Truist, Suffolk Chiefs Rethink AI Strategy as Compute Costs Surge

算力成本激增，Truist 与 Suffolk 高管重新审视 AI 战略 企业领导者正从独立的软件工具转向基于成果的定价模式与人机协同的 AI 工作流，以弥合高昂算力支出与可衡量业务投资回报率（ROI）之间的差距。 该研讨小组探讨了《财富》250强企业及高度受监管行业如何成功整合并规模化应用生成式 AI。核心讨论聚焦于攻克定制化的“最后一公里”、采用基于成果的定价模式、权衡自建与外购策略，以及运用领域优先的架构。演讲嘉宾强调，AI 的目的并非取代人类员工，而是用于解决以往无法规模化扩展的运营短板，并推动员工转型为智能体工作流的人机协同编排者。 企业 AI 落地并非简单的自研与采购二选一，企业应掌握核心数据壁垒并打通最后 1 英里的定制化，而底层平台与基础设施则可选择租赁或采购。 (4:29) 初创企业若想快速切入大型企业市场，应作为模块化插件融入微软、Salesforce 和 Workday 等现有大型生态，而非试图替换它们。 (11:06) 由于高昂的实时推理计算成本，生成式 AI 的商业模式正逐渐超越传统按席位付费的 SaaS 模式，转向基于实际业务结果的定价机制。 (17:48) 企业采购 AI 软件通常基于 3 至 5 年的 NPV 模型进行评估，尽管前期有现金支出，但相比高昂的人力成本，数以亿计的 LLM 推理具有显著的长期成本效益。 (19:42) 随着 AI 从单纯工具走向全流程落地，企业员工正逐步转型为监督与协调 AI 智能体的指挥者，使企业在不缩减人力的前提下覆盖以往难以顾及的业务需求。 (36:38)

Scaling generative AI across Fortune 250 enterprises requires abandoning tool-first experimentation in favor of domain-led process re-engineering, hybrid build-and-lease infrastructure, and human-guided agent orchestration, according to enterprise leaders and technology founders at Ai4. Navigating the Build-Versus-Buy Dilemma and Last-Mile Integration (4:29) Enterprise technology leaders caution that adopting generative AI should not be treated as a binary choice between building in-house or buying commercial software. Steve Mahoney of Legion Intelligence emphasizes that organizations must retain ownership of their proprietary data moat and user experience while leasing or buying underlying platform infrastructure, indexing, and governance layers. While basic retrieval-augmented generation (RAG) prototypes are straightforward to assemble internally, maintaining 24/7 availability, scaling vector indexes, and enforcing continuous data governance drain specialized technical resources that could otherwise support core competitive advantages. Simultaneously, startups selling into regulated environments face strict integration and privacy constraints. Chandra Kapireddy of Truist Bank notes that financial institutions cannot freely export enterprise data to outside vendors for model training or evaluation. Consequently, external solutions achieve faster enterprise adoption when they act as modular extensions to incumbent platforms—such as Microsoft Copilot, Salesforce, and Workday—rather than standalone systems attempting to replace enterprise tech stacks. Domain-Forward Implementation and Pragmatic Task Automation (11:06) Rather than leading with technology demos, Aleksey Chuprov of Suffolk Construction advocates for a domain-forward framework: identifying high-value business units, re-engineering workflows end-to-end, deploying conventional deterministic automation first, and reserving generative AI strictly for complex, unstructured tasks. In construction engineering, 95% of design workflows involve technical placement—such as sizing HVAC ducts and calculating lighting wattage—rather than creative drafting. Suffolk coordinates discrete point solutions using an underlying agentic orchestration layer to prevent structural clashes. In banking, Kapireddy identifies operational efficiency and loss reduction as primary ROI drivers. High-impact areas include automated developer code generation, performance review drafting, and preliminary assembly of anti-money laundering (AML) reports, Know Your Customer (KYC) documentation, and commercial credit memos. Direct customer-facing revenue workflows, by contrast, maintain strict human-in-the-loop validation to satisfy compliance mandates. Inference Unit Economics and Outcome-Based Pricing Structures (15:13) The economic mechanics of enterprise generative AI diverge sharply from conventional SaaS due to ongoing compute overhead. Quinn Favret of Tavus observes that high-throughput multimodal and interactive video models incur direct marginal inference costs, running several cents per minute of real-time interaction. To bridge backend operational expenses with customer budgets, vendors increasingly look beyond rigid per-seat licensing. Mahoney and Favret argue that pricing models are shifting toward outcome-based frameworks aligned with specific client metrics, allowing AI vendors to tap into varied operational budgets in traditional industries like construction and banking. From the procurement standpoint, Kapireddy explains that corporate buyers evaluate AI tools against human capital expenses using 3-to-5-year Net Present Value (NPV) horizons. Because an annual salary of $100,000 matches the cost of billions of inference tokens, enterprises willingly absorb upfront implementation cash outflows if productivity compounding yields significant long-term returns. Workforce Adaptation, Latent Demand, and Human Orchestration (26:22) Panelists highlight that early enterprise rollouts often suffer steep post-launch usage declines when employees lack clear use cases or fear job redundancy. Chuprov counters this through a four-part change management model: executive KPIs on automation, bottom-up operational buy-in, targeted training, and reinforcement systems to track adoption and mitigate process drift. Favret adds that AI avatars and automated agents often target previously unserved operational demand—such as off-hours patient support or low-margin inquiries in healthcare—expanding operational capacity without direct staff reductions. To categorize enterprise automation, Kapireddy outlines a four-bucket agent model comprising human agents, deterministic rule-based systems (RPA), probabilistic generative models, and physical or robotic systems. The panel concludes that as autonomous tools mature from initial deployment surges into production stability, organizational structures will increasingly rely on human generalists serving as prompt orchestrators and auditors who monitor multi-agent workflows for compliance and visibility. The panel concludes that sustained enterprise generative AI adoption depends on moving past generic productivity assistants toward auditable, domain-specific systems that bridge infrastructure costs with clear business outcomes, while transitioning human workers into supervisory orchestrators across deterministic and probabilistic tools.

在由 Ai4 主办的研讨会上，来自金融、建筑和 AI 基础设施领域的高管指出，生成式 AI 在大型企业的落地已从简单的工具试用转向重塑业务流程的系统化整合，核心在于解决“最后一公里”定制、算力经济模型重构以及组织变革管理。 自研与外购抉择及最后一公里定制 (4:29) Legion Intelligence 产品副总裁 Steve Mahoney 认为，企业 AI 战略不应被视作非黑即白的“自研或外购”对立。数据构成了企业最核心的护城河，因此企业应当界定哪些技术层需要自主拥有、长期租赁或逐步采购。虽然通用的检索增强生成（RAG）或聊天原型易于搭建，但全天候运行、向量索引扩展与持续数据治理需要耗费高额工程资源，这些底层架构更适合采购。相反，企业必须集中内部资源攻克具体的“最后一公里”，即结合特定业务流程对领域工作流进行定制开发。 针对初创企业的切入方式，Truist Bank 负责 AI、机器学习和分析的 Chandra Kapireddy 提出，在受强监管的金融机构中，数据在未经严格合规审批前不能移出安全边界进行外部微调。初创企业如果试图替代现有的核心软件将面临极大阻力，更可行的商业化路径是作为模块化扩展，无缝嵌入微软 Copilot、Salesforce 或 Workday 等现有大型企业平台中，以降低合规审查摩擦。 流程优先转型与采购经济模型 (11:06) Suffolk Construction 数据分析与 AI 副总裁 Aleksey Chuprov 警示企业应避免“技术优先”的误区，即盲目引入新技术再去寻找应用场景。他主张采用以业务领域为先的四步路线：首先锁定高业务价值且易于推动组织变革的场景；其次对端到端流程进行重新梳理与再造；随后优先采用规则系统与传统数据集成解决基础自动化，仅将生成式 AI 用于处理复杂非结构化缺口；最后才根据具体定义评估工具采购或自研。在建筑工程中，95%的设计工作实为暖通管道尺寸计算、灯具瓦数分配等工程排布，公司正借助自主研发的智能体作为连接纽带，协同各专业点状工具以防系统冲突。 在采购与财务评估维度，Chandra Kapireddy 指出人力成本仍然是企业最大的支出项，例如一名年薪 10 万美元员工的支出相当于数以十亿计的模型推理 Token 成本。因此，银行采购团队会基于 3 至 5 年的净现值（NPV）模型评估长期回报，愿意承受前一两年的净现金流出以换取长期的复合运营效率提升。同时，Steve Mahoney 与 Tavus 联合创始人 Quinn Favret 补充道，AI 的实时 GPU 推理成本打破了传统 SaaS 的固定毛利模式，定价正在摆脱单纯的按席位或按算力收费，逐步转向与客户实际业务目标挂钩的基于成效定价模式。 破解采纳瓶颈与组织变革管理 (26:22) Steve Mahoney 观察到，过去半年多以来，大量企业级 AI 项目在经历初期爆发式增长后出现急剧的使用率下滑。导致这一停滞的核心原因在于员工在体验完初始“惊艳时刻”后缺乏持续落地的场景，或者由于担忧岗位被取代而产生抵触情绪。针对这一瓶颈，Aleksey Chuprov 提出了包含四个维度的组织变革管理框架：一是业务主管必须背负明确的 AI 渗透率与自动化考核 KPI；二是推动自下而上的认知转变，让员工成为智能体输出结果的评估者和指导者；三是开展标准作业程序培训；四是建立持续监控与防范流程偏离的强化机制。 在劳动力结构演变方面，Quinn Favret 与 Chandra Kapireddy 指出，AI 当前主要被用于承接以往因成本或人力瓶颈而无法覆盖的潜在业务需求，例如让医疗客服团队实现 24 小时不间断响应并处理低客单价咨询，从而在不裁员的前提下提升客户留存。Kapireddy 还提出了涵盖人类、确定性（RPA）、概率性（生成式 AI）和物理机器人四类智能体的分类框架，并指出企业角色正从单一领域的专精人员转向擅长提示词工程、上下文管理与多智能体监督的通用型“人类协调者”。 与会者一致认为，企业部署 AI 的重心已从 2025 年对自主智能体工具的探索，转向 2026 年以“人类作为协调者”为核心的治理体系，但如何在保障合规与审计透明度的同时规模化释放业务价值，仍是高度监管行业需持续验证的核心挑战。

