics personnel, but will instead function as "digital twins" augmenting human operators who possess decades of uncodified tradecraft. By automating routine administrative overhead, intelligent systems aim to capture meaningful productivity dividends—analogized by Dianat as unlocking a "25-hour day"—while preserving the vital personal relationships that underpin carrier and merchant networks.

来自 Rackhouse Venture Capital、MODE Global、ShipStation Global 与 Quantum Computing Inc. 的行业专家指出，物流企业正在跳过传统 SaaS 软件，加速转向以数据与智能为核心的决策体系；然而，算法模型本身并非技术瓶颈，跨系统的数据质量割裂、深厚的遗留架构以及一线组织的落地适配才是决定成败的关键。 现实瓶颈：数据质量与遗留系统架构的冲突 (7:08) Karthikeyan Ilangovan 与 Arjun Srinivasan 指出，限制物流 AI 发挥效能的核心难题并非算法本身的精确度，而是底层数据的质量与割裂。在多式联运场景中，单一运输触点往往需要与承运人、港口码头、仓储服务商和保险机构等 10 至 15 个第三方系统对接。在跨国海运长达数周的航程中，实时可见性极易因地缘政治扰动而中断，导致集装箱因清关或短途转运不畅在港口滞留数周，对冷链医药等敏感货品构成重大腐败风险。 在底层架构上，物流行业受制于严重碎片化的遗留系统。大型企业往往同时运行 McLeod、Aljex、Descartes、2Ship 等多套特定运输管理系统（TMS），且行业依然严重依赖被称作“技术栈中的 COBOL”的 EDI 电子数据交换标准。为打破孤岛并将记录系统（systems of record）升级为行动系统（systems of action），企业正转向 Databricks 和 Snowflake 等集中式 Lakehouse 架构，并探索模型上下文协议（MCP）与 Agent 封装层，以逐步解构并绕过老旧的数据接口。 运营执行优化：动态定价与仓储视觉监控 (4:43) 在轻资产的货运经纪业务中，企业商业模式类似于网约车平台，依赖撮合货主需求与承运人运力。由于单票货物的利润空间极其微薄，平台必须依靠规模化交易盈利。ShipStation 的机器学习定价引擎整合了多维度特征：包括历史运价走势、载重与特定线路规律，GDP 与失业率等宏观经济变量，以及墨西哥湾飓风季等极端天气指标，从而实现现货运价的动态调整，既能保护平台的毛利水平，又能帮助货主抵御运费剧烈波动。 在被称为供应链改造“零号现场”的仓储环境中，非计划停机的成本十分巨大。Arjun Srinivasan 依据此前在亚马逊仓库技术团队的经验说明，单个履约中心拥有长达 15 至 20 英里的连续传送带，每停机 1 小时即会导致约 3200 万美元的销售额损失。为此，行业运维已从被动的人工故障报修，演变为利用卷积神经网络（CNN）与视觉语言模型（VLM）进行基于状态的连续监控。在安全措施方面，室内巡检无人机与自主移动机器人（AMR）同样遵循先在物理隔离沙盒中验证、再随传感器成熟逐步与人类共享作业空间的演进路径，并在高端定制零售的视觉单品识别与防伪质检中显现成效。 量子计算协同：求解组合优化与极端情景建模 (9:38) 传统的机器学习模型本质上依靠历史经验数据进行归纳，难以对从未发生过的高破坏性事件进行有效推演。Quantum Computing Inc. 的 Pouya Dianat 阐述道，量子计算并非要取代经典 AI，而是通过大幅拓宽数学解空间对其形成补充。在突发区域冲突导致航道封锁等低概率但影响巨大的“黑天鹅”情境下，量子模型无需历史先例即可对供应链网络执行压力测试并提前制定备用路由。 在算法特性上，量子计算的核心优势在于具备高度的完全可解释性。相比深度学习常伴随的“黑箱”机制与生成式 AI 的模型幻觉，量子系统具备明确的参数调节机制，能够提供可审计的确定性结果，更符合监管合规要求。目前，该技术正切入旅行商问题（TSP）等组合优化场景，并在 AT&T 实时网络维修车队调度以及信用卡反欺诈分类等业务中展开早期探索；与此同时，埃森哲等咨询机构正将量子硬件部署于展厅，并利用生成式 AI 构建自然语言交互界面，以降低终端业务人员的操作门槛。 组织变革管理：快速跟随策略与人机协同价值 (22:55) 企业推进 AI 与前沿技术的最大阻碍在于“最后一公里”的落地采纳。Arjun Srinivasan 强调，诸如将运输报价转化为订单并生成提单（BOL）等业务，涉及环环相扣的多步骤流程，目前尚无法实现端到端全自动运行，必须依赖人机协同。若技术团队脱离实际闭门研发，上线后的采用率必然遭遇断崖式下跌，因此必须在一开始就将一线操作人员吸纳为领域专家（SME），深度参与需求定义与用户验收测试（UAT）。 面对高昂的研发与试错风险，在技术路径上采取“快速跟随者”策略往往优于激进探索。通过观察亚马逊或 Uber Freight 等先行者在大规模重构中踩过的架构陷阱，企业能以更低成本提炼出成熟的落地方案。多位专家一致认为，物流业属于紧密依赖人际关系与信任沉淀的行业，包含数十年无法被纯算法直接复制的行业隐性经验；未来的技术演进并非为了裁员，而是通过 Agent 工作流充当销售与调度人员的“数字孪生”，消除冗杂认知负担，为从业者赋予犹如“一天多出第 25 个小时”的高效价值。 本场研讨表明，物流智能化的实质是对数据整合底座、运筹优化算力与业务工作流的整体重塑。尽管 EDI 与传统 TMS 等遗留资产在短期内依然不可替代，但将高鲁棒性的数据接入、确定性优化算力与深耕一线的行业专家深度结合，已成为供应链应对全球化不确定性的核心演进方向。

## Background 本内容由 Ai4 频道于 2026 年 8 月 22 日发布在 YouTube 平台上。该圆桌论坛由 Rackhouse Venture Capital 合伙人 Brendan Baker 主持，汇聚了来自 MODE Global、ShipStation Global 和 Quantum Computing Inc. 的多位高管与技术专家，共同探讨人工智能、预测分析及量子计算在路径优化、仓储自动化、延误预测与高弹性供应链建设中的实际落地与未来变革。 ## TL;DR 本期圆桌深入剖析了物流行业从单点算法迈向端到端智能化的实际落地路径与核心瓶颈。嘉宾们指出，制约当前供应链预测精度的并非算法模型本身，而是极端碎片化的多方数据源、源头数据质量以及落后的 EDI 遗留系统。讨论不仅对比了计算机视觉在亚马逊高频流水线与奢侈品非标定制中的迥异场景，还展示了量子计算在打破历史数据依赖、探索极端边缘场景及路径优化上的独特互补优势。与会者一致认为，未来的物流智能化并非用机器彻底取代人，而是依托湖仓底座、Agentic AI 包装器与数字孪生，将数十年的行业老兵经验沉淀为一线人员的超级助手，在维系深厚人际信任纽带的同时实现渐进式蜕变。 ## Atmosphere & style 整场讨论节奏明快且极具工业实务的厚重感，既有前沿科技的宏大视野，又充斥着货运一线充满油污与灰尘的真实细节。主持人 Brendan 巧妙以风投视角穿针引线，三位技术领袖则毫无科技布道者的空泛浮躁，反而屡屡直面意外停机损失数千万美元、数据割裂难以打通以及老旧协议“死于千刀万剐”的残酷现实。现场互动幽默而务实，在谈笑间层层剥开从炒作泡沫走向商业 ROI 的深层阻力，散发出一种历经实体工业严苛敲打后所独有的从容、清醒与敬畏感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=4s) 企业在物流与实体工业场景中正跳过传统单纯的 SaaS 软件，以极快速度直接投资于智能化（Intelligence）与智能体能力。 - [02:35](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=155s) 现代物流供应链涉及跨越十余家供应商的多方系统割裂，单点机器学习模型难以应对动态冲击，推动行业转向构建端到端智能体决策层。 - [04:43](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=283s) 轻资产货运经纪依赖全要素特征工程进行动态定价，融合历史运单规律、宏观经济指标与局部极端气候以精细化平衡供需与利润空间。 - [07:04](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=424s) 预测分析落地的限速瓶颈在于源头数据质量、多系统实时集成以及外部事件因果归因的盲区，而非模型算法本身的能力不足。 - [09:38](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=578s) 量子计算通过指数级拓宽解空间打破传统 AI 对历史数据的强依赖，可在无先验数据下模拟并推演破坏力巨大的黑天鹅与极端边缘场景。 - [11:28](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=688s) 相比深度学习的黑盒与幻觉困境，量子计算具备确定性与可解释性优势，能高效求解旅行商问题（TSP）并实现毫秒级动态路径与调度优化。 - [13:40](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=820s) 仓储设备维护正从被动抢修转向计算机视觉与视觉大模型（VLM）的主动式预测，大幅减少超大型履约中心因传送带断裂等故障造成的巨额停机损失。 - [16:10](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=970s) 高客单价与非标定制化奢侈品物流无法适配现成标准化 WMS 软件，倒逼企业结合计算机视觉与自研追踪系统管理复杂个性化流程。 - [18:23](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=1103s) 仓储无人机与协作机器人的落地遵循“物理沙盒与作业空间隔离”的演进路径，在专属隔离区域闭环验证后再逐步融入复杂工况。 - [22:55](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=1375s) 组织变革的“最后一公里”需要业务专家（SME）从第一天起深度介入需求定义与 UAT 测试，在保留人类在环（Human-in-the-Loop）的前提下推动技术采纳。 - [25:54](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=1554s) 企业级 AI 落地面临高并发、旧系统集成与输出确定性等苛刻要求，企业需以严格的 ROI 论证识别并穿透供应商的“AI 包装”营销泡沫。 - [28:39](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=1719s) 咨询巨头积极引入量子与 AI 争夺存量市场，而重运营的物流企业更倾向于采取“快速跟随者”策略，汲取行业先锋的试错经验以规避落地风险。 - [34:27](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=2067s) 面对多品牌碎片化的运输管理系统（TMS），企业通过集中式现代湖仓（Lakehouse）与 MCP 网关统一构建智能层，推动记录系统向行动系统演进。 - [36:00](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=2160s) EDI 等遗留技术生态在供应链中具有较强黏性，未来 TMS 将继续作为源头记录系统存在，而跨系统工作流将交由外层的 Agentic AI 包装器协同执行。 - [39:14](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=2354s) 物流行业本质上由深厚的人际信任关系驱动，技术的终极方向是通过 Agentic AI 与数字孪生将行业老兵的专业知识转化为生产力助手。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=4s) 论坛开场与嘉宾背景介绍 主持人 Brendan Baker（来自 Rackhouse Venture Capital 的合伙人）拉开论坛序幕。他提到，Rackhouse 专注于在企业成立后的第一轮极早期阶段进行投资，并在物流与供应链等现实工业场景中进行了大量布局。从早期客户的动向中，他们观察到一个显著趋势：许多企业正在跳过单纯的传统 SaaS 软件，以比以往任何时候都快的速度直接投资于智能化（Intelligence）能力。随着会议进入午后时段，Brendan 调动现场气氛，呼吁大家打起精神、保持高能量状态，并邀请三位行业专家用简短的一两句话做自我介绍。 Mode Global 的数据分析与 AI/ML 副总裁 Karthikeyan Ilangovan 率先发言。他指出物流行业节奏飞快、充满挑战，正在经历深刻的数字化与智能化转型。随着 AI、生成式 AI（Generative AI）以及智能体 AI（Agentic AI）的兴起，坚实的数据底座变得前所未有的关键，尤其在当前充满不确定性的市场环境与地缘政治局势下，这正是 Mode Global 正在着力解决的核心命题。 Quantum Computing Inc. 的首席产品官 Pouya Dianat 接着介绍道，虽然许多人常开玩笑认为量子计算是一项遥不可及的未来技术，但它实际上正在融入当下的现实应用中。他此行旨在探讨量子计算如何与 AI 形成技术互补，并赋能物流这一最适合由量子计算发挥优势的核心问题领域。最后，ShipStation Global（由 Worldwide Express 和 Unishippers/Octane 联合构建的物流巨头）AI 与数据科学高级副总裁 Arjun Srinivasan 做了介绍，指出其业务覆盖从包裹快递（Parcel）到大宗货运（Freight）的全链路，致力于通过智能化物流平台服务广大的中小企业（SMB），改善其端到端运输体验。 --- ### [02:35](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=155s) 供应链冲击下的端到端复杂度与预测分析 Brendan 抛出核心议题：过去几年中，供应链唯一的“恒量”就是高度的复杂性与不可预测性，各种外部冲击频发。他询问预测分析技术（Predictive Analytics）如何帮助企业实时管理运输模式，并进一步赋能下游的商品陈列与需求决策。 Karthikeyan 剖析了物流供应链中盘根错节的系统整合难题。他表示，传统的经典机器学习模型在定价、路线规划和网络优化中依然扮演着重要角色，但在如今极端多变的市场环境下，仅靠单点模型已远远不够。物流供应链中任何一个触点，都可能牵扯到 10 到 15 家不同的外部供应商与合作伙伴系统；从港口转运、仓储管理到货物保险，整个端到端流程极度碎裂。举例来说，一批货物即使顺利从上海运抵休斯顿，但若在港口滞留等待海关清关长达两到三周，就会引发灾难性后果——特别是对医药等温控敏感型商品，极度依赖实时冷藏监控与动态异常决策。因此，Mode Global 今年正在正面迎战系统集成瓶颈，积极构建智能体系统（Agentic Systems），以提供端到端的自动化决策层。 ```mermaid flowchart LR accTitle: 物流供应链实时预测与决策流程 subgraph S1["多方数据整合与追踪"] A["多方系统集成<br/>(10-15家外部系统/ERP/CRM)"] --> B["实时状态监控<br/>(冷链温控/港口报关)"] end subgraph S2["智能预测与业务决策"] B --> C["智能体系统与分析层<br/>(Agentic Decision Layer)"] C --> D["端到端业务响应<br/>(路线重调/异常处置)"] end ``` --- ### [04:43](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=283s) 货运经纪撮合与动态定价策略 针对供应链定价的高波动性，Arjun 深入分享了货运经纪（Freight Brokerage）模式下的定价机制。他将 ShipStation 的货运模式类比为“货运领域的 Uber 或 Lyft”——作为轻资产经纪平台，企业本身不拥有实体卡车资产，核心价值在于高效匹配托运人（Shippers）与承运人（Carriers），业务形态涵盖整车（Full Truckload, FTL）、冷藏车（Refrigerated）、零担（Less Than Truckload, LTL）及包裹服务。 在定价层面，平台既需要在微观上为货运经纪公司创造利润空间并凭借货量实现规模效应，又必须确保托运人不承受过高溢价，本质上是一场精细化的供需容量规划（Demand Capacity Planning）。数据科学团队从历史运单数据中提炼出针对特定路线（Lane，如达拉斯至拉斯维加斯）、特定载重级别和货运类型的特征库（Features），并深度融合宏观经济指标（如 GDP 增速、失业率）与局部微观/环境指标（如墨西哥湾沿岸秋季的飓风季节等极端天气模式），从而在托运人询价下单时动态生成兼顾市场竞争力与利润率的最优价格。 - **轻资产匹配模式**：类比打车软件算法，在托运人需求与承运人运力之间进行动态调度。 - **全要素特征工程**：将内部历史运载规律与外部宏观变量（GDP、就业）及局部气象环境（夏秋飓风、季节温差）相结合。 - **动态边际收益平衡**：在确保托运人成本可控的前提下，最大化每笔交易的 Brokerage Margin。 --- ### [07:04](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=424s) 预测落地的核心瓶颈：数据质量与外部归因 Brendan 随即追问：在推高预测分析技术前沿的过程中，当前真正的“限速步骤（Rate-limiting step）”究竟是什么？是数据源整合不足、数据质量欠缺、实时性不够，还是模型本身的架构瓶颈？ Karthikeyan 与 Arjun 一致指出，“模型本身绝非瓶颈所在，真正卡脖子的是源头数据质量与实时集成能力”。Karthikeyan 以跨国海运为例，一个集装箱在港口间航行通常需要 2 到 4 周，一旦遭遇中东地缘政治危机或局部战争，航线往往被迫临时改道，而绝大多数国际物流根本无法实现高频、实时的全链路轨迹追踪。因此，清理并打通来自 ERP、CRM 和第三方系统的数据质量，其技术复杂度远高于模型算法的构建本身。 Arjun 补充道，虽然行业内拥有如 DAT（运价指数平台）或 FreightWaves SONAR（追踪招标拒绝率 Tender Rejection 等指标）等成熟的数据提供商可通过 API 接入，但当客户临时致电要求预订一趟即时货运时，行业目前依然无法完整捕捉所有外部数据源与单笔货运表现之间的深层因果关系（Causal Effect），这也构成了定价与预测的盲区。 --- ### [09:38](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=578s) 量子计算如何打破“历史数据依赖” Pouya 从量子计算的角度重新审视了物流预测问题，并阐述了量子计算如何与经典 AI 形成底层能力互补。他指出，传统 AI 模型本质上极度依赖历史训练数据，AI 所给出的“智慧（Wisdom）”完全建立在过往经验之上；然而，物流领域充斥着前所未有的黑天鹅事件。 量子计算带来的核心突破在于能够从指数级层面扩大“解空间（Solution Space）”，成倍增加预测场景的多样性。即便在缺乏历史数据支撑的前提下，基于量子物理特性的建模也能够推演并预见那些概率极低但破坏力巨大的极端场景（如突发的中东地缘冲突），帮助企业建立预案。Pouya 强调，这种场景预测能力是经典计算架构或纯 AI 无法单独实现的，其在车队预测性维护（Predictive Fleet Maintenance）和高弹性物流网络规划中具有不可替代的战略价值。 - **AI 的本质约束**：依赖历史数据训练，输出基于过往经验的经验性判断。 - **量子计算的核心赋能**：指数级拓宽解空间，在“无先验历史数据”的情况下探索极端边缘场景（Catastrophic Edge Cases）。 - **物流应用价值**：提前预演地缘冲击与设备故障，变“被动事后响应”为“前瞻预案调度”。 ### [11:28](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=688s) 量子计算的可解释性与物流路径优化 Pouya Dianat 深入探讨了量子计算相较于传统人工智能的核心技术优势——“可解释性”（Explainability）。在当今以深度学习为主导的 AI 体系中，模型常常面临“幻觉”（Hallucination）和黑盒困境，工程师往往“处于黑暗之中”，无法完全解释模型为何会得出某一特定输出；而在量子计算中，系统内部发生的过程是清晰可见的，工程师可以通过调节参数旋钮来直接改善结果质量。从法律与行业合规的角度来看，这种确定性与可解释性为企业在严格监管的环境下提供了更坚实的合规基础。 在被主持人 Brendan Baker 追问具体应用场景时，Pouya 提到了行业竞争对手公开的一个经典案例：电信巨头 AT&T 正在探索利用量子计算优化其全网运维团队（Maintenance Crew）的实时调度与部署。这种调度本质上是一个需要根据动态数据进行实时反应的复杂优化问题。Pouya 进一步解释，量子计算早期最受关注的原因之一就是求解著名的“旅行商问题”（Traveling Salesman Problem, TSP），而该问题与物流路径规划具有天然的映射关系。此外，量子计算在欺诈交易检测等高维分类问题中也展现了巨大的商业潜力。 - **确定性 vs 黑盒**：量子计算消除了 AI 的幻觉问题，运行逻辑可追溯，为企业在法律监管和审计上提供安全保障。 - **动态调度**：AT&T 探索利用量子算法实现网络维护工程师团队的毫秒级实时动态排班与路径调度。 - **核心数学模型**：旅行商问题（TSP）在量子计算下可实现指数级或大幅效率提升，直接赋能干线运输与末端派送路线规划。 --- ### [13:40](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=820s) 仓储技术突破：计算机视觉与设备的主动式预测维护 Brendan Baker 将话题引向仓储场景，指出风险投资机构普遍将仓库视为物流技术变革的“风暴中心”（Ground Zero），并向嘉宾询问各自在当前技术周期中最看好或看空的具体应用。Arjun Srinivasan 结合自己在亚马逊领导仓储技术数据科学与工程团队的经历，坚定看好利用计算机视觉进行“状态监测”（Condition-based Monitoring）与预测性维护。 ```mermaid flowchart TD accTitle: 亚马逊仓储传送带维护模式对比 subgraph 传统被动维护 ["传统被动响应模式 (Reactive)"] A1["传送带机械疲劳断裂"] --> B1["仪表盘触发警报"] B1 --> C1["人工现场排查与抢修"] C1 --> D1["停机损失巨大 (约 $32M/小时)"] end subgraph 视觉预测维护 ["计算机视觉主动维护 (Proactive)"] A2["部署高清摄像头 / 无人机巡检"] --> B2["CNN / 视觉大模型 (VLM) 实时图像采样"] B2 --> C2["识别传送带微小裂纹与磨损特征"] C2 --> D2["提前排班精准维修 / 零意外停机"] end ``` Arjun 回忆道，大约四五年前在亚马逊，一个大型履约中心（Fulfillment Center）内从入库到出库连接着长达 15 到 20 英里的传送带系统。过去，巡检主要依赖极其耗费人力的人工排查，一旦传送带发生断裂，系统在仪表板上报警后才能被动派人抢修。对于亚马逊这样体量的平台，仓库停摆 1 小时的业务中断损失高达约 3,200 万美元。团队当时通过卷积神经网络（CNN）提取图像样本训练模型，精准预测传送带的断裂风险。随着大语言模型和视觉语言模型（VLM）以及巡检无人机的引入，设备维护已经从被动抢修彻底转变为“打个响指般迅速”的主动式预防。 - **停机代价**：亚马逊履约中心每发生 1 小时意外停机，造成的销售额损失高达约 3,200 万美元。 - **传送带规模**：单个超大型仓储中心的内部传送带总长度可达 15 至 20 英里。 - **技术演进**：从早期的 CNN 图像特征分析，逐步演进到结合无人机巡检与视觉语言模型（VLM）的实时全自动预警。 --- ### [16:10](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=970s) 奢侈品供应链的极致定制化：波肯包与婚纱的履约挑战 Karthikeyan Ilangovan 随后分享了他在高端奢侈品百货内曼·马库斯（Neiman Marcus）的工作经验，勾勒出与亚马逊“标准化海量流通”截然相反的仓储世界。在奢侈品零售领域，虽然也应用了 OCR 和条形码扫描等基础路由技术，但其核心痛点在于极端复杂的“非标定制化”履约流程。 在 Neiman Marcus，客户购买的可能是价值 50,000 美元的定制婚纱，或是售价 20,000 美元且需要个性化字母刺绣（Monogramming）的爱马仕柏金包（Birkin Bag）。这类商品无法走传统的自动化分拣流水线，必须分流至专门的特种包装台进行精细化人工加工与打包。传统的现成（Off-the-shelf）仓储管理软件完全无法支撑这种高度个性化的业务逻辑，技术团队必须自研专用系统与定制化追踪工具，确保每一件高定商品从打样、走秀展示到包装出库的全流程实时可见与准时交付。 - **高客单价商品处理**：单价数万美元的婚纱与柏金包等奢品需要专门设立定制化包装工作站。 - **自研系统必要性**：传统“即买即用”的标准化 WMS 软件无法适配多变的个性化加工节点，倒逼企业开发定制化追踪系统。 - **业务场景差异**：奢侈品物流不仅要服务日常订单，还必须无缝对接周末时装周走秀与高净值客户的个性化营销活动。 --- ### [18:23](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=1103s) 仓储无人机与人机协作的安全隔离演进 针对 Arjun 提及的无人机应用，Brendan 提出了仓储行业内长期存在的顾虑：许多仓库管理层出于安全规范和法规限制，对在库内飞行作业持怀疑态度，更倾向于在现有地面工作流和自动化设备上做渐进式改造。 Arjun 以亚马逊早期引入移动搬运机器人的历程作类比，消除了这种误解。四五年前，亚马逊在试验用于“拣选、打包与发货”（Pick, Pack and Ship）的机器人时，同样面临巨大的人身安全隐患——过去工人必须亲自爬梯子在货架通道中穿梭取货。为了解决协作冲突，亚马逊首先设立了严格的“物理沙盒区”（Sandbox Environment），将机器人作业区域彻底隔离封锁，禁止人类进入，实现了“人机作业空间分离”。仓储无人机的发展也遵循完全相同的成熟路径：先在划定好的专属无人工空域内执行闭环巡检，待技术成熟并完成相关安全标准与法规认证后，再逐步探索更加复杂的作业场景。 --- ### [20:03](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=1203s) 计算机视觉在新兴电商：非标单 SKU、内容策展与二手鉴真 Brendan 结合自己的风投投资组合展开讨论，他投资了针对 Shopify 生态的退货管理平台 Loop Returns，以及致力于对球星卡、名表和手袋等二手收藏品进行智能鉴真、评级与定价的平台 Verdera。他指出，计算机视觉目前在处理“单 SKU（Single-SKU）、高配置、非标商品”方面展现出了极高的数据捕获效率，为非标准化商品贸易带来了突破性的基础设施能力。 Karthik 对此深表赞同，并补充了 Neiman Marcus 在 AI 与视觉层面的实际应用： ```mermaid flowchart LR accTitle: 奢侈品视觉AI与内容工作流 subgraph 前端营销 ["商品数字化与内容策展"] A["高定商品 / 走秀款"] --> B["AI 辅助产品详情页 (PDP) 快速生成"] B --> C["缩短 6 个月产品目录制作周期"] end subgraph 二手流通 ["二手奢品交易与品控"] D["名表 / 经典手袋回购"] --> E["计算机视觉质检与特征比对"] E --> F["快速鉴真与成色定级"] end ``` 在前端，AI 被广泛用于自动化生成和优化商品详情页（PDP, Product Detail Page）的内容。由于高定时装的季节性极强，过去制作一份覆盖秋冬或春夏的大型产品目录需要耗时长达 6 个月——其中甚至包括为特定活动定制的“粉色半百万美元劳斯莱斯”等极致单品；AI 的介入极大压缩了内容生产周期。在后端，面对近年来蓬勃发展的二手奢品转售市场，团队利用计算机视觉对名贵手表和中古包进行成色质检与瑕疵识别，有效提升了商品上架定级与流转的整体效率。 --- ### [22:35](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=1355s) 走向深水区：从技术落地到组织变革的阻力 在充分探讨了量子计算、预测性维护、无人机巡检与计算机视觉等技术工具后，Brendan 将讨论重心转向实施层面的硬骨头——“组织变革”（Organizational Change）。他强调，技术本身往往只占成功的一半，无论是企业内部团队对新工作流的接纳，还是推动客户改变长期固化的运营习惯，组织与文化层面的阻力往往才是决定 AI 与自动化能否真正释放 ROI 的关键所在。 ### [22:55](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=1375s) 组织变革管理与“最后一公里”的技术落地 在探讨如何通过组织变革将正确的技术应用于当下最关键的问题时，专家们聚焦于技术采纳的“最后一公里”难题。Arjun Srinivasan 引用了李飞飞（Fei-Fei Li）在早间主题演讲中的观点，指出自动化业务流程从来不是单一任务，而是一整串连续任务的集合。以物流领域的“从报价到订单创建”（Quote to Order Creation）及生成海运提单（Bill of Lading）为例，当前流程依然必须保留人类在环（Human-in-the-Loop），无法实现全流程端到端的完全自治。传统上执行这些重复性操作的业务人员，必须从项目立项第一天起就深度介入。如果技术团队选择闭门造车、直到后期才让业务人员接入，落地采纳率往往极低；只有让一线业务领域专家（SME）担任 Beta 测试员并参与用户验收测试（UAT），变革才能平稳推进。 ```mermaid flowchart LR accTitle: 流程自动化与业务人员协同推进流程图 subgraph S1["传统业务流程"] A["客户询价与报价<br/>Quote"] --> B["人工审核与信息录入"] B --> C["生成提单<br/>Bill of Lading"] end subgraph S2["变革管理与技术协同"] D["业务专家 SME 与 Beta 测试"] -.->|"全程参与 UAT"| E["AI / 自动化系统"] E -->|"Human-in-the-Loop"| C end ``` 针对量子计算这类前沿技术，Pouya Dianat 分享了更深层次的采纳阻力。企业对量子技术普遍存在认知误区，误以为这是遥不可及的未来技术，因而忽视眼下的应用价值。企业通常抱有两大顾虑：一是担心必须高薪聘请量子科学家才能驾驭系统，二是担忧引入自动化效率提升后会导致核心优秀员工被裁减。对此，Pouya 强调实际落地方案无需客户自建量子科学家团队，技术带来的效率提升并非为了裁员，而是重新分配员工职责，让最优秀的员工以更高效、不同的模式开展工作。 - **工作流拆解与协同**：物流订单生成涉及报价、数据核验、合规与提单生成，现阶段仍离不开人类在环把关。 - **业务专家（SME）早期接入**：一线操作人员必须在需求定义与 UAT 测试阶段全面参与，避免技术上线后的抵触情绪。 - **量子技术认知纠偏**：无需专门招聘量子科学家即可使用商业化量子方案，技术升级的核心是职能重塑而非裁减关键员工。 --- ### [25:54](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=1554s) 从ROI出发与企业级AI落地的复杂性 谈及文化与业务转型的深层逻辑，Karthikeyan Ilangovan 强调让业务部门与领域专家全程参与是取得成果的关键。MODE Global 采取了高度以投资回报率（ROI）和业务价值为核心的推进策略，将业务利益相关方与项目成败风险紧密绑定。他对比了个人 AI 与企业级 AI 的巨大鸿沟：虽然如今每个人都能在个人层面使用 ChatGPT 甚至快速搭建专属 Agent，但企业级场景面临着极高的系统复杂性、高并发扩展性、旧系统集成难题以及对输出确定性和准确性的苛刻要求。与此同时，市场上充斥着各类软件供应商生硬打上“AI Agent”标签的营销噱头，企业必须在喧嚣中保持清醒，通过与业务端共同定义问题、明确用例与衡量回报来驱动实质采纳。 Brendan Baker 对这一观点深表赞同，直言在技术泡沫（Vaporware）与真正具备商业影响力的技术之间，清晰的 ROI 论证是说服客户并促成闭环的核心关键。他提及组织变革领域的权威学者约翰·科特（John Kotter）及其经典的组织变革八步法，并顺势向嘉宾们抛出一个尖锐问题：面对传统成熟企业必须加速拥抱技术的迫切需求，在他们各自接触的客户群体与行业生态中，哪些类型的企业在技术采纳上展现出了最积极、最值得看好的行动力？ --- ### [28:39](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=1719s) 咨询巨头与货运巨头的技术采纳模式：从“快速跟随”到业务颠覆 在探讨哪些机构正在积极拥抱前沿技术时，Pouya Dianat 指出大型管理咨询公司表现出了极高的敏锐度。由于咨询行业的商业模式同质化严重，各家都在极力寻找差异化的竞争优势，因此积极将量子计算等前沿技术纳入自身方案库，以便向客户标榜技术领先地位。例如埃森哲联邦服务（Accenture Federal Services）在其位于华盛顿特区的展厅中直接部署了量子计算设备，并利用 AI 构建用户交互界面，帮助最终用户对接量子算力以解决复杂的实际业务问题。 ```mermaid flowchart TD accTitle: 技术颠覆者与快速跟随者生态对比 subgraph Innovators["行业颠覆者与先驱"] A["Amazon / Walmart"] B["Uber Freight"] end subgraph FastFollowers["快速跟随与实务落地"] C["传统货运与供应链企业"] D["吸收经验教训<br/>规避试错风险"] end subgraph Consultancies["咨询机构与生态伙伴"] E["Accenture 等咨询巨头"] F["利用 AI + 量子技术构建差异化方案"] end Innovators -->|"规模化实践输出"| FastFollowers Consultancies -->|"高阶解决方案赋能"| FastFollowers ``` 在物流与货运经纪领域，Arjun Srinivasan 与 Karthikeyan Ilangovan 则表达了更加务实的“快速跟随者”（Fast Follower）策略。结合在亚马逊（Amazon）、AWS 以及埃克森美孚（ExxonMobil）的过往经历，Arjun 表示他们高度关注 Uber Freight 和 Amazon Supply Chain 等先驱如何运用前沿技术重塑行业，但在货运经纪这类重运营领域，盲目成为最初的技术试错者风险过高，更理性的路径是观察行业颠覆者的试错经验与失败教训，快速迭代并适配到自身业务中。 Brendan Baker 提到其合伙人 Kevin 正是 Uber Freight 的创始团队成员之一，并深刻反思了纯科技创投视角对传统、复杂而混乱的实体工业所存在的盲目假设与天真认知。同时，Brendan 指出了 AI 时代咨询业面临的深层商业模式冲突：当前越来越多的智能 AI 解决方案不再以每年 10 万美元的传统 SaaS 软件来定价，而是直接锚定替代 500 万至 1000 万美元的咨询服务预算。面对软件对传统以计费工时（Billable Hours）和项目制交付为生的咨询业务的自我蚕食（Cannibalization），咨询巨头能否完成底层思维方式的转型，将成为未来技术格局的重要看点。 - **咨询业的差异化争夺**：咨询公司积极引入量子与 AI 技术作为抓手，以此争抢存量客户；埃森哲联邦服务已在其华盛顿展厅部署量子设备并配套 AI 交互界面。 - **货运业的“快速跟随”策略**：沃尔玛、亚马逊与 Uber Freight 在超大规模场景下探索技术边界，供应链中后方企业更倾向于汲取其成功经验与试错成本，实现低风险高效跟进。 - **软件对咨询预算的锚定与颠覆**：新一代企业级 AI 软件正在打破“SaaS 订阅价”逻辑，直接以替代数百万美元传统咨询项目预算的方式进行商业化变现。 ### [33:38](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=2018s) 传统巨头与 AI 原生企业的节奏之争，以及 TMS 的未来定位 在讨论物流核心系统前，主持人 Brendan Baker 分享了一个生动的对比案例：他们投资了一家从事“AI 原生”财务审计的初创公司，目标是将传统“四大”会计师事务所格局转变为“五大”。传统四大事务所制定的自动化目标通常是到 2030 年实现 25% 的自动化率；而这家初创企业仅用了 7 个月就达到了 31% 的自动化水平。这种剧烈的速度差异揭示了底层创新者与大型传统企业之间的鸿沟。随后，Brendan 将话题引向物流供应链的核心骨干——运输管理系统（TMS），抛出关键问题：将智能化工具整合进 TMS 最大的阻碍是什么？未来 TMS 是否还会以我们熟悉的形式继续存在？ ```mermaid flowchart TD accTitle: TMS 与现代化智能架构演进图 subgraph LegacyLayer["传统系统与记录源 (System of Record)"] direction TB TMS["碎片化 TMS 阵列<br/>(McLeod / Aljex / Descartes / 自研)"] EDI["遗留集成协议<br/>(EDI / COBOL / 自定义 API)"] end subgraph DataLakeLayer["统一数据与上下文层 (System of Intelligence)"] direction TB Lake["实时湖仓 / 湖基底<br/>(Lakehouse / Databricks / Snowflake)"] MCP["模型上下文协议 / 网关<br/>(MCP Gateway)"] end subgraph ActionLayer["智能执行与交互层 (System of Action)"] direction TB Agents["Agentic AI 工作流包装器"] Twin["数字孪生与业务决策赋能<br/>(Digital Twin)"] end TMS --> Lake EDI --> Lake Lake --> MCP MCP --> Agents Agents --> Twin ``` ### [34:27](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=2067s) 碎片化的记录系统与统一智能层：湖仓与 MCP 的破局点 面对 TMS 的整合难题，ShipStation Global 的 Arjun Srinivasan 从集团架构的角度给出了见解。ShipStation 旗下拥有大约 11 个品牌，日常运营中使用了大量不同的 TMS，既包含整车运输行业耳熟能详的 McLeod、Aljex 等老牌系统，也有自研的定制化系统。