## Background 本视频由 Ai4 于 2025 年 8 月 29 日发布于 YouTube 平台。视频记录了一场关于“在企业中大规模应用生成式 AI 的战略方法”的专家研讨会，多位来自金融、建筑及科技领域的资深高管与投资人共同探讨了生成式 AI 从概念验证（POC）走向企业级全面落地所面临的组织架构挑战、战略规划与最佳实践。 ## TL;DR 研讨会深入剖析了企业推动生成式 AI 大规模落地的核心路径与痛点。嘉宾们指出，企业应厘清自研与外购（Rent vs. Own）的动态平衡，将数据与核心业务流程作为自身护城河，同时借助成熟的基础设施快速实现业务价值。针对AI落地的实际挑战，与会者提出了领域优先（Domain-forward）的转型方法论、灵活的基于结果的定价模型，以及包含高层赋权与基层转型在内的四步走组织变革框架。未来，随着不同类型 AI 智能体的融合应用，人类将逐渐转变角色，成为掌控全局与协同多智能体运行的“关键协调者”。 ## Atmosphere & style 整场研讨会呈现出极其务实、专业且富有远见的探讨氛围。主持人与四位来自不同行业的嘉宾互动频繁、节奏紧凑，思想碰撞不断。讨论既不乏对底层技术与算力成本的深刻剖析，又充满了接地气的真实企业案例与财务实操算账。嘉宾们的分享理性而不失热情，既展现了对 AI 商业化与效率革新前景的强烈信心，也坦诚地揭示了企业在面对合规、组织抗性以及“人”的关卡时的复杂心态与真实挑战。 ## Key takeaways - [07:35](https://www.youtube.com/watch?v=aQUzwx53bxs&t=455s) Legion Intelligence 认为，AI 落地不应视作非黑即白的“自研还是外购”（Build vs. Buy），而应视为“租赁还是持有”（Rent vs. Own）：企业必须长期持有（Own）自身的数据护城河与核心业务流程，但对于高复杂度的通用底层基础设施，则应租用（Rent）成熟平台。 - [09:06](https://www.youtube.com/watch?v=aQUzwx53bxs&t=546s) Tavus 指出，AI 初创公司要推动企业客户买单，必须一方面通过真实用例展示“艺术的可能性”（Art of the Possible），另一方面提供极低门槛的试用，让用户在“分钟级”时间内感受到价值与 ROI。 - [11:08](https://www.youtube.com/watch?v=aQUzwx53bxs&t=668s) Suffolk Construction 倡导采用“领域优先”（Domain-forward）的 AI 转型方法论：首先锁定高价值且低阻力的核心业务领域并重构流程，优先使用传统数据集成与自动化手段，仅在必要时才选型引入 AI。 - [13:15](https://www.youtube.com/watch?v=aQUzwx53bxs&t=795s) 强监管的金融机构（如 Truist Bank）在引入 AI 时极度注重合规与数据安全，初创公司不能假设有自由测试的数据集，而应主动提供可与 Microsoft、Salesforce 等既有巨头生态深度融合的协议或插件。 - [01:16:18](https://www.youtube.com/watch?v=aQUzwx53bxs&t=4578s) 生成式 AI 产品的定价模式正从传统的 SaaS 按席位/人头计费，加速转向“价值与结果导向”（Outcomes-based），并逐渐演变成为驱动品牌知名度与渠道分发（Channel Partnerships）的战略杠杆。 - [01:19:42](https://www.youtube.com/watch?v=aQUzwx53bxs&t=4782s) 大型企业在评估 AI 项目的 ROI 时采用净现值（NPV）方法推演 3 到 5 年的财务模型；相比于极其高昂的人力成本，前期的 GPU 推理与 Token 费用在长期规模化应用中展现出巨大的经济效益优势。 - [01:21:34](https://www.youtube.com/watch?v=aQUzwx53bxs&t=4894s) 建筑业中 95% 的工程设计（如暖通、电路）过程极为繁琐，Suffolk Construction 通过构建“智能体编排层”（Agentic Connecting Tissue）作为粘合剂，将各单点 AI 串联以实现工程设计的全流程自动化。 - [01:21:34](https://www.youtube.com/watch?v=aQUzwx53bxs&t=4894s) Truist Bank 将金融业 GenAI 场景归纳为四类：收入最大化、损失最小化、生产力提升（如评估总结、辅助代码）和运营效率优化（如信用备忘录撰写），其中生产力与运营效率提升是最容易落地的突破口。 - [01:27:04](https://www.youtube.com/watch?v=aQUzwx53bxs&t=5224s) 企业在将 AI 推向生产环境后常遭遇使用率陡降（Drop-off）的“人”之瓶颈，Suffolk 提出了高层 KPI 归属、基层转变为 AI 监督者、系统化培训及持续强化机制的“四步走”变革管理框架。 - [32:03](https://www.youtube.com/watch?v=aQUzwx53bxs&t=1923s) AI 很大程度上填补了过去人类因成本或时间限制而无法覆盖的业务空白（如全天候客服与低价值工单处理），帮助企业在保持员工规模不变的情况下实现爆发式业务增长。 - [33:20](https://www.youtube.com/watch?v=aQUzwx53bxs&t=2000s) Truist 将企业智能代理分类为人类代理、确定性代理（RPA/逻辑）、概率性代理（GenAI/LLM）以及物理机器人代理，大型企业级应用往往需要将这四类代理进行协同组合。 - [33:38](https://www.youtube.com/watch?v=aQUzwx53bxs&t=2018s) AI 工具的普及推动组织招聘从极度专业化转向通用型人才（Generalist），人类的角色正从具体的执行者快速转变为掌控全局、协调多个 Agent 运行的“人类协调者”（Orchestrator）。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=aQUzwx53bxs&t=4s) 研讨会背景与与会嘉宾介绍 本次研讨会由 WndrCo 合伙人 Justin Wexler 主持，主题为“在企业中大规模应用生成式 AI 的战略方法”。WndrCo 是由 DreamWorks（梦工厂）创始人 Jeffrey Katzenberg 参与创立的风险投资机构，重点布局部署于 Fortune 250（财富 250 强）企业的科技初创公司。Justin 本人也是 Writer（本次活动赞助商之一）、Alyic 和 Nomi 等 AI 企业的董事会成员，长期致力于帮助初创公司在 Delta（达美航空）、American Express（美国运通）以及 Pfizer（辉瑞）等大型企业中落地。 在简短的开场后，四位来自业内不同领域的重磅嘉宾依次进行了自我介绍： * **Steve Mahoney**：Legion Intelligence 的产品副总裁。该公司源自美国国防部（DoD），旨在为企业提供构建 AI 的平台，帮助连接和治理非结构化及结构化数据，并安全、高效地部署智能体（agentic）工作流。 * **Aleksey Chuprov**：Suffolk Construction 的数据分析与 AI 副总裁。Suffolk 是一家大型建筑公司，承建住宅、商业地产、数据中心、高校和实验室等。公司将数据与 AI 视作行业变革的关键，经历了从建立数据仓库、商业智能（BI）、预测性机器学习到如今智能体自动化（agentic automation）的演进之路。 * **Quinn Favret**：Tavus 的联合创始人。Tavus 自 2020 年起便深耕该领域，致力于研究“人机自然交互”。其核心产品是“AI 数字人”（AI humans），能在类似于 Zoom 的虚拟会议室中以极具逼真度的方式与人类对话，广泛应用于销售、市场、客户成功、教育及招聘面试等场景。 * **Chandra Kapireddy**：Truist Financial（美国第七大银行，总部位于夏洛特）的 AI 与机器学习负责人。在加入 Truist 之前，他曾先后供职于 Amazon 和 JPMorgan Chase（摩根大通），拥有丰富的大数据与数据分析经验。 --- ### [00:57](https://www.youtube.com/watch?v=aQUzwx53bxs&t=57s) 从 DoD 基因到企业最后里: Legion 的落地哲学 针对企业级 AI 解决方案的落地，Steve Mahoney 分享了 Legion Intelligence 的发展历程与核心理念。由于公司起源于美国国防部（DoD），在 ChatGPT 问答爆火之前，他们就已经在无外网接入的环境中解决复杂的文档检索、报告生成与数据连接问题。 Steve 指出，过去企业级搜索和数据索引领域一直试图提供标准化工具，但每个企业都有其极具特色的“最后一公里”（last mile）需求。即使在传统的 BI 时代，套用标准仪表盘也是不够的。Legion 的策略是“承包最硬核的底座基础设施”，把数据连接、治理等复杂底层做好，同时把“最后一公里”的定制化能力交还给企业，让客户快速搭建符合自身业务流程的专用应用。 - **核心价值点**：帮助企业摆脱仅停留在“演示阶段”的临时 RAG 引擎或原型，将其平滑推向工业级应用。 - **技术积淀**：结合 Steve 过去 20 年在网络与数据基础设施方面的经验，将 AI 打造为连接数据与人类的全新桥梁。 --- ### [00:50](https://www.youtube.com/watch?v=aQUzwx53bxs&t=50s) Suffolk 的建筑数字化实践与 Tavus 的交互革命 Aleksey Chuprov 详细阐述了 Suffolk Construction 如何在传统实体建筑行业中落地 AI。建筑业的痛点在于项目交付的安全性、成本控制与工期。Suffolk 的 AI 演进路径十分清晰： ``` +------------------+ +------------------+ | 数据仓库建设 | --> | 商业智能 BI | +------------------+ +------------------+ | v +------------------+ +------------------+ | 智能体流程自动化 | <-- | 预测性 AI/ML | +------------------+ +------------------+ ``` 他们首先打牢数据基础，随后利用预测性机器学习分析历史趋势以降低项目交付风险，如今则聚焦于利用智能体（agents）重新定义施工流程，实现部分环节的端到端自动化。 与此同时，Tavus 联合创始人 Quinn Favret 则带来了纯技术前沿的视角。Tavus 在“AI 还没有成为热点”的 2020 年就已入局，主攻生成式视讯与对话 AI。通过打造具有人类外观、语言习惯和交互逻辑的数字人，Tavus 帮助企业将原本极其消耗人力资源的交互过程（如面试、导购、培训）实现规模化复制。 --- ### [07:35](https://www.youtube.com/watch?v=aQUzwx53bxs&t=455s) 自研还是外购？解析企业 AI 战略与“数据护城河” 主持人 Justin Wexler 抛出了一个极其现实的矛盾：很多非科技企业（如房产公司、银行）的 IT 团队在获得 AI 预算后，倾向于“重新发明轮子”，把所有基础设施从零开始自研，甚至排斥外部成熟供应商的介入。 对此，Steve Mahoney 提出了非常深刻的观点：**这不应该是一个非黑即白的“Build vs. Buy”（自研还是外购）问题，而是一个“Rent vs. Own”（租赁还是持有）的动态权衡。** * **数据是真正的护城河**：AI 让企业第一次感知到“数据即产品”。企业的核心竞争优势在于自身的数据、独特的业务流程以及客户体验，这些是企业必须**长期持有（Own）**的部分。 * **底层基础设施应当租赁（Rent/Lease）**：自研一个简单的 RAG 原型很容易，但要维持 24/7 高可用、处理复杂的权限治理、弹性扩容索引，需要消耗极其庞大的工程人力。企业如果将所有精力耗费在通用底座上，就会分散打造“数据护城河”的核心注意力。因此，正确的策略是在初期租用成熟平台，待业务成熟后再评估是否收回自研。 --- ### [09:06](https://www.youtube.com/watch?v=aQUzwx53bxs&t=546s) 降低交付门槛：“艺术的可能性”与“分钟级见效” Quinn Favret 从初创公司视角补充了企业对 AI 采纳的痛点。他坦言，如今在企业内部建立一套全新的 AI 系统门槛极高——不仅招聘顶尖人才困难，确定战略路线也充满不确定性。因此，绝大多数企业最终依然会选择购买现成方案。 要让企业买单，初创公司必须做好两件事： 1. **展示“艺术的可能性”（Art of the Possible）**：大家普遍体验过 Microsoft Copilot 等通用工具，虽然概念很好，但在具体业务场景中往往效果有限。初创公司需要通过极具说服力的真实用例，帮助客户拓宽想象力，明确 AI 在特定场景下的巨大价值。 2. **极速的价值实现时间（Time to Value）**：展示完愿景后，必须提供极低门槛的试用与接入体验。不能让客户等上几个月或几周，而是要让使用者在**几分钟内**就能登录、上手并体验到“神奇时刻”（magic moment），直接看到明确的 ROI 指标。 - **用例聚焦**：拒绝“火烧连营”式的全面推翻重来，而是锁定初创公司最擅长、ROI 最明确的细分场景。 --- ### [11:08](https://www.youtube.com/watch?v=aQUzwx53bxs&t=668s) 业务领域驱动的 AI 转型方法论 面对繁杂的 AI 工具，Aleksey Chuprov 提醒企业切忌“本末倒置”。很多企业常常看到某个 Cool（炫酷）的 AI 工具，就试图在公司内部为其寻找应用场景，这种做法很难带来真正的战略变革。 Aleksey 建议企业采用**领域优先（Domain-forward）**的转型方法论： 1. **锁定核心领域**：不要试图在全公司范围内修修补补，而是选定一个在“业务价值”与“实施难度/变革阻力”之间取得最佳平衡的业务领域。 2. **重新梳理端到端流程**：明确现有流程的痛点，设计出未来的重构蓝图。 3. **阻力最小路径原则**：如果简单的“数据集成”或“传统自动化”就能解决问题，优先使用传统手段；只有在传统手段无法满足时，才引入 AI 技术。 4. **最后才选型工具**：在明确了具体的 Use Case（用例）后，再评估是自行开发还是采购市面上的现有工具。 --- ### [13:15](https://www.youtube.com/watch?v=aQUzwx53bxs&t=795s) 金融巨头的考量：合规、安全与生态融合 作为传统强监管行业（金融业）的代表，Truist Bank 的 Chandra Kapireddy 分享了大型银行在引入 AI 初创公司时的核心关切。他直言，初创公司带给银行业的创新激情令人兴奋，但要在银行落地，必须理解大型机构的特殊文化与合规要求。 Chandra 给初创公司提出了几点中肯建议： * **切勿假设数据可以自由用于测试**：在严格监管的环境下，敏感数据的提取与测试受到极严格的保护，初创公司不能假设有现成的数据集可供随意跑模型。 * **注重与现有巨头生态的融合**：像 Truist 这样的银行，内部已经深度集成了 Microsoft（如 Copilot 生态）、Salesforce、Workday 等大型平台。初创公司如果能够提供可扩展这些既有平台的协议或插件，而不是试图替换它们，将大大降低接入门槛，从而更容易进入大型银行的供应链。 --- ### [15:13](https://www.youtube.com/watch?v=aQUzwx53bxs&t=913s) AI 时代的定价难题与渠道合作 研讨会随后转向了 AI 初创公司面临的商业化挑战。主持人 Justin Wexler 提到，目前 AI 技术的定价模式正在经历剧烈演变，同时初创公司与 Big Tech（科技巨头，如 Microsoft、Google、Salesforce）之间既存在竞争，又存在渠道合作的复杂关系。 Tavus 的 Quinn Favret 坦言：“在 AI 领域，定价真的非常困难。”不同于传统 SaaS 软件近乎零边际成本的扩容，AI 应用需要运行推理（Inference）、调用 GPU 资源，有着极高的硬件与算力成本。