他强调，TMS 的核心本质是“记录系统”（System of Record）和业务的“单一事实来源”（Source of Truth），但现实中这些系统极度碎片化；与之相对的是，“智能层绝不能是碎片化的”（Your intelligence cannot be fragmented），它必须能够轻松跨越多个 TMS 进行统一调取。 Arjun 进一步指出，传统老牌 TMS 缺乏现代化接口能力，企业往往不得不耗费巨资为其搭建 API 层，更不用说支持像 MCP 网关（Model Context Protocol Gateway）这样的前沿协议。为了打破这一僵局，企业正在推进架构转型，将数据向 Snowflake、Databricks 等现代湖仓（Lakehouse/Lake Base）集中，力求在碎片化的记录系统之上构建一个统一的数据与智能层，从而让系统从“记录”平滑走向“行动”（System of Action）。 - **多品牌多系统并存**：单一集团旗下容纳十余个品牌，兼具自研 TMS 与 McLeod、Aljex 等传统商业系统。 - **智能层必须全局打通**：记录数据可以分散在各子系统，但决策与洞察智能必须具备全局视角的统一调取能力。 - **湖仓底座承接转型**：通过构建集中式 Lakehouse 实时汇聚数据，弥补传统系统缺乏现代化 API 与 MCP 网关的缺陷。 ### [36:00](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=2160s) 多模态业务差异与遗留技术枷锁：EDI 时代的“慢死”与智能体演进 MODE Global 的 Karthikeyan Ilangovan 对此深表赞同。MODE Global 也是通过并购扩张发展起来的，旗下拥有 6 个投资组合品牌，除自研 TMS 外还接入了 Descartes、2Ship、Triton 等第三方系统。他解释道，不同 TMS 本身就是为特定运输模态（Modality）量身打造的——包裹快递的业务逻辑与国际海空运截然不同，业务诉求差异巨大。然而，技术本身往往不是最大的绊脚石，真正的阻力来自生态协作机制：即便企业自身配备了最新的现代化 API，但对接的客户或承运人却依然只能用 EDI 传输，“这就像是又冒出来一个 COBOL 系统”。 在谈到 EDI 是否会被 AI 彻底淘汰时，现场展开了风趣的讨论。大家联想到 Databricks 大会上飞机拉着“SaaS 已死”横幅的戏谑场景，一致认为老旧技术不会一夜崩塌，而是会经历一个漫长的过程——“死于千刀万剐”（Dead by a thousand cuts）。针对 TMS 的未来形态，Karthikeyan 认为 TMS 依然会作为源头记录系统稳固存在，但其独立范围可能会收缩，外围则会被 Agentic AI（智能体自动化）作为包装器（Wrapper）层层包裹，负责执行跨系统的复杂工作流。 - **模态差异导致系统分割**：零担、快递、国际货运等业务需求异构，决定了各专用 TMS 难以被单一系统暴力替换。 - **生态集成壁垒凸显**：客户与承运人技术能力参差不齐，落后的 EDI 与类 COBOL 架构依然在供应链中扮演“钉子户”。 - **系统架构外包演进**：TMS 负责保留数据起源与事实记录，而上下文构建、智能洞察与工作流驱动则交由外层的 AI 智能体包装器接管。 ### [39:14](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=2354s) 未来前沿展望：人际关系底色、量子的去炒作化与数字孪生赋能 在小组讨论尾声，主持人 Brendan 邀请嘉宾们眺望未来 3 到 4 年的技术前沿，探讨哪些变化令人充满敬畏。Arjun 首先指出，交通物流行业是一个极其依赖人际关系（Relationship-driven）的领域，无论上层部署了多少高级分析或 AI 工具，业务的真正驱动力依然是与小微企业客户及承运人之间的信任纽带。技术的演进绝不能抛下生态伙伴孤军奋战，必须在维持人际连接的同时，耐心地带领仍停留在 EDI 时代的伙伴一同数字化升级。 Quantum Computing Inc. 的 Pouya Dianat 则从量子计算与 AI 浪潮的角度提出了反思。他提到，AI 行业经历过过度炒作后如今正面临现实的反弹阻力，量子计算在落地过程中必须吸取这一教训，保持克制与务实。对于新技术的终极价值，Pouya 给出了一个直观的比喻：它们相当于给人们每天 24 小时的时间里“额外赠送了第 25 个小时”，释放出的生产力如何转化为真实价值，最终取决于使用者自身。 Karthikeyan 总结道，物流行业虽然高度碎片化，却有着紧密的社群文化，沉淀着从业数十年的专业领域知识。未来的胜局不在于用机器彻底取代人，而在于通过 Agentic AI 和数字孪生（Digital Twin）把专家经验、本体知识和人际洞察沉淀为数字能力，成为一线业务人员与销售网络的超级助手。正如主持人 Brendan 最后的总结："通过领域专业知识与人际关系赢得未来，这是最完美的落脚点。" - **关系为本不可替代**：供应链的高效运转依赖深厚的人际信任网络，AI 工具是赋能关系而非切断连接。 - **技术务实防范过度炒作**：量子与前沿 AI 技术的价值在于消除繁冗工作，实质性地为从业者创造“每天的第 25 个小时”。 - **数字孪生协同进化**：将 30 年行业老兵的专业知识、语境理解与 AI 数字孪生结合，是实现物流智能化的最优路径。 ## Notable quotes > [07:04](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=424s) 模型本身绝非瓶颈所在，真正卡脖子的是源头数据质量与实时集成能力 — Karthikeyan Ilangovan > [09:38](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=578s) AI 的本质约束在于依赖历史数据训练，输出基于过往经验的判断；而量子计算的核心赋能在于指数级拓宽解空间，在无先验历史数据的情况下探索极端边缘场景 — Pouya Dianat > [13:40](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=820s) 对于亚马逊这样体量的平台，仓库停摆 1 小时的业务中断损失高达约 3,200 万美元 — Arjun Srinivasan > [22:55](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=1375s) 自动化业务流程从来不是单一任务，而是一整串连续任务的集合，现阶段仍离不开人类在环 — Arjun Srinivasan > [22:55](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=1375s) 技术带来的效率提升并非为了裁员，而是重新分配员工职责，让最优秀的员工以更高效、不同的模式开展工作 — Pouya Dianat > [25:54](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=1554s) 在技术泡沫与真正具备商业影响力的技术之间，清晰的 ROI 论证是说服客户并促成闭环的核心关键 — Brendan Baker > [34:27](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=2067s) Your intelligence cannot be fragmented（智能层绝不能是碎片化的），它必须能够轻松跨越多个 TMS 进行统一调取 — Arjun Srinivasan > [36:00](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=2160s) 即便企业自身配备了最新的现代化 API，但对接的客户或承运人却依然只能用 EDI 传输，这就像是又冒出来一个 COBOL 系统 — Karthikeyan Ilangovan > [39:14](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=2354s) 交通物流行业是一个极其依赖人际关系的领域，无论上层部署了多少高级分析或 AI 工具，业务的真正驱动力依然是与小微企业客户及承运人之间的信任纽带 — Arjun Srinivasan > [39:14](https://www.youtube.com/watch?v=E4pgFlhSlSA&t=2354s) 通过领域专业知识与人际关系赢得未来，这是最完美的落脚点 — Brendan Baker ## Mentioned Brendan Baker, Karthikeyan Ilangovan, Pouya Dianat, Arjun Srinivasan, Rackhouse Venture Capital, MODE Global, Quantum Computing Inc., ShipStation Global, Worldwide Express, Unishippers, Octane, Amazon, AWS, ExxonMobil, Uber Freight, Walmart, Neiman Marcus, AT&T, Accenture Federal Services, Databricks, Snowflake, Shopify, Loop Returns, Verdera, DAT, FreightWaves SONAR, McLeod, Aljex, Descartes, 2Ship, Triton, 李飞飞 (Fei-Fei Li), 约翰·科特 (John Kotter), MCP (Model Context Protocol), EDI, COBOL, TMS, VLM, CNN, Agentic AI, Digital Twin ## Video description Brendan Baker - Partner, Rackhouse Venture Capital Karthikeyan Ilangovan - Vice President - Data Analytics and AI/ML, MODE Global Pouya Dianat - Chief Product Officer, Quantum Computing Inc. Arjun Srinivasan - SVP AI & Data Science, ShipStation Global From freight to fulfillment, logistics players are turning to AI to navigate complexity. This panel features operators and innovators discussing how AI is improving route planning, automating warehouses, and predicting delays. We’ll also explore how leaders are using AI to hit sustainability targets and build more resilient global supply chains. We will discuss the impact of intelligent robotics on logistics, and how leaders are taking advantage of these innovations. Join this panel to hear from leading transportation and logistics executives and cutting edge technologists on how AI will shape the future of transport. [Source](https://www.youtube.com/watch?v=E4pgFlhSlSA)

[7:04] Karthikeyan Ilangovan: 模型本身绝非瓶颈所在，真正卡脖子的是源头数据质量与实时集成能力 [9:38] Pouya Dianat: AI 的本质约束在于依赖历史数据训练，输出基于过往经验的判断；而量子计算的核心赋能在于指数级拓宽解空间，在无先验历史数据的情况下探索极端边缘场景 [13:40] Arjun Srinivasan: 对于亚马逊这样体量的平台，仓库停摆 1 小时的业务中断损失高达约 3,200 万美元 [22:55] Arjun Srinivasan: 自动化业务流程从来不是单一任务，而是一整串连续任务的集合，现阶段仍离不开人类在环 [22:55] Pouya Dianat: 技术带来的效率提升并非为了裁员，而是重新分配员工职责，让最优秀的员工以更高效、不同的模式开展工作 [25:54] Brendan Baker: 在技术泡沫与真正具备商业影响力的技术之间，清晰的 ROI 论证是说服客户并促成闭环的核心关键 [34:27] Arjun Srinivasan: Your intelligence cannot be fragmented（智能层绝不能是碎片化的），它必须能够轻松跨越多个 TMS 进行统一调取 [36:00] Karthikeyan Ilangovan: 即便企业自身配备了最新的现代化 API，但对接的客户或承运人却依然只能用 EDI 传输，这就像是又冒出来一个 COBOL 系统 [39:14] Arjun Srinivasan: 交通物流行业是一个极其依赖人际关系的领域，无论上层部署了多少高级分析或 AI 工具，业务的真正驱动力依然是与小微企业客户及承运人之间的信任纽带 [39:14] Brendan Baker: 通过领域专业知识与人际关系赢得未来，这是最完美的落脚点

[0:04] Brendan Baker: skipping past SaaS and they're investing in intelligence a lot faster than they used to. [0:04] Brendan Baker: no low energy, high energy [0:04] Pouya Dianat: it's actually in the future, if not today, [7:08] Karthikeyan Ilangovan: The model definitely isn't the problem. The challenge is the quality of the data. [9:38] Pouya Dianat: AI basically gives you wisdom based on the experience. [11:28] Pouya Dianat: There are knobs that you can turn to basically get better results, better quality of results as opposed to what happens in AI which sometimes you don't know why you are getting the results that you are getting. You are very much in the dark. That doesn't happen in quantum. [13:40] Arjun Srinivasan: For a company like Amazon or anyone who has a warehouse or a fulfillment center, there's about 15 to 20 miles of conveyor belt that runs all the way from inbound to outbound... And for a company like Amazon, 1 hour of downtime translates to about $32 million of lost sales. [16:10] Karthikeyan Ilangovan: One of our toughest challenges was everything at Neiman Marcus is high-end luxury goods... specialized packing stations that handle delivery of specialized goods. Talk about like, you know, $50,000 wedding gown or your $20,000 Birkin bags that needs your monogram enabled. [22:55] Karthikeyan Ilangovan: slaps an agent name to their system. [32:04] Brendan Baker: eat their own business [38:08] Karthikeyan Ilangovan: SaaS is dead [38:08] Karthikeyan Ilangovan: death by a thousand cuts [39:14] Pouya Dianat: one extra hour for your 24-hour day