以 Tavus 的视频数字人技术为例，每运行一分钟就会产生数美分的直接硬性成本。如何在保证自身毛利率的同时，给企业客户提供清晰、合理且具备规模效应的定价模型，是所有 AI 初创公司必须解决的核心命题。 <ElicitationsGroup message="想要深入探索本次研讨会提及的生成式 AI 企业落地策略？你可以选择以下方向："> <Elicitation label="分析‘Build vs. Buy’在企业 AI 落地中的决策框架" query="请详细分析企业在部署生成式 AI 时，如何评估自研（Build）与采购（Buy）的决策框架，包括数据安全、基础设施成本与长期 ROI 维度。"/> <Elicitation label="了解传统行业（如建筑或金融）的 AI 智能体应用案例" query="请结合 Suffolk Construction 和 Truist Bank 的案例，详细说明 AI 智能体（Agentic AI）在传统行业和强监管行业中的落地实践与挑战。"/> <ElicitationsGroup>` ### [01:16:18](https://www.youtube.com/watch?v=aQUzwx53bxs&t=4578s) 商业化模式演进：从 SaaS 人头计费转向“价值与结果导向” 在讨论生成式 AI 产品的定价与商业化策略时，**Quinn Favret**（Tavus）指出，当前的定价正经历一次深刻的范式转移。传统的 SaaS 模式通常依赖于按席位（Seat-based）或订阅费计费，而现在 AI 企业的定价机制越来越注重与**实际创造的价值（Value Alignment）**紧密贴合。AI 产品的运行机制包含明确的底线成本，如何将这些复杂的底层成本转化为客户能够清晰理解、且能无缝融入其现有采购生态系统的定价模型，是 AI 创业公司必须承担的翻译责任。 定价不再仅仅是保护 SaaS 毛利率（SaaS Margins）的财务工具，它已跃升为推动品牌知名度、增长飞轮以及渠道分发（Channel Partnerships）的核心战略杠杆。通过巧妙地融入现有生态渠道，AI 产品能以更低的门槛获得初期市场渗透和品牌认知。 ``` 传统 SaaS 定价机制 生成式 AI 新型定价机制 ┌─────────────────────────┐ ┌─────────────────────────┐ │ • 按席位/人头计费 │ ─► │ • 紧密贴合业务价值与结果 │ │ • 关注固定毛利率 (Margin)│ │ • 灵活的底层算力/Token换算│ │ • 模式僵化，脱离实际产出│ │ • 成为品牌与渠道分发杠杆 │ └─────────────────────────┘ └─────────────────────────┘ ``` **Steve Mahoney**（Legion Intelligence）从公共部门与企业端经验补充道，政企客户的市场拓展完全是另一个“独立的异类世界”，通常需要完全不同的团队和流程去应对。然而，现在也是产品走向市场（GTM）和定价策略最富创造力的时代。以往超大规模云厂商（Hyperscalers）按算力（Compute）计费、软件厂商按人头计费，这些指标往往与客户获得的最终价值存在断层。在当前 AI 发展早期，企业拥有极大的空间去探索**基于业务结果（Outcomes-based）**的灵活定价——例如设定基础保底费用以覆盖平台成本，再根据实际交付的业务成效进行浮动结算。这种灵活性使得 AI 厂商能够随时跟进不同行业（如建筑业或金融服务业）的预算分配结构。 - **直销与渠道组合**：Quinn 透露，虽然渠道合作（Channel Partnerships）是敲开客户大门、建立初期信任的良好补充，但对 Fortune 500 等大型企业而言，**直销（Direct Sales）**依然是目前最核心、最有效的增长引擎。 - **动态预算适应性**：由于各行业（如金融、建筑）对 AI 的采购预算可能分布在不同的部门，灵活的定价结构让厂商能够“随预算所在而变”，打破了传统软件采购的固有藩篱。 --- ### [01:19:42](https://www.youtube.com/watch?v=aQUzwx53bxs&t=4782s) 采购端的 ROI 考量与算力/人力成本博弈 **Chandra Kapireddy**（Truist Bank）从大型金融机构采购和审批者的视角拆解了 AI 技术的成本逻辑。传统企业在评估项目成本时，通常将其拆分为**软件、硬件与人力**三大块，其中人力成本永远是企业最昂贵的支出。面对生成式 AI，虽然前期存在昂贵的 GPU 硬件与推理成本，但如果与人力成本进行横向对比，AI 的经济效益便极其显著。 Chandra 现场算了一笔账：在企业内部，雇佣一名年薪 $100,000 美元的普通员工所产生的成本，如果转化为生成式 AI 的 Token 推理费用，足以运行**数以百亿计（Billions of Tokens）**的模型推理。因此，当 AI 初创公司向大型企业推销时，企业采购部门首要关注的是**基于结果的 ROI 计算**。 ``` 企业内部成本评估方程 (NPV 视角) 初始投入 (Year 1-2) 中长期回报 (Year 3-5) ┌───────────────────────┐ ┌───────────────────────┐ │ • AI 平台与工具采购 │ ──────► │ • 显著减少人力工时 │ │ • 系统的集成与调优 │ 现金流净 │ • 业务流程自动化运营 │ │ • 净现金流流出 (Outflow)│ 转正 │ • 净现金流流入 (Inflow) │ └───────────────────────┘ └───────────────────────┘ ``` 为了评估 AI 项目的可行性，大型企业通常会采用**净现值（NPV）计算法**，进行 3 到 5 年的长期财务推演： - **前 1-2 年**：处于系统搭建与试错期，呈现净现金流出（Cash Outflow）。 - **第 3-5 年**：随着规模化应用与效率提升，产生显著的净现金流入（Cash Inflow）。 - **长期主义战术**：有了清晰的长期财务模型，采购部门就不会在项目初期因为短暂的成本波动而纠结，从而为 AI 的落地争取到宝贵的时间窗口。 --- ### [01:21:34](https://www.youtube.com/watch?v=aQUzwx53bxs&t=4894s) 行业突破口：建筑业设计自动化与金融业四大 AI 场景 在探讨 AI 落地最成熟的业务场景时，**Aleksey Chuprov**（Suffolk Construction）与 **Chandra Kapireddy** 分别分享了建筑与金融领域的颠覆性实践。 #### 建筑业：从“炫酷效果图”到“复杂工程设计自动化” Aleksey 揭秘道，大众谈到建筑设计时，脑海中浮现的往往是建筑师画的效果图（Renderings）。但实际上，**95% 的设计工作量完全不在建筑外观上，而是极其枯燥的工程系统设计**——例如在一个房间内如何配置灯具的位置与功率、暖通空调（HVAC）管道与格栅的尺寸计算，以及数以百计相互交织的建筑子系统设计。过去这些全靠工程师人工计算与绘制，而 Suffolk Construction 目前正通过 AI 实现这些全流程的自动化。 - **点解决方案（Point Solutions） vs. 智能体胶水（Agentic Architecture）**：初创公司往往擅长做单点突破（如专做照明设计或专做冷却系统的 AI），而大企业的核心任务是将这些单点工具串联起来。Suffolk 构建了一套**智能体编排系统（Agentic Connecting Tissue）**，作为不同单点 AI 之间的“粘合剂”，自动在不同系统间传递数据并协调依赖关系，从而快速生成多套互不冲突的完整设计方案。 ``` ┌────────────────────────────────────────┐ │ 智能体编排层 (Agentic Tissue) │ └───┬────────────────┬───────────────┬───┘ │ │ │ ▼ ▼ ▼ ┌────────────┐ ┌────────────┐ ┌────────────┐ │ 照明设计 │ │ 暖通系统 │ │ 电路管道 │ │ (Point) │ │ (Point) │ │ (Point) │ └────────────┘ └────────────┘ └────────────┘ ``` #### 金融业：生成式 AI 的四大应用场景桶（Buckets） Chandra 将金融机构的 AI 与生成式 AI 用例归纳为四大核心范畴： 1. **收入最大化（Revenue Maximization）**：如智能获客与交叉销售。由于直接面向客户，在没有“人机协同”（Human-in-the-loop）的情况下存在合规风险，目前全自动化推行的谨慎度最高。 2. **损失最小化（Loss Minimization）**：重点应用于反欺诈（Fraud Detection）与商业邮件欺诈（BEC）的风控识别。 3. **生产力提升（Productivity Improvements）**：**最容易落地、无需犹豫的突破口**。例如辅助程序员写代码，若运用得当可直接节省数百万美元；又如员工每半年/年度的绩效评估（Performance Evaluations），原本极其耗时的人工总结，利用 Gen AI 处理只需消耗原本 **1/10 的时间**。 4. **运营效率优化（Operational Efficiencies）**：金融业极具潜力的领域。涉及大量的反洗钱（AML）、了解你的客户（KYC）审核，以及为每个商业客户撰写极度消耗人力的**信用备忘录（Credit Memos）**，生成式 AI 在这些重度依赖文档处理的岗位上展现出惊人的替代效益。 --- ### [01:27:04](https://www.youtube.com/watch?v=aQUzwx53bxs&t=5224s) AI 规模化落地的“人”之关卡与四步走管理框架 谈及 AI 工具引入后对组织架构与员工角色的冲击，**Steve Mahoney** 指出，企业在部署 AI 时遇到的最严峻挑战往往不是技术本身，而是“人”的因素。 在过去半年与近 50 家企业的交流中，Steve 发现绝大多数企业都成功将 AI 项目推到了生产环境（Production），但随后往往会遭遇陡峭的**曲线跌落（Drop-off）**。员工在经历初期的“惊艳时刻”（Magic Moment）后，要么因为缺乏具体业务场景的结合而放弃使用，要么因为**亲眼看到 AI 正在吞噬自己的岗位**而产生抗拒与恐惧。如果企业没有提前规划好对“人”的处理方案，AI 规模化落地必然触壁。成功的企业在推行 AI 之初就会制定清晰的计划：例如明确将获得的生产力转化为减少 20%-30% 的招聘预测，或是将员工技能升级（Upskill）以承担更高阶的任务。 ``` [生产力爆发期] ──► [技术/组织瓶颈] ──► [使用率骤降 (Drop-off)] │ ▼ 必须引入“以人为本”的变革策略： ┌──────────────────────────────────────┐ │ 1. 高层明确 KPI 与自动化渗透率 │ │ 2. 基层转变为“AI 监督者与反馈者” │ │ 3. 系统化的技能重塑与应用培训 │ │ 4. 强化机制：监控使用率与合规性 │ └──────────────────────────────────────┘ ``` 针对这一挑战，**Aleksey Chuprov** 提炼出了一套**AI 规模化落地的四步走管理框架**： - **高层支持与责任归属（Leadership Buy-in & Ownership）**：业务部门领导（BU Leaders）必须承担明确的 KPI，将 AI 在该部门的渗透率和自动化比例直接纳入业绩考核。 - **自下而上的认同与角色转变（Bottom-up Buy-in）**：基层员工必须清晰意识到新标准作业程序（SOP）的价值。员工的角色不再是手动执行重复性工作，而是转变为**批判性审视 AI Agent 输出结果的“监督者”**，通过给 AI 提示与反馈来迭代优化产出。 - **持续的技能培训（Training）**：建立常态化培训机制，确保团队掌握工具的正确使用方法。 - **强化与监督机制（Reinforcement Mechanisms）**：建立完备的管理机制，持续衡量 AI 的实际使用率与价值捕获（Value Capture），及时发现不合规行为并给予纠正反馈。 ### [32:03](https://www.youtube.com/watch?v=aQUzwx53bxs&t=1923s) AI 补充人类空白：从客服案例看企业效率提升 Quinn Favret 指出，生成式 AI 的引入虽然提升了效率，有时也会降低招聘预算，但大部分应用实际上是被填补到了人类此前**无法覆盖或无法按比例扩展**的空白领域。他举例说明：一家总部位于纽约 Healthcare 公司，其客户成功团队希望每位用户都能更早被激活、获得升售并在遇到支持工单时被有效转化。然而，团队无法做到 24 小时全天候有人在线。如果用户在半夜遇到问题却无法联系到人工支持，往往会选择直接取消订阅。此外，由于很多低价值工单在交易金额上无法承担人工处理的高昂时间成本，这些需求过去一直被忽略。 AI 的出现让这些兼具经济价值却无法通过扩充人力来解决的场景变得可行。企业与初创公司正通过这种方式，在保持现有员工规模不变的同时，借由 AI 实现业务的爆发式增长与效率提升。 ### [33:20](https://www.youtube.com/watch?v=aQUzwx53bxs&t=2000s) 消费级 AI 的倒逼与 TRUIST 的代理四分类法 Chandra Kapireddy 提到，当前企业正处于一个非常特殊的时期：**消费级技术的发展速度超越了企业级技术**，这给企业 IT 团队带来了巨大的交付压力。TRUIST 拥有 4 到 5 万名员工，大家在手机上都能随时使用 ChatGPT，因此自然会质问公司为什么内部不能提供同样的体验。为了应对这一需求，企业必须推动 AI 的民主化，让每名员工都掌握使用工具的能力。 他观察到一个有趣的心理现象：即使员工知道某个问题的答案（比如给老板写邮件回复或撰写 RFP），也会倾向于先让 GPT 类的工具跑一遍，看看 AI 能否提供更好的版本。这种“先交由 AI 检查”的习惯已成为事实上的标准，意味着未来每位员工都会拥有属于自己的智能代理（Agent）。 Chandra 将企业中的 Agent 划分为四大类（并幽默地表示谁能提出第五类就给谁 100 万美元）： - **人类代理（Human Agents）**：人类依然至关重要，永远不会消失。 - **确定性代理（Deterministic Agents）**：基于传统的 `if-then-else` 逻辑，即 RPA（机器人流程自动化），也是目前绝大多数软件的构建方式。 - **概率性代理（Probabilistic Agents）**：基于 AI 的智能代理，具备概率推断能力。 - **物理机器人代理（Robotic Agents）**：实体形态的机器人，用于处理物理世界中的任务（如拆封实体信件和支票）。 在实际业务应用中，企业需要组合使用这几类 Agent。他和 TRUIST 的 CIO 更喜欢将 AI 定义为**增强智能（Augmented Intelligence）**，而非单纯的“人工智能”。 - 员工心理变化：即使已知答案，也会倾向于先用 AI 生成/校验一次，形成新的工作流习惯。 - 架构组合：构建复杂企业级应用通常需要结合人类、确定性、概率性及物理机器人等多种 Agent 协同工作。 ``` +-------------------+ | Enterprise Agents | +---------+---------+ | +------------------+--------+--------+------------------+ | | | | v v v v +--------------+ +---------------+ +--------------+ +---------------+ | Human Agent | | Deterministic | | Probabilistic| | Robotic Agent | | (Employees) | | (RPA/Logic) | | (GenAI/LLM) | | (Physical) | +--------------+ +---------------+ +--------------+ +---------------+ ``` ### [33:38](https://www.youtube.com/watch?v=aQUzwx53bxs&t=2018s) 组织架构演变与“人类协调者”时代的到来 与会者探讨了 AI 对组织结构和岗位职责的深远影响。当前企业在招聘时正呈现出**从极度专业化转向通用型人才（Generalist）**的趋势。通用型人才借助 AI 工具能够发挥出远超以往的效能。未来的重点不再仅仅是观察员工总数的变化，而是观察**组织架构、岗位职责与工作流程**如何进化。 以 sales engineers 等岗位为例，员工的工作重心正转移到上下文管理、脚本撰写和提示词工程（Prompt Engineering）上。人类的角色正在从具体的执行者转变为**协调者（Orchestrator）与管理者（Operator）**。