[0:04] Brendan here, partner from Brickhouse. I  am excited for this because we invest in [0:09] companies at the earliest stages, like  the first round after their formation, [0:13] and we do a lot of real world in logistics and  supply chain. And we just see from early customers [0:20] that they're skipping past SaaS and  they're investing in intelligence a [0:23] lot faster than they used to. So, uh for that  many reasons, I'm excited for this panel. Um [0:30] we're we're in the middle of the afternoon.  We're going to have to get through this together. [0:34] No low energy. High energy. All right. I'm going  to start by asking our panelists uh for one or two [0:40] sentences on their background, and then we'll get  we'll kick into it. Hi everyone. I'm Karthikeyan [0:46] Elangovan. Um I lead the you know, data analytics  and AI strategy for Mode Global. Um it's a very [0:53] interesting industry. It's very fast-paced, um and  it requires a lot of um transformations, truly. [1:00] And the data foundation seems to be the all more  focus with the AI, generative AI, and agentic AI. [1:07] It's one of the fundamental problems we are  tackling at Mode Global, um especially in this [1:12] current market conditions, um and geopolitical  condition, it becomes the center of attention. [1:18] Thank you. Um good afternoon, everyone. I'm  Pooya Dianat. I'm the chief product officer [1:23] at Quantum Computing Inc, and as the name of the  company suggests, we are in the quantum computing [1:27] space. But, um we were just joking about it that  it sounds like a technology that is so far in the [1:33] future, and I said it's actually in the future,  if not today. So, that's why I'm here to uh [1:39] talk about uh what this technology can bring in  and complement AI, uh which is something that we [1:45] do actually in in the products that we we deploy  today. And um how quantum computing can enhance [1:52] uh the logistics as one of the core uh problems  that it can solve. So, happy to be here. [2:01] I'm Arjun Srinivasan. Uh I'm the SVP of AI and  data science at ShipStation Global. Uh so, we are [2:08] um a lot logistics powerhouse of two different  companies, Worldwide Express and Octane. [2:14] Uh operate all the way from parcel to freight.  Uh and we have intelligent logistics platform [2:22] uh that we serve to the small and mid-size  businesses primarily. So, really focused on [2:28] how can we make our customers' life better, uh you  know, like for their end-to-end shipping needs. [2:35] Um so, really excited to be here. One  of the constants in the last few years [2:41] is complexity and unpredictability in the  supply chain. So, I want to start there. [2:46] Uh we're seeing constant supply supply chain  start uh shocks. How are predictive analytics [2:52] helping to manage modes in real time? And then  downstream from that into sort of merchandising [2:57] demand decisions, etc. Karthik, maybe I'll start  with you. Uh that's a good question. So, the [3:03] classic machine learning models on pricing, route  planning, and optimization are still in demand. [3:10] Um but with the current market conditions how it  is and there's so many factors that contributes [3:15] to the variables, uh we need that intelligence  layer where you can tie out different um you know, [3:23] um source systems and being able to predict  what's coming. That's the challenge being [3:30] we work with a lot of integration partners,  we work with a lot of third-party systems [3:34] um because any supply chain touch point  you think of, there could be like 10 to [3:40] 15 different vendor partners that we work with. Um  there are so many touch points where your shipment [3:46] gets carried over from one port to the other, one  place to the other, and we also do solutions like [3:53] warehousing, we do insurance, we do you know,  everything end-to-end for supply chain. So, [3:59] there goes the complexity. Your shipment can get  from Shanghai to Houston, but if it's going to sit [4:04] in the port waiting for customs clearance and  you know, sitting there for 2 weeks, 3 weeks, [4:10] especially when you're working with highly  sensitive goods like pharmaceutical, it becomes [4:14] a lot more problem. It needs refrigeration, it  needs real-time insights as to how we can solve [4:21] problems, and that's kind of the challenge that  we are tackling with when it comes to integrations [4:26] with so many different systems that can provide  data for us in real time, and that's kind of [4:32] one of the head-on challenges we are facing this  year, building out new agentic systems that can [4:39] help automate and provide decision layer for  business. I'm going to come back to part of [4:43] that in a sec, but Arjun, I want to get your  thought on pricing and pricing volatility in [4:48] the supply chain since you're so close to it. How  are you seeing managing complexity around pricing, [4:54] particularly around freight brokerage, which  I think you guys are super familiar with? [4:59] Yeah, so I'd say think of you know, like you all  would have used Uber or Lyft or any ride-share [5:06] app, think of it that we are an equivalent of  that in freight. By that, what I mean is we are [5:13] we don't own any assets, but then we kind of  match our customers who are shippers to the [5:19] carriers, whether that is a full truckload or a  refrigerated truckload or a less than truckload or [5:25] parcel services. So, where pricing plays a key  role is that, you know, like how can us as a [5:33] as a corporation as a freight brokerage make  more margin because you know, like that's how [5:39] we survive. And then, more volume that you sell,  these margins start to add up. And at the same [5:46] time, we want to make sure that our shippers uh  don't bear the cost. So, it becomes kind of a [5:53] demand capacity planning at that point. Um so,  a few things that we are doing from a pricing [5:58] standpoint um is, you know, like how are we using  I mean, like we are actually using historical data [6:04] uh for a certain lane or for a certain load or  for a certain weight class. So, we kind of uh have [6:11] come up with a set of so-called features from  a data science standpoint and then using these [6:16] features, which are like our own historical  pattern in how we have moved loads from one place [6:22] to another, like say from Dallas to Las Vegas  or whatever that is. And uh we look at certain [6:27] customer segments and say for, you know, like with  changing market patterns, with macroeconomic uh [6:35] things like GDP and unemployment rate and all that  playing into effect, with local weather patterns, [6:41] right? Like right now it is it is like peak of the  summer. Uh you know, like Gulf Coast for instance, [6:47] as you get into the fall, that's the hurricane  season. So, we look at all the macro and the local [6:52] uh market indicators and environmental indicators  and we try to come up with the right pricing uh [6:58] for that particular load to move um you know, like  to help a shipper move that load from one place to [7:04] the other. Maybe I'm going to go off schedule  immediately, which is great. Maybe briefly cuz [7:09] I want to come to to Pouya on predictive uh  maintenance in a second. But you made me think [7:15] when we're pushing the frontier of predictive  analytics in your spaces, what is the current rate [7:21] limiting step? Is it you're not getting enough  integrations with data sources? Those are not [7:27] of high enough quality or real enough time? Is it  an execution through those integrations? Is it a [7:33] model quality problem? I'm just curious if there's  one It's a little bit of a simplistic question, [7:37] but what's keeping you up at night, any of you uh  in this area? The model is definitely isn't the [7:43] problem. The challenge is the quality of the data,  right? Uh let's take your international shipping. [7:49] It takes anywhere between two to four weeks  for any container to reach from port to port. [7:55] Um and with example like war in the Middle  East, the container gets routed, you don't [7:59] have real-time tracking for most international  shipments. So, that the quality of the data um [8:06] and the integration challenges with the different  third-party systems that we work with, that proves [8:12] to be a lot more complex than the model itself.  Um and framing the model is a lot more easy. Now, [8:18] these variable factors that you have to put in um  and make sure you are also tackling the head-on [8:25] challenge of addressing the data quality at the  source system, whether it's your ERP, your CRM, [8:31] or whatever the source system is. Uh that seems  to be a lot more challenge than the model itself. [8:39] I Yeah, I'm like I can I can add on to that. Like  there are a lot of um data and analytics providers [8:44] from the transportation and freight industry. Uh  like DAT is a good source. There is another source [8:50] called Sonar, uh which looks at all the tender  rejection. A lot of external players who kind of [8:57] provide API calls to integrate with. Um so So,  like we can tap into a lot of this data. But then, [9:03] at that certain point in time, when you have to  book a load, like when your customer calls for [9:07] that earlier example, you know, like if you have  to book a load from Dallas to uh Las Vegas for [9:13] their freight needs, the that that real-time  data and then all the historical patterns, [9:20] I would say, are not completely captured yet. So,  there is a bit of the data quality problem that [9:26] Karthik was talking about. But uh we have  we have not even seen that causal effect of [9:32] every single data source that we can tap into and  that affects our load. Great point. We I want to [9:38] bring you into this at this point. Maybe starting  from a high level on where you see some of the [9:44] nearest injection points for quantum in  some of the stuff we're talking about. [9:48] But I believe you were mentioning earlier that  predictive fleet maintenance is a particularly [9:51] interesting area. So happy to have you zoom in  on that, too. Yes, 100%. So [clears throat] Oh, [9:58] the thing about I mean, I'm going to make a  quick distinction between why why quantum is also [10:03] why I'm speaking of quantum logistics. So with  the AI, there is an inherent dependency on the [10:09] past data because of the training that goes into  it. And and that AI basically gives you wisdom [10:15] based on the experience. Now, what quantum adds  to this equation is that it it in increases your [10:22] solution space, basically. It increases the number  of scenarios for predictability. So it can foresee [10:28] scenarios without past information to be to be for  whoever is doing the maintenance on the network or [10:34] to be able to foresee what is going to happen.  