业内因此流传着一个生动的说法：**2025 年是 Agent 之年，而 2026 年则是“人类协调者”之年**——人类站在中央掌控全局、协调多个 Agent 运行，从而在保持企业审核能力（Auditability）和透明度（Visibility）的前提下，实现前所未有的规模化效率。 - 招聘趋势：企业越来越倾向于招聘善于利用 AI 工具的通用型人才（Generalists）。 - 岗位转型：技术销售与运营人员逐渐演变为上下文与提示词的调控者。 - 治理保障：通过人类作为 Orchestrator 的把关，确保 AI 自动化流程的可审计性与透明度。 ## Notable quotes > [07:35](https://www.youtube.com/watch?v=aQUzwx53bxs&t=455s) 数据是真正的护城河 — Steve Mahoney > [07:35](https://www.youtube.com/watch?v=aQUzwx53bxs&t=455s) 租赁还是持有 — Steve Mahoney > [09:06](https://www.youtube.com/watch?v=aQUzwx53bxs&t=546s) 艺术的可能性 — Quinn Favret > [09:06](https://www.youtube.com/watch?v=aQUzwx53bxs&t=546s) 价值实现时间 — Quinn Favret > [11:08](https://www.youtube.com/watch?v=aQUzwx53bxs&t=668s) 领域优先 — Aleksey Chuprov > [01:16:18](https://www.youtube.com/watch?v=aQUzwx53bxs&t=4578s) 实际创造的价值 — Quinn Favret > [01:16:18](https://www.youtube.com/watch?v=aQUzwx53bxs&t=4578s) 基于业务结果 — Steve Mahoney > [01:19:42](https://www.youtube.com/watch?v=aQUzwx53bxs&t=4782s) 基于结果的 ROI 计算 — Chandra Kapireddy > [01:19:42](https://www.youtube.com/watch?v=aQUzwx53bxs&t=4782s) 净现值（NPV）计算法 — Chandra Kapireddy > [01:21:34](https://www.youtube.com/watch?v=aQUzwx53bxs&t=4894s) 智能体编排系统 — Aleksey Chuprov > [01:27:04](https://www.youtube.com/watch?v=aQUzwx53bxs&t=5224s) 曲线跌落 — Steve Mahoney > [01:27:04](https://www.youtube.com/watch?v=aQUzwx53bxs&t=5224s) 监督者 — Aleksey Chuprov > [33:20](https://www.youtube.com/watch?v=aQUzwx53bxs&t=2000s) 增强智能 — Chandra Kapireddy > [33:38](https://www.youtube.com/watch?v=aQUzwx53bxs&t=2018s) 2025 年是 Agent 之年，而 2026 年则是“人类协调者”之年 ## Mentioned Justin Wexler, Steve Mahoney, Aleksey Chuprov, Quinn Favret, Chandra Kapireddy, WndrCo, DreamWorks, Legion Intelligence, Suffolk Construction, Tavus, Truist Bank, Truist Financial, Writer, Alyic, Nomi, Delta, American Express, Pfizer, Amazon, JPMorgan Chase, Microsoft, Microsoft Copilot, Salesforce, Workday, Google, RAG, Agentic AI, RPA, NPV ## Video description Chandra Kapireddy, Firmwide Head of Gen AI, ML and Analytics, Truist Bank Justin Wexler, General Partner, WndrCo Quinn Favret, Tavus Aleksey Chuprov, VP Data Analytics and AI, Suffolk Construction Steve Mahoney, VP, Product, Legion Intelligence Abstracts Scaling generative AI such as LLMs and multimodal AI from POC to enterprise-wide deployment requires not just technical readiness, but strategic vision. This session will focus on the broader organizational challenges and opportunities of integrating AI at scale. Panelists will discuss theories, best practices, and innovative approaches to embedding AI systems across business units to drive value. Explore how companies are navigating cross-functional alignment, ensuring leadership buy-in, and fostering a culture that supports widespread AI adoption. Learn about the strategic frameworks that can maximize the impact of generative AI on a large scale. [Source](https://www.youtube.com/watch?v=aQUzwx53bxs)

[7:35] Steve Mahoney: 数据是真正的护城河 [7:35] Steve Mahoney: 租赁还是持有 [9:06] Quinn Favret: 艺术的可能性 [9:06] Quinn Favret: 价值实现时间 [11:08] Aleksey Chuprov: 领域优先 [76:18] Quinn Favret: 实际创造的价值 [76:18] Steve Mahoney: 基于业务结果 [79:42] Chandra Kapireddy: 基于结果的 ROI 计算 [79:42] Chandra Kapireddy: 净现值（NPV）计算法 [81:34] Aleksey Chuprov: 智能体编排系统 [87:04] Steve Mahoney: 曲线跌落 [87:04] Aleksey Chuprov: 监督者 [33:20] Chandra Kapireddy: 增强智能 [33:38] 2025 年是 Agent 之年，而 2026 年则是“人类协调者”之年

[17:48] Steve Mahoney: This is like one of the most unique times in the world in terms of pricing and strategy around bringing a product to market... we get to be creative. [33:20] Chandra Kapireddy: If you tell me the fifth bucket, I'll probably give you a million dollars

[0:04] Hello, great to see everybody. I'm [0:06] Justin Wexler, general partner at [0:07] Wonderco. Uh we're a venture capital [0:10] firm started by DreamWorks founder [0:11] Jeffrey Katzenberg. So, uh enterprise AI [0:14] is a huge focus for us. We invest in [0:16] technology startups that are deployed at [0:18] Fortune 250 enterprises. I'm on the [0:21] board of Writer, which is one of the [0:23] sponsors of this event, as well as Alyic [0:25] and Nomi. So spend a lot of time with [0:27] these founders helping them scale in [0:29] fortune 250 environments like Delta and [0:33] American Express and Fizer and many [0:35] others. So this is a topic very near and [0:38] dear to my heart. We have some great [0:39] panelists here that are from industry [0:41] both at larger organizations deploying [0:43] AI as well as representing technology [0:46] startups that are getting deployed in [0:48] enterprise. So we'll go right down. love [0:50] to hear a little bit about each of you, [0:52] your names, the company you work for, [0:54] uh, and then we'll go on to different [0:55] topics from there. [0:57] >> Sweet. Yeah. So, I'm Steve Mahoney. I am [0:59] the VP of product for a company called [1:00] Legion Intelligence. Uh, so we're born [1:03] from the DoD and we deliver a platform [1:05] that enables organizations to build AI [1:09] within their enterprise. So they have a [1:11] lot of the tools to connect and govern [1:13] all of their uh unstructured and [1:15] structured data within their [1:16] organization and then build [1:18] applications, build workflows, agentic [1:20] workflows and deploy them securely, [1:22] safely, efficiently within the [1:23] enterprise. Um I've spent the past 20 [1:26] years building data infrastructure, [1:28] network infrastructure, and this is yet [1:30] another platform that uh is taking [1:32] advantage of data and bringing it to [1:34] humans. So yeah, great. [1:37] >> Hi everyone, my name is Alex Chuprof. I [1:39] am uh VP data analytics and AI at SAFK [1:43] construction. So we built agents but we [1:46] also built real buildings. So we are [1:49] construction company. Uh we build [1:51] residential and uh commercial real [1:53] estate, we build data centers, uh [1:56] universities, uh labs and many many [2:00] other types of assets. And we are [2:02] strategically betting on data and AI to [2:06] be the gamecher for the industry for us [2:09] to deliver the projects safer uh more [2:13] cost efficiently and uh faster. So our [2:17] journey started probably like for many [2:19] of you from u building the data house in [2:23] order right so then uh building the [2:25] business intelligence on top of that and [2:28] but then after that we started uh [2:31] investing a lot in predictive AI. So how [2:35] do we leverage machine learning to learn [2:37] from the past trends and uh to reduce [2:40] the risk on uh delivery of our projects [2:44] and then a lot of uh today a lot of our [2:48] action is in agentic world. How can we [2:52] actually re uh re reimagine the how [2:55] construction is being done and uh [2:57] automating the process flows uh [3:00] sometimes end to end. [3:04] Howdy friends. My name is Quinn Favre. [3:05] I'm the co-founder of Tavis. So, we've [3:08] been uh building in the space since [3:10] about 2020, long since before AI was [3:12] cool. We're a research lab focused on [3:14] teaching machines how to interact and [3:17] talk to humans just how humans talk to [3:19] humans. And so the way that that [3:20] actually materializes is this concept of [3:23] AI humans that you can have in a Zoom [3:25] room equivalent where it looks like a [3:27] human, it talks like a human, it [3:29] converses like a human and actually [3:31] ultimately interacts like a human at the [3:33] end of the day. And these applications [3:34] are deployed across the board from sales [3:36] and marketing to customer success to [3:38] education all the way to recruiting and [3:40] interviewing. So