So that the number of scenarios and different [10:40] scenarios that are not necessarily something  that you can gain through the AI analysis, [10:44] but through quantum with the right modeling, you  can see those scenarios that may not be likely, [10:49] but it may happen, which may be catastrophic.  So it gives you the capability of being able [10:55] to plan accordingly. So if if a war in the middle  is all of a sudden is happening and no one knows [10:59] and you have some your your company has been able  to predict this scenario and have a plan for it, [11:05] then you can react. So that's what quantum can add  to the to the pool of solutions that are available [11:12] that is uniquely quantum and it's not something  that you can classically do or you can do use AI [11:18] computation to do to do that. And that  that becomes important in logistics and in [11:22] many other applications, but but particularly for  logistics, the the capability to predict. Also, [11:28] with quantum, you can you have a feature called  explain ability. That's doesn't exist in in in AI [11:36] because now you can clearly know what is happening  in your computer. It's not hallucination and you [11:41] can explain what is going on. So there are knobs  that you can turn to basically get better results, [11:46] better quality of results as opposed to  what happens in AI which sometimes you [11:51] don't know why you are getting the results  that you are getting. You are you very much [11:55] you are in the dark. That doesn't happen in  quantum. So from regulatory point of view [12:00] you can explain and be be able that from the legal  standpoint you also have a safer ground basically. [12:06] I'm not sure why explain ability would be  important in logistics and supply chain but [12:11] you know of course it is. Can you do you have a  specific example with like early customers, early [12:15] design customers, early prototypes, something  you've seen just to drill it down for us? [12:20] Some of it I can disclose  some of it I cannot but but [12:24] but I disclose something actually from a  competitor of us which deals with the logistics. [12:31] AT&T just started exploring the use of quantum  computing in deploying of their maintenance [12:37] maintenance crew for for their networks. That's a  sort of like optimization that needs to happen and [12:42] the reason that that they are going after that  it's it gives you the real time capability of [12:47] in real time based on the new data be able to  react. So that's that's one of the the information [12:52] is publicly available from our competitors of I  wish I could disclose ours but ours we have for [12:58] instance used our quantum computer to to to track  the fraudulent data in any credit card transaction [13:08] in data that that we have. Similar  concept it's it's it's basically it's a [13:13] classification problem that happens  in in in AI that can be mapped into [13:19] many other areas of application logistics being  one of them. People always I mean the reason that [13:24] people one of the reasons that people became  interested in quantum was a traveling salesman [13:29] problem. That's that can be solved more  efficiently with with quantum. That's one of the [13:35] motivations of going after that. So, that's that's  immediately finds application in the logistics. [13:40] Very cool. I'm going to shift gears. We have a  number of portfolio companies that are targeting [13:46] the warehouse. So, we kind of see warehouses as  ground zero for some of this change. I want to hit [13:51] it from a technology point of view and I want to  hit it from an organizational point of view. First [13:56] technology, I'm curious for each of you, starting  with Arjit and Karthik and Pooja, if you if you [14:01] have a point of view on it. What's one technology  or application that you're most bullish on [14:08] early on in this cycle? Or if you prefer which one  one that you'd short or that you're bearish on? [14:17] I can say you know, like if I draw from my  experience working at Amazon in warehouse [14:23] technology space using computer vision for  predictive or proactive maintenance. We call [14:28] it condition-based monitoring. That is a proven  technology. You know, like I was overseeing a [14:34] data engineering data science team for warehouse  tech for Amazon about like five four to five years [14:39] back. At that time we were using deep neural  networks. I mean, the GPTs are like we were [14:44] exploring some transformer architecture, but we  were actually using convolutional neural networks [14:49] using the classic example of taking samples from  your uh I mean, trying to collect a training data [14:58] set from the samples for like like to assess  when a conveyor belt in a warehouse would snap. [15:05] And for a company like Amazon or or anyone who has  a warehouse or a fulfillment center, there's about [15:11] 15 to 20 miles of conveyor belt that runs all the  way from inbound to outbound. And this was a very [15:17] manual process that companies like even Amazon  was solving for back then. So, I can see that [15:24] um and and with um you know, like with  large language models and particularly [15:27] vision language model, that landscape is  quickly changing. And while I exited Amazon, [15:32] they were already exploring usage of drones to  do that. So, just imagine uh that, you know, [15:38] um your your ability to do proactive maintenance  is just like come at a snap of a finger, [15:44] whereas it used to be very reactive in the past,  you know, like when the belt snapped, somebody [15:49] in the warehouse would get alerted on their  dashboard, then they have to go and fix that. [15:53] And for a company like Amazon, 1 hour of downtime  translates to about $32 million of lost sales. So, [15:59] from a business perspective, like this is a  no-brainer. So, I'm I'm like really really looking [16:03] forward to hear how that uh you know, like how how  that is evolving with the current day and age. Um [16:10] One of the things that we're seeing as investors  is the ROI from intelligent technologies is so [16:16] high because they're operational it's a collection  of operational technologies in my opinion, [16:20] so the ROI of preventing $132 million stoppage uh  seems obvious. I'm curious if you have any points [16:28] of view. Wow, nothing in scale of Amazon.  Um my previous company um at Neiman Marcus, [16:35] the warehouse solutions is kind of  little bit different. We do have [16:39] um you know, the OCR technology. We do have like  the barcode scanners that helps route packages. [16:46] One of our toughest challenges was everything  at Neiman Marcus is high-end luxury goods. [16:52] And when you cater to that kind of a clientele,  we have specialized packing stations that handle [16:57] delivery of specialized goods. Talk about like,  you know, $50,000 wedding gown or your $20,000 [17:05] Birkin bags that needs um you know, your monogram  enabled. So, being able to separate these kind of [17:12] specialized orders required customization  to existing warehouse technologies, and [17:17] that's one of the key challenges for Neiman Marcus  because everything is customized to a point where [17:23] the traditional you know buy customize doesn't  work. So we had to build you know specific tools [17:31] and systems that helps operate at that level. So  not anything on the Amazon scale but again the [17:36] more customization it is the more complex it is  and that's one of the challenges that I saw in the [17:41] luxury retail because it's not your average uh you  know end of the week sale. It's customized runway [17:49] ramp walk shows. That's your weekend marketing  team even. So everything needs to be customized [17:57] to our specific clientele and we had a  unique set of data and unique set of tools [18:02] and warehousing solutions was one of the key  because for on time delivery of these goods it's [18:08] key that you able to deliver the customization  and real time tracking of some of those uh where [18:15] those packing stations handle those customized  goods was a very interesting use case back then. [18:23] I want to I have comments about a couple of  your elements of your answer. I want to actually [18:29] go back to drones for a second because I'd love  if you can clear up a misconception for me that [18:33] I understand from warehouse environments where  there's a hesitation for drone based solutions [18:37] because you're not allowed to or you don't want  to be flying something through a warehouse whereas [18:43] if you go in and tackle a workflow from  your existing team or from your existing [18:49] automation it works a lot  better. Do I have that wrong? [18:54] Or what did you see at Amazon  with respect to sort of drones? [18:57] I would say it's I mean I can draw a parallel  to the robots that are in places like Amazon [19:04] where you know like while again like four or  five years back that area where the robots would [19:11] operate for pick back and ship That is like  you place an order, then you know, like [19:15] in the past a person had to go, climb a ladder up  the aisle, pick the order, pack that, and ship. [19:20] Robots are now doing that left and right.  Uh but then when they were experimenting it, [19:24] that area was cordoned off. So, humans and robots  were not interacting. Like like there was an area [19:29] where they used to experiment. There was only  like robots, and then humans were handling. So, [19:35] kind of like a sandbox environment. So, I can draw  the same parallel to drones. And like if there is [19:40] an area like in a fly zone, uh then you know,  like drones would be operating in an area where [19:45] there would not be any human intervention, and  vice versa. And over the time this could mature, [19:51] and then you know, like you have to go through a  lot of regulatory and uh compliance requirements [19:56] to make sure that um you know, like drones  can operate safely within a warehouse. [20:03] Interesting. And then I want to go to the second  thing. So, you mentioned Neiman Neiman Marcus [20:07] is a fascinating example because it's it's so  high-end. It is so bespoke. Um we have a couple [20:13] touch points. I've invested in a company called  Loop Returns, which is I would say the leading [20:17] Shopify-based returns management platform. But we  also have one called Verdera, which helps sort of [20:23] authenticate, rate, and price collectible goods  like cards, you know, handbags, watches, etc. [20:29] The thing that ties these together, which you  almost got to, is vision is so interesting right [20:34] now as a way to efficiently capture information  and data on unique goods, single SKU goods, highly [20:41] configured purchases, anything that isn't like  massive SKU count, highly searchable, optimizable [20:46] commerce. And we're seeing just the beginnings  of those types of inroads as vision-based [20:53] intelligence into these these areas. I'm curious  if that brings up anything for for you all. [20:59] So, Neiman Marcus does few things with the  eye, especially their PDP, their product pages. [21:06] Um there's a lot of work in terms of curating that  content. AI helps expedite and we were able to see [21:14] significant gains in that space. We also dealt  with luxury resales as well. Your luxury watches [21:22] and those for that to look at the  quality control, the vision you know, [21:29] you know, and AI kind of helped a little bit,  but again, that was few years ago. We were still [21:34] in the early phases of experimentation,  but those are some of the things that [21:40] where AI could help us not only bring quality,  but also expedite the way we put output to because [21:48] again, fashion with especially with fast fashion,  right? Again, there's a demand for goods at a very [21:53] faster pace. Neiman still goes through traditional  like fall, spring, summer catalogs, but even [21:59] curating that catalog is a 6-month effort for us,  right? Because again, you have your customized [22:07] half a million-dollar Rolls-Royce in pink that  needs to be ready for time for your fall fashion [22:13] event. So [laughter] that's that's kind of  like the complexity and customization and that [22:20] is the need of the industry. So again, there's  little flexibility with which we can work with, [22:26] but that requires a lot of innovation in terms  of AI and that's kind of a unique space to be in. [22:35] Awesome. I want to shift out of the technology  for a second into the organizational challenges [22:39] and bring the three of you in and make sure that  Puya gets a chance to to address this as well. Um [22:46] technology is only part of the battle. A lot  of it is organizational change right now, [22:50] whether it's within our own orgs or especially  for our customers. What is the most critical [22:55] thing from an organizational change point of  view to get the right technology tackling the [23:00] right problems right now? I'll be start with  Arjun, but I'd love to hear all three of you. [23:06] Yeah, sure. I mean like that last mile adoption  is what you are kind of like referring to here, [23:11] right? Like I see that day in and day out. Um  all the way from say like very simple operations [23:17] workflow. When we automate an operational  workflow, I think like Feifei Li like she kind of [23:22] touch base on that at keynote this morning. When  you automate a task, it's not one task, it's a [23:26] series of multiple tasks. So say like when you are  automating a task like say you know like for us, [23:31] um I can talk about like a quote to order  creation. You know like you get a quote [23:35] from a customer and from that you need to  create something called a bill of lading. [23:39] Uh people in transportation industry know what  I'm talking about. So like that's that's how an [23:42] order is actually made. And you have like humans  in the loop. You need human in the loop. It's not [23:47] a completely autonomous process end-to-end yet.  All right, so like where I see change management [23:53] is that people who have been traditionally doing  this work, which is mostly like routine work, [23:59] you know like they're um I mean like you have to  bring them along when you start your journey when [24:05] you automate this. If you're already working on  automated like automation up front and then bring [24:10] them in the later part of it, then I you know like  I'm I'm I have seen and experienced firsthand that [24:16] the adoption is going to be low. So they have to  be the beta testers, they have to be involved in [24:20] even UAT testing. You know like they are your SMEs  and they have to carry it along with a technology [24:26] group uh within any company, any enterprise across  any industry vertical. So that is my experience. [24:33] Um I can also speak to that. The you mentioned  adoption. Uh actually that's that's the challenge [24:38] that the main the main challenge that that we have  for for our technology. I mean there there are the [24:42] perceptions about the technology that a quantum  technology that is that exists that um people [24:47] think that it's a technology of future. Why should  I look at it now? And that needs to be addressed. [24:52] Then the other part is the adoption within that  institution. What what we go we first need to [24:57] show the value that that that it brings to that  organization and that's that's a key part and then [25:03] the other question that people have is, okay, do  I need to like go hire now quantum scientist to [25:09] be able to use this? And the answer is no, you  don't need to do that. So that's another thing [25:14] that needs to get addressed. And then the other  thing is, okay, now that what we have, we bring [25:18] in operational efficiency to any organization,  that's part of the value proposition, [25:22] what does it mean for that organization? Do  they lose people or do they like reassign [25:25] people or actually the same people can continue  working but differently. So that's another [25:30] challenge that we need to address that we've  experienced with some of our customers that [25:34] they are like, okay, if I do this, am I going  to to lose like two of my best best employees? [25:40] The answer is no, it's just that we're assigning  people to do that. So these are some of the things [25:43] at least from the quantum perspective given that  it's it's a little bit in the earlier stage, [25:48] we need to go through organizations  and help them understand this better. [25:54] Talking about change management,  the culture and the transformation, [25:58] one of the big wins we've seen is bringing in the  business and SMEs into the process from the start. [26:06] One of the key transformation things that we did  was working on very ROI and value-based approach [26:16] and having the business stakeholders  and the business owners part of [26:25] the risk of not being able to deliver a value.  