it's been quite an [3:42] exciting journey and excited to talk [3:43] more about it today. [3:45] >> All right. Um my name is Shandra Capert. [3:48] I'm the head of AI uh and machine [3:50] learning at uh Truist Ris Financial [3:53] Services is the seventh largest bank in [3:56] United States. Um the headquarters in [3:58] Charlotte. So previous to joining Truis [4:01] I've been with Trace for one year. Uh [4:03] previous to TR I was with JP Morgan [4:05] Chase and previous to that I was with [4:08] Amazon and uh so my entire life has been [4:12] with the data and analytics. The only [4:14] thing that has changed is the definition [4:16] of analytics and of course the data uh [4:19] as you all know has fundamentally [4:21] changed. Uh and uh so I'm definitely [4:24] looking forward to this uh fantastic [4:26] panel conversation that we are having [4:28] here. [4:29] >> Great. Okay. So great to have you all [4:31] here. I want to start with Steve and [4:34] Quinn. So you both represent [4:36] technologies that are extremely powerful [4:39] and can provide a ton of value in [4:40] enterprise. They're also pretty uh [4:43] horizontal solutions that could you know [4:47] add value to a number of stakeholders, a [4:49] number of different departments within [4:50] organizations and are multifunctional by [4:53] their kind of definition. So we'd love [4:55] to have a you know consideration on [4:58] really how do you get started where do [5:00] you try to sell into uh and how do you [5:02] show value one part of the org and then [5:04] scale it across an enterprise. [5:06] >> Yeah. Yeah. A lot of the bet that so uh [5:10] Legion has been around for about 3 years [5:12] or so and as I mentioned we were born in [5:13] the DoD and a lot of that was you know [5:16] before ChachiBT especially was and [5:18] that's not been available inside of the [5:20] DoD. How do we take these concepts and [5:23] ideas and apply them to the workforce [5:25] and the you know the people that are [5:27] trying to push papers yes but do [5:29] reporting and do search etc etc. And a [5:32] lot of the the thesis was we can build [5:36] generalized tools and we have for years [5:37] for enterprise search and indexing and [5:39] you know uh delivering user experiences [5:42] but there's like a last mile that is [5:45] always custom to every single [5:46] organization. And this is true back in [5:48] the data days as well. You can build [5:50] dashboards and libraries and have all of [5:52] these canned solutions, but the reality [5:54] is there's this like last mile that's [5:56] really important and unique to every [5:57] organization. And so a lot of what we [6:00] have done is cover the hard stuff, [6:02] right? Like what's the hard stuff and [6:04] we'll talk about this when we get into [6:05] scale even more is like think about the [6:07] things that you know an organization it [6:09] doesn't make sense for them to want to [6:11] build uh and give them tools so they can [6:13] deliver that last mile. Give them the [6:15] ability to you know customize and and [6:17] tailor and and build those use cases [6:19] specifically for them. So from like our [6:21] perspective in how we build and sell the [6:23] product, that's how we engage is uh [6:25] let's talk about those specific [6:27] workflows and how fast you can actually [6:28] put them together and solve that very [6:30] unique problem. Uh and compare that to [6:33] the really cool chatbt prototype or rag [6:36] engine prototype that we built [6:37] originally in this company. Um they're [6:39] they're miles apart in terms of how they [6:41] get used at scale. [6:43] >> Just for one followup before we [6:44] continue. So sounds extremely powerful [6:47] and you're providing tons of value to [6:48] great enterprises and in the really [6:51] regulated government space [6:53] at the same time at least in my seat as [6:56] an enterprise AI investor we've seen [6:58] organizations I want to hear the [7:00] perspective from folks in industry too [7:02] uh where tech teams really want to [7:05] embrace building everything even the [7:07] hard stuff right when maybe their [7:09] organization they're not a tech company [7:10] maybe they're a real estate company [7:11] maybe they're a bank right and yet uh AI [7:14] has given free license to just say this [7:17] is going to take years to build but I'm [7:19] going to you know have like road map for [7:21] the next few years off of all the work [7:24] like you coming in and actually doing it [7:25] is a negative for me as a tech leader or [7:28] a tech you know um you know [7:30] professional. So how do you respond to [7:32] that when you're looking to help these [7:33] organizations holistically? [7:35] >> Yeah, it's I mean straightforwardly it's [7:37] not a build versus buy. That's you're [7:39] touching on the build versus buy like [7:40] everybody like you should want to own [7:42] your AI strategy. You should want to own [7:43] AI. Why? Because it is the closest thing [7:46] that we've ever had to the data is the [7:48] product and and it is very much that is [7:50] your special sauce as an organization. [7:52] That's your moat is your data within [7:54] your organization. So the question [7:55] shouldn't be build versus buy. It should [7:57] be where do I want to own right now and [7:59] where do I want to lease and then [8:01] consider where I buy down the road. [8:03] Right? So it's the strategy of like how [8:05] do I set a plan? How do I set a strategy [8:07] around the moat around the data, the [8:09] moat around the maybe the user [8:10] experience, the moat around our unique [8:13] value prop that we give? And then what [8:15] am I going to lease? What am I going to [8:17] take advantage of for now and build [8:18] towards later? And that ends up being a [8:20] lot of what that's really how people are [8:23] thinking about their strategy. And it's [8:24] not as clear-cut as a I want to own [8:26] this, I don't want to own that. Because [8:28] once you get it in house, once you build [8:29] that rag prototype and you're like, "All [8:31] right, cool. I'm going to need one whole [8:33] person to think about governance to [8:35] every one of these tools to think about [8:37] scale in terms of the indexes to think [8:39] about availability and this tool has to [8:41] be available 24/7 because people are [8:43] using it 24/7. So it ends up being a [8:45] really quick conversation around like [8:46] that might be cool. Why don't we take [8:48] the burden off of you for now for those [8:50] kinds of things and down the road you [8:52] can decide maybe this is something we do [8:54] want to own or maybe it's a good thing [8:56] that we were able to divert all of our [8:57] talent and resources to the thing that [8:59] was actually our mode. [9:00] >> Got it. Okay, that makes a ton of sense [9:02] and yeah, I would love to hear your [9:03] perspective particularly around all the [9:05] use cases. [9:06] >> Yeah, it's super interesting. We see the [9:07] build verse buy decision pop up a lot, [9:09] right? And what's the the fun part for [9:11] us about it is that to build something [9:14] today, it takes a lot of effort. It's [9:16] really hard. It's really hard to figure [9:18] out who to hire. It's really hard to [9:19] figure out what is the strategy around [9:21] it, which means that oftent times it [9:23] makes sense to actually go out and buy [9:25] something instead, right? And so when we [9:26] think about how do we showcase what [9:28] we're doing into enterprises and you [9:30] know show