The other thing is AI, everybody is using it at a [26:31] personal level, right? Everybody has cloud GPT,  they are able to even build their own agents, [26:37] but when you look at it from an enterprise level,  it's a completely different ballgame because of [26:43] the complexity, the scalability, the integrations  and being able to provide the right output. [26:49] On the other hand, you have vendors, partners,  every tool kind of slaps an agent you know [26:56] name to their system. So it's kind of a fine  balance between those when you are trying to drive [27:02] an adoption. And the biggest thing is bringing  the business into the framework has helped us [27:08] define what's the problem statements, what's  the use case, what's the value you're going [27:12] to drive with it. And it becomes easier for us  for people um you know, to be able to leverage [27:20] um technology and uh drive higher adoption. Yeah,  I love that you said ROI because I think that is [27:27] I emphatically believe that that's true. And I  also believe in technology that actually has an [27:33] impact versus vaporware and we're kind of in  an era where the ROI is so exciting from an [27:38] operational point of view that making that final  link and making sure that we believe it ourselves [27:44] but can convince customers of it, I think is one  step that is very very powerful right now. Um [27:50] organizational change is something that I've cared  about for a long time. There's some good research [27:54] on it. I think it's John Kotter KOTTER basically  has studied this for his entire academic career. [27:59] Has eight steps to how you would change  your organization or or a customer's. [28:04] I'm going to put you three on the spot.  You may want to answer this or not but if [28:09] if you believe that it is important for existing  incumbent companies to rapidly adopt technology [28:15] right now which I I think it is. And then you  have a customer view across your customer set. [28:21] What names are you particularly excited  about where you said, oh yeah, I believe [28:25] in that company because I see great behavior  in terms of adoption from the vendor side. [28:32] Is anybody willing to answer that  question? Or from past roles. [28:39] Um I can actually talk talk  on this one. It's Let's go. [28:43] I cannot name names but I can say the  market segment. Actually got it got it. [28:47] And the market cap and the hair color of  the CEO. Well, it's interesting enough um [28:53] the consulting companies are are eager to  adopt the technology. And the reason is that um [29:00] their business models across the board it's very  similar to each other. That they are looking [29:04] for that like even even a small incremental  difference that the distinction that they can make [29:10] and capture new customers. So um some consulting  firms are working very closely with us on adopting [29:16] this new technology because that enables them  to go to their customers and be like, "Hey, [29:21] uh I can give you this this thing that these  other guys are not are not giving to giving it [29:25] to you." So um they are um although their internal  process processes and procedures are very slow, [29:32] but some of them are very eager into getting  their hands on this this technology on the [29:36] quantum technology and and adopt it and deploy it.  Um some of them are bragging how how how about how [29:42] they are the first one getting into it and make  sure that their customers are ready for adopting [29:46] this technology when when it becomes mainstream.  So that's that's been my experience on that. [29:51] I can say I mean I've been in tech with Amazon  and AWS in the past and I've also been um [29:59] you know, like with them in market um like how I  am right now and then in the past with ExxonMobil. [30:04] So I can draw from my experiences that I always  look up to tech companies who are doing very [30:09] similar things. So for now uh in ShipStation world  I look into like say Uber freight or even Amazon [30:17] their supply chain solutions and how  they use tech to solve business problems [30:21] are kind of forefront. So you know, like always  love to uh get I mean like I'm always inspired [30:29] and love to learn from what they are doing.  Um but that also means that you know, you are [30:34] a fast follower most of the case and I I think  like honestly speaking I don't want to be an early [30:40] adopter. Um you know, like especially in a freight  brokerage kind of an industry, I mean, like, [30:45] not mean to say that we don't do any innovation,  but I would rather be a fast follower and see, [30:49] you know, how other industries who are at  the forefront of technology solve that. Um, [30:55] and then quickly learn from them and adapt  to our business needs. Like Arjun said, um, [31:00] the supply chain space, there's a space for  everyone. Um, there are industry disruptors, um, [31:07] but there comes a challenge with, like now,  being the first in kind of disrupting. So, um, [31:13] in companies like Walmart and like now, um,  Uber Freight, they are redefining the industry [31:19] in a way [snorts] that, like now,  it's massive scale, like Amazon. [31:24] But the learnings comes out of it, both positive  and what worked and what failed. Um, that's an [31:29] interesting learning opportunity for all of us  in the industry to, you know, take it forward, [31:35] um, and make sure we don't repeat the  mistakes again. My Rafa's partner, Kevin, [31:40] was one of the founding team at Uber Freight. So,  I will tell him that he got multiple shout-outs [31:45] shout-outs on the on the panel. Um, that's  actually validating because as somebody who comes [31:49] more from pure tech, I think that we often make  assumptions in tech and startups about what the, [31:55] like, more traditional and messier industries  need, and oftentimes we're wrong, um, and we're [32:00] naive. And so, when we actually jump over the  fence and listen and get it right eventually, [32:04] I think that's super validating. I also want  to go back to the consulting one for a second. [32:10] One of my lines these days is that Accenture  should buy a third of our portfolio right [32:14] now. It's really interesting watching the  consultancies work and not work in this era of AI, [32:21] particularly for one reason. Like, we're seeing  more and more intelligent solutions actually [32:26] anchor their pricing against consulting spend.  So, instead of anchoring against a 100K per [32:32] year SaaS spend, you're anchoring against 5 or 10  million in consulting spend to solve this problem. [32:38] Um and it'll be interesting to see if any  of them can sort of eat their own business, [32:43] which is often billable hours or project outcomes.  Accenture seems to be the one that is most deeply [32:50] engaging of the big consultancies, but  but uh we'll see. I mean Um you mentioned [32:56] Accenture. Accenture Federal Services is  the company one of one of the consulting, [33:00] but is the one this one is also public, so I can  is has has one of our computing machine in their [33:07] in their showroom in in Washington, D.C. actually,  and they are exploring it. And um interesting [33:13] enough, they are actually using AI to build the  interface for for to for for people for the end [33:19] users to interface with our quantum computer and  solve their problem there. So, um Accenture is [33:24] the default to Yeah, a leader in that field. I  think they're I think they're doing a good job [33:30] and still they'll have to switch mindset to sort  of really cannibalize some of their business. [33:34] Uh another quick example on the consulting stuff  and then I've got a question about TMS. I'm going [33:38] to plant that meet that that seed in your heads.  We have a we have a company that does AI native [33:42] financial audits essentially sort of take turning  the big four accounting firms into a big five. [33:46] And their goals for automation, the sort of  big four accounting firms, is I think 25% [33:52] automation by 2030. Our company is at 31% in  7 months. So, there's a difference in pace [33:59] between the people that are really building  a lot of these core technologies and some of [34:03] the incumbents and larger larger orgs, but  we'll see how it plays out. I'm curious to [34:08] get everybody's take on TMS. It's kind of one  of the, if not the, backbone data systems. [34:14] Um what are some of the biggest hurdles when  integrating intelligent tools into TMS? Uh maybe [34:20] a punchier question, does TMS continue to exist  as we've known it? Or what happens in that corner? [34:27] I can um give a perspective from my side.  