folks folks like you all like [9:32] how can you use this? What can you do? [9:33] It really comes down to two things. The [9:34] first thing is showcasing the art of the [9:36] possible right what can this technology [9:39] actually do right everyone's tried [9:40] Microsoft copilot right it it sounds [9:42] lovely doesn't actually work great [9:44] concept of AI but like how can we help [9:46] people dream of what you actually do [9:48] with it. So part of it is showing really [9:50] amazing use cases, having people promote [9:52] from the ground up, um helping people [9:54] understand where this can provide value [9:56] and having really strong thesis around [9:58] that, right? It's, you know, not [9:59] necessarily a torched earth policy where [10:01] you're trying to do everything, but [10:02] these are the areas that we play really [10:03] well in. The second side of that though [10:05] is once you show the art of the [10:06] possible, you need to make it really, [10:08] really easy to bring that to fruition, [10:10] right? How can you actually put the [10:11] tools within someone's hands? How can [10:13] you give them the access or the [10:15] capabilities where they can build this, [10:17] launch this, log into it, start using [10:19] it, have a magic moment, start seeing [10:20] value from it, whatever the metric is in [10:22] minutes, not hours, not days, not weeks, [10:24] not months in many cases. How can we [10:27] really show that? And there's a lot of [10:28] ways that that we go into that and and I [10:30] won't delve too much into those, but [10:32] really what we found is the key to [10:33] actually um you know, showing the value [10:36] of AI and bringing it to life is helping [10:38] people understand what it can do and [10:39] then actually living up to that promise [10:41] of making it easy. And the product [10:43] speaks for for itself when it comes to [10:44] that usually. [10:45] >> Got it. No, makes sense. And yeah, [10:47] definitely want to double click on time [10:48] to value. That's, you know, for a [10:50] startup is absolutely key to getting the [10:52] attention. But love to turn it over to [10:54] the two of you as well and hear just how [10:57] do you consider your relationship with [10:58] the startup AI ecosystem uh relative to [11:02] other efforts, you know, with the [11:04] broader big tech or with internal [11:05] efforts as well. Yeah, look in order to [11:08] address this, I want to start from uh [11:10] where do we start with the AI strategy [11:13] or with any kind of strategy, right? So, [11:16] we shouldn't start like with a tool, [11:18] right? With the something what is [11:19] exciting and then decide and build and [11:21] buy. We start from like where what what [11:25] are the areas we need to transform. So, [11:28] essentially domain forward approach. you [11:30] you don't want to boil the ocean and try [11:33] to do small things everywhere in the [11:35] organization. You need to define the [11:37] domain where your uh you have the best [11:41] balance of value versus like time it [11:44] will take for you to implement right the [11:46] change management you will need to do. [11:48] Then secondly you uh really need to [11:52] think through the process like what is [11:54] your process end to end and how do you [11:57] want to change it and then after that [12:00] only it comes right. So once you uh lay [12:02] out the process and the value map once [12:05] you understand how do you re-engineer it [12:08] then think about the tools and going [12:10] like from the uh kind of path of the [12:14] less resistance if you can do that data [12:17] integration and automation instead of AI [12:19] do it first right and then you uh select [12:22] and then then you kind of convert it to [12:25] the sets of the use cases and then only [12:28] at that point you need to think about [12:30] like whether you build or or buy right [12:33] so whether anything exists what solves [12:36] this problem or you need to uh reinvent [12:40] it yourself right so uh essentially many [12:44] companies I see they start with oh this [12:46] tool looks cool and then let us find the [12:49] uh application for that it's not like [12:52] how the uh strategic transformation [12:54] works strategic transformation need to [12:57] work like from domain from process uh [13:00] angle and then tools is only uh [13:04] essentially the the last part right so [13:07] when uh you uh decide like what use [13:10] cases and how you solve them [13:14] >> yeah I think uh so where can startups [13:16] help um especially for banks [13:20] um I think I think we truly value lots [13:23] of partnerships that we have and know I [13:26] think uh one thing that I always have uh [13:31] communicated with my startup friends is [13:33] really understand the big company [13:36] culture, [13:38] understand the the nuances that these [13:41] big banks especially care about because [13:45] of the regulated environments and do not [13:48] assume that data is going to be out [13:51] there freely available for testing. [13:54] Right? I think to me this is where the [13:56] biggest challenges are and and also the [13:59] prototypes and the state-of-the-art [14:01] possibilities are fantastic. When [14:03] startups come to any of our meetings, [14:05] the kind of energy levels that we have [14:07] inside the room is so high because oh my [14:09] god we could do that. Awesome. The [14:12] immediate question is really how [14:13] practical is it? Can we actually do it? [14:16] So this is where I highly recommend uh [14:19] all of the startups to think a broader [14:21] set of protocols that works with [14:24] existing enterprise systems and most big [14:27] banks like us we already have players [14:30] like Microsoft I mean obviously the [14:33] ecosystem that we have the co-pilot has [14:35] some advantages but obviously you got to [14:37] weigh the pros and cons we also have [14:39] Salesforce we also have workday right we [14:42] we have all these you know mega [14:44] platforms s and they're also competing. [14:47] They're always have they always have [14:48] some kind of a need to extend those [14:50] platforms. Think about how you can kind [14:53] of add more capabilities to those [14:55] platforms as well. It makes it easier [14:57] for you all to really come to us make it [14:59] make our life much simpler and I think [15:02] that's how I kind of see startup as a [15:03] huge value ad that I think uh you all [15:06] can actually provide and we heavily rely [15:08] on them today to innovate and integrate [15:10] with our exist existing ecosystem. [15:13] >> Great. Well, I'd love to dive in further [15:15] on that as well. So, for startups, [15:19] uh there's really a couple ways that [15:21] they can get deployment in enterprise, [15:22] right? They're selling direct. There's [15:24] also partnering with channel providers [15:26] like Microsoft or Google or Salesforce. [15:29] Uh at the same time, we've seen a lot of [15:31] development and evolution in pricing [15:33] models for AI technologies. So, I'll [15:37] hand it back to the the vendors first. [15:39] love to hear how you're thinking about [15:41] pricing the product and over time how [15:44] channel providers can be more friend [15:46] than foe as you're looking to get [15:47] deployment at scale and maybe we start [15:49] with you first. [15:50] >> Yeah, happy happy to. Um pricing's hard. [15:53] It's really hard in the AI world. You [15:55] know, we're we're going from previous [15:57] SAS companies where they're really [15:59] wasn't necessarily scaling unit [16:00] economics at least to the degree or the [16:02] magnitude there is today to now you know [16:04] running running inference running uh you [16:06] know different GPUs. It's really [16:08] expensive uh you know to run our [16:10] technology it costs several cents per [16:12] minute right and so what we have to do [16:14] and our responsibility um to to your [16:16] point earlier is how can we meet [16:17] enterprises in the middle with that [16:18] pricing right we know okay we have you [16:21] know x cost basis it costs you know ABC [16:23] to run but how can we translate that how [16:26] can we actually apply you know a [16:28] function to that where someone [16:29] understands it for different use cases [16:31] where it works in existing pricing [16:32] models or existing ecosystems where you [16:35] know the translation burden is on us [16:36] versus other folks [16:38] So a lot of the way that we we think [16:39] about pricing is how can we align it [16:41] very closely with value and speak about [16:43] it in really simple terms that align [16:45] with existing ecosystems. And I think [16:47] that gets really interesting though when [16:48] you talk about channel partnerships is [16:49] that's where the distribution comes in [16:50] and that's where standard economics come [16:52] in. So if you're able to mold yourself [16:54] into a channel partnership that serves [16:56] really really well for initial [16:57] distribution and brand recognition. So [16:59] that's one of the big things that I [17:01] think pricing is now becoming a lever [17:02] for. Whereas previously it used to be [17:04] about SAS margins. It used to be about [17:05] things like that. Now, pricing is really [17:07] becoming a lever for brand awareness, [17:08] recognition, and different flywheels [17:10] when it comes to marketing. So, it it's [17:12] pretty fascinating. I'm curious everyone [17:13] else's thoughts on the panel as well, [17:14] how they're seeing pricing evolve. Um, [17:16] but seeing that today has been quite [17:17] quite the trend. [17:18] >> And has channel partnerships been a [17:20] meaningful uh growth lever today or [17:23] something being considered for the [17:24] future? [17:26] >> It's been beneficial today. I I wouldn't [17:28] necessarily say it's been the the top of [17:30] the list by by any means. I think uh [17:32] direct sales has by far and a way been [17:34] been most interesting for us. Uh we see [17:36] most of our enterprise and fortune 500s [17:38] come directly to us or we go directly to [17:40] them. Uh the channel partnerships has [17:41] been really helpful for getting an [17:42] initial foot in the door and then [17:44] helping lever a direct sale from there. [17:46] >> Curious for you especially with the [17:48] government. [17:48] >> No, totally. And I mean the government's [17:50] a weird beast and anybody who is either [17:52] in the government or worked with the [17:53] government knows that you just kind of [17:54] like draw a box around that. That's a [17:56] whole different world. You almost treat [17:57] it and you have different humans even [17:59] that deal with it. Um, but I actually [18:01] think um I'm not saying that that uh AI [18:04] doesn't introduce complexity or or it [18:06] isn't hard to think about pricing in the [18:08] AI world, but I think this is like one [18:10] of the most unique times in the world in [18:13] terms of pricing and strategy around [18:15] bringing a product to market. Not just [18:17] because technology is there and so it [18:19] makes it so you can do it fast, but more [18:21] specifically like this is a time when we [18:22] get to be creative, right? Like [18:24] previously we were we always talk about [18:26] pricing to value. It's like the phrase [18:28] you use. And the question is like what [18:30] what does value mean right like if I'm [18:32] if I'm a hyperscaler I'm I'm going to [18:34] price based on compute and I don't know [18:36] how compute is tied to value when it [18:38] comes to AI uh it's it's I can't say [18:40] it's one to one right and SAS some a lot [18:43] of times with seats and I don't know how [18:45] seats are always priced directly to [18:46] value and so like there's a creative [18:48] component to this that we get to you get [18:50] to mold the pricing in a way that makes [18:52] sense so maybe and what we explore is [18:54] like let's talk about outcomes right we [18:57] can have something that makes us whole [18:58] we can have something that makes it so [19:00] that the platform, you know, is paid for [19:02] in some capacity. Um, but let's talk [19:04] about outcomes. Let's talk about the the [19:05] kinds of things that you want to [19:06] accomplish. Let's mold and adjust the [19:08] pricing in some way, shape, or form. And [19:10] by the way, because this is like a [19:12] unique and and early part of this [19:14] industry, it's probably okay. Think the [19:16] ball's going to drop at some point in [19:18] time and we're going to figure this all [19:19] out. But that also opens up the door to [19:22] industries that you might go into and [19:24] partners and being creative with them as [19:26] well because as you go into something [19:28] like construction or you go into [19:29] financial services, they have a lot of [19:31] different pricing components and [19:32] structures and they might even have [19:33] different teams that are taking this on. [19:35] So being able to kind of uh pivot to [19:37] wherever the budget is is is something [19:39] that I've never been able to do in all [19:41] of my years until now. [19:42] >> Yeah. So you're on the other side of it. [19:44] You have to work with procurement teams. [19:45] So when you hear these different pricing [19:47] models, what do you think? [19:48] >> Yeah. Look, I I think the pricing I mean [19:51] if you take uh three different [19:53] components of pricing traditionally it's [19:55] software, hardware and uh human cost. I [19:59] think we always think that you know [20:01] humans are really the most expensive [20:04] component no matter where you go. And so [20:06] when we think about the generative AI, I [20:09] know there's a little bit of a hardware [20:11] component right now with the GPUs and [20:13] all of that. I think that plays a role [20:15] when you scale the implementation of uh [20:18] the applications. So we kind of do care [20:20] about that. But I think we kind of spend [20:22] most of our time and energy in terms of [20:24] really you know if we have to do [20:26] in-house how much would it cost for us? [20:29] Would we have uh would we really hire 10 [20:31] people and train them and so on so [20:34] forth? I mean if you take uh you know [20:36] just uh you know $100,000 employee right [20:40] with $100,000 you can run billions of [20:43] tokens from influence perspective right [20:46] I mean just one employee one one kind of [20:48] so I think uh what we do in uh in terms [20:51] of really when you look at the when [20:52] startups come to us the very first thing [20:55] we look at is really the outcome base I [20:57] think as as rightly mentioned like if we [21:00] do this what will be the outcome in [21:02] order for us to do this how much money [21:04] does it really cost? By the way, we take [21:05] a long-term approach. We take a [21:07] threeyear, 5year approach. We kind of do [21:10] the NPV calculations. So, initially it [21:12] may be cash outflow for the first one or [21:14] two years, but we fully know that the [21:16] next 3 years we're going