So, ShipStation Global is a conglomerate [34:32] like we have about like 11 brands under the  hood. That means that a lot of TMSs we use [34:38] some of the TMSs that the truckload industry or  transportation industry is aware of McLeod, Aljex, [34:44] and we also have our own custom-built TMSs.  I think the main challenge that I see is [34:48] that your TMS is your source of truth, so they  are your systems of record. And they are still [34:53] fragmented. Some consolidation has happened  over the years and it continues to happen. [34:57] Your intelligence cannot be fragmented. Right?  Like it has to, you know, like it has to exist in [35:04] an easy to tap across multiple TMSs. So the  main challenge I see that is with integration. [35:10] And you know, like with the very legacy TMSs  like McLeod or Aljex and and others out there, [35:17] you know, like things like API is I mean,  like you have to go through a challenge of [35:22] having to build an API layer to them. They still  don't have things like MCP Gateway. You know, [35:27] like if if I had all my data within Snowflake or  Databricks, for instance, that would have solved [35:31] the problem, right? But we are going through  a transformation of building our lake house or [35:36] lake base these days as companies like Databricks  call, so we can bring as much of or like we can [35:42] try to solve for that fragmented system of records  as much as possible and put that intelligence [35:48] layer on top of it. So that way access to  information and then, you know, like when [35:53] when you have to drive from system of record to  system of action, it becomes an easy task for us. [36:00] Like Arjun, I'm in the same space, so Mode grew  through acquisition and we have six different [36:05] portfolio brands. We have our own proprietary TMS  systems, but we do work with Descartes, 2Ship, [36:13] Triton, and other TMS systems. Those were  built for specific modalities for specific [36:19] reasons in mind, right? Because the parcel side  light of business is completely tangential to your [36:25] international cargo, um, mode of transportation.  So, the needs are widely different, but like [36:34] Arjun mentioned, right? The integration is the  key challenge, but often times technology is not [36:40] the biggest hurdle. Um, you work with a customer  and you have, you know, the latest API, but [36:46] EDI is the only way to integrate. So,  there comes another COBOL, right? So, [36:51] the the process, the framework, um, you know,  the capabilities widely vary across your [36:57] customers and carriers. And, um, like Arjun  mentioned, like, you know, the lake base having [37:04] every data available in real-time streaming into,  um, you know, these systems can help benefit, [37:11] but the biggest challenge is how does  how widely different these data, um, [37:18] layers are. And that's why, you know, we tackle  one layer at a time because, um, but you you asked [37:25] about the TMS in the future, definitely there  is a place for it. The scope might be narrow, [37:31] depending on very specific use cases. And then you  will have agentic automations as a wrapper around [37:37] it to facilitate a lot of the workflow automation,  uh, but TMS is is still going to be there, um, [37:45] as your source of truth, as your, um, you  know, origination point for a lot of your, uh, [37:52] process and your data. But the integration and the  eventual landing space and from there, where you [37:59] can do the context layer and your, uh, intelligent  insights, that's drastically changing by the day. [38:08] I was thinking about this in your answer,  and then you mentioned it. Are we at the [38:11] place where AI can finally kill EDI, or are we  going to be stuck with it for 800 more years? [38:18] Uh, it's a very interesting question. I remember  seeing in the Data Bricks conference, one of those [38:22] flights with banners where SAS is dead. [laughter]  Similar question. These things die slowly. [38:30] Yeah, it's going to be dead by a thousand cuts. [38:33] Uh that's how it looks right now  because there's so much of appetite to [38:38] build your own applications to be able to like  customize it. Um but we will know in few years. [38:47] The other thing you said is the other three-letter  acronym you mentioned is MCP and MCP is a very, [38:53] very interesting landscape. I have a pretty strong  opinion that everybody should have one. The real [38:58] question is what you put in yours and what you  support and and how that changes over time and [39:04] that is sort of the integration strategic  landscape. That's where a lot of it's starting [39:08] to play out. Um we got a couple minutes. I want  to pull us out of this conversation by asking [39:14] what's on the next frontier? This is going  to be like a question for Puya for sure, [39:19] but I suspect all three of you have a perspective.  If we think two or three, four years out, what's [39:24] something that is like a little bit scary to think  about in terms of what we solve or a component of [39:30] the systems that we kill off or something that we  can do that we're not able to tackle right now. [39:37] I I can say that um you know, like primarily  in transportation logistics industry, it's very [39:44] relationship-driven. While um I love  working with my business partners [39:49] and then enabling digital tools, whether that  is automation, AI, advanced analytics tools, [39:55] uh the relationship they have with the customers  and they have with the carriers on both sides, [40:00] that's what moves this industry. So, um one  thing I'm cognizant and a little bit concerned, [40:07] but then hopefully we can solve this like how do  we maintain the relationship as we evolve on the [40:11] technology footprint and that requires that  you know, like you talked about EDI earlier, [40:17] there are still a lot of carriers who are still  stuck in the EDI world. We have to bring everybody [40:21] along with us. So, you can't work in isolation and  you have to work with your customers. In our case, [40:27] it is small business owners. Right, so everybody  has individual personalized needs. And on the [40:32] other front, work with your carriers, whether  that for parcel or freight or other modality, [40:37] where a lot of them are still stuck  in the legacy ways of working. So, [40:41] it's not a one-size-fits-all, but then  you have to slowly bring everybody along. [40:46] So, actually on the on the quantum front, one  thing that I've learned and we've learned from the [40:51] AI and its deployment is the pushback that came on  on has come become real on on AI and how it's like [40:58] the the hype that built around it is now receiving  lots of pushback. And that's something that we [41:03] on the quantum front, while we are deploying it,  we have to be cognizant of not over-hyping it to [41:09] the level that then the uh the pushback on it to  be too big. But, what I'm looking forward to is [41:16] uh with the adoption of all these uh new  technologies or for AI becoming more mainstream [41:21] and for quantum to be integrated as a new next  new front. Um is uh all of these at the end of the [41:29] day, what are value that they provide? The main  thing is that they provide I say one extra hour [41:35] for your 24-hour day, basically. So, it makes your  24 25 gives you make it a 25-hour day for you. So, [41:41] you you'll have more time to generate more value  and it's up to the end user how they want to use [41:46] it and how they want to get more productivity out  of it. So, that's what I'm looking forward to. [41:52] The transportation industry is highly fragmented,  uh but the beauty of it is a very tight-knit [41:57] industry. You see a lot of folks who've been  in the industry for 30 years, went on to build [42:03] amazing products um and services that they are  offering. Um so, definitely I see the future where [42:12] you're able to standardize a lot of these  workflows and automated with agent AI, [42:18] but still the core strength is the the knowledge  layer and you talk about ontologies and context [42:24] and everything. It's it's the relationship like  Arjun said the the knowledge of and how the inner [42:31] workings with specific types of industries and  customers that we work with that's really hard to [42:37] replicate and but we can make it better with the  AI and it can help augment and be the digital twin [42:45] that will help our agents and like now sales and  now our network make it much more better for the [42:53] future and that's kind of like more optimistic  way of looking at it I would say in my opinion. [43:00] That we can win through domain expertise and  relationships is I think an excellent place [43:04] to leave this. Please welcome in Please  join me in thanking our panelists today.