to have more [21:18] cash inflows. That's kind of the [21:19] calculation we do internally. That makes [21:21] it a lot easier for communication. So, [21:24] we kind of don't have to worry about [21:26] sweating somehow make it work initially [21:28] because we always think long term. [21:31] >> Got it. Um well I want to stay with with [21:34] you two for a bit longer and talk a bit [21:36] about use cases that you think are most [21:38] ripe for AI to play a role in disrupting [21:40] within your orgs. Obviously much [21:42] different companies of scale industry uh [21:44] but maybe there's some similarities and [21:46] you know areas to tackle for AI. [21:49] >> Yeah I guess uh there will be a lot of [21:51] similarities especially in the back [21:52] office operations right. So on in the [21:55] functional areas but uh the use case I [21:58] want to talk about like or the the big [22:00] transformational [22:02] uh program uh which we are running now [22:05] is about the design and design [22:07] automation right so thinking about every [22:09] building and think about like this room [22:12] uh in order like when every uh everyone [22:16] has hears about design it's typically [22:19] you think about like architects and [22:21] renderings and these flashy kind of [22:23] objects But surprisingly like 95% of [22:26] work in design is actually in not in the [22:30] architectural space but it's about like [22:33] looking around this room where do you [22:36] place the lights uh what shall be the uh [22:39] wattage of those lights shall be the [22:42] right uh grills and like for the hack [22:46] system how the hike system need to be uh [22:49] uh need to be uh sized for this kind of [22:52] room. There are like kind of hundreds of [22:55] the different systems in this room you [22:57] need to design and all of that is [23:00] designed by human and uh what we are [23:03] doing we are actually automating those [23:05] flows and this is how like sometime uh [23:08] this example on how we work with [23:10] startups. The startups are great as a [23:13] point solutions. Think about like there [23:15] would be the startup which can do the [23:17] lighting design. There would be the [23:19] startup which can do like uh design of [23:22] cooling system but like where our role [23:25] is is to think about the whole process [23:27] and how do you orchestrate those things [23:30] and uh what we are building we're [23:32] building like the aentic connecting [23:35] tissue which brings all those point [23:38] solutions together right because all [23:41] those designs have its own dependencies [23:45] and all those therefore you need to have [23:47] like agent in between which orchestrates [23:50] everything and which uh takes the [23:53] information from one system to another. [23:56] So then you can generate multiple [23:58] designs and all of them like not [24:00] clashing with each other. So it's like [24:02] this is one of like cool cool examples [24:05] like which are quite easy to understand. [24:08] Um I want to add for banking side. Yeah, [24:12] I think broadly I mean the we kind of [24:15] put AI use cases into four buckets [24:18] broadly speaking. Uh one is revenue [24:21] maximization bucket. Uh the gen I'm [24:24] talking about the gen AI use cases and [24:26] also generally AI use cases. Uh the [24:29] second bucket is a loss minimization [24:31] bucket and uh the third bucket is really [24:34] about productivity improvements and the [24:38] fourth is operational efficiencies and [24:41] where we think you know without even [24:43] thinking you can readily go into number [24:46] three and number four which is a [24:48] productivity improvements across the [24:50] board no matter what role you play in [24:52] the company you can actually use geni to [24:55] improve your productivity. [24:57] uh the for example developers if your [25:00] companies have it developers if they [25:02] want to improve the productivity that's [25:04] going to translate to significant [25:07] millions and millions of dollars if we [25:08] if done right and and similarly if you [25:11] are I'm sure if you're working in [25:13] full-time you know jobs and if you have [25:15] to do performance evaluations every 6 [25:18] months and 12 months end of the year how [25:20] much time you actually spend to simply [25:22] write to take all the feedback summarize [25:24] and write now with the gen AI we can do [25:26] it a lot you know at least onetenth of [25:29] the time is what I'm hearing and and uh [25:31] in my own experience depends on how much [25:33] time you want to spend and so [25:35] productivity is a no-brainer but [25:37] operational efficiencies I think is a [25:39] big area in banking and for example AML [25:42] operations anti-moneyaundering [25:43] operations and KYC you have to write a [25:46] credit memos for every single commercial [25:49] client that we serve and all of those [25:51] are very human labor intensive with gen [25:55] AI we got attack So those two are super [25:57] important. Loss minimization you can do [25:59] on fraud especially the you know if you [26:02] have a business email compromise [26:03] scenarios and things like that but uh [26:06] revenue maximization you got to be a [26:07] little bit more careful around you don't [26:10] want that to be in front of the customer [26:11] without human in the loop kind of [26:13] scenario. Attraction is not there yet in [26:15] terms of automating fully the whole life [26:17] cycle without human in the loop. So [26:19] those are kind of four broader [26:20] categories that we actually look at. [26:22] >> Super helpful. Uh so at least in our [26:25] portfolio the the the startups that we [26:27] see uh a combination of you talk about [26:30] human in the loop where AI can play an [26:32] augmenttor for productivity and then [26:35] we've seen full automation solutions as [26:37] well. I mean you both are kind of in the [26:40] space where uh the role of humans and [26:42] these orgs are going to rapidly change [26:44] you know whether it's using these tools [26:46] or perhaps some AI will be doing the [26:49] work that maybe humans have been doing [26:50] you know for the last several decades. [26:52] So love to hear the perspectives of all [26:54] four of you where you see the role of [26:57] team members in your orgs or your [26:59] customers orgs evolving as AI continues [27:02] to get deployed in the organizations. [27:04] >> Yeah, I'm I'm happy to start. I think [27:06] like I mean we're you know we're 405 and [27:08] I think this is going to be probably the [27:10] place where scale actually hits a lot of [27:12] the conversation. Like I I think people [27:14] is the first part of where scale [27:16] challenges are for a lot of [27:18] organizations. I think over the past six [27:20] or seven months, I've probably talked to [27:22] close to 50 different companies, and a [27:25] large portion of those companies have [27:26] gotten to production at some point in [27:28] time with whatever project they were [27:30] working on, but they either hit a [27:32] roadblock from a technology perspective [27:34] or they hit a roadblock from a people [27:36] perspective. Like they saw that hockey [27:37] stick growth really quickly of like, oh, [27:39] this is such a cool tool. Thanks for [27:41] giving me access to this. and then the [27:43] precipitous drop off where a whole slew [27:46] of things happened. Whether it was, [27:48] okay, I don't actually know how to use [27:49] this. I had my magic moment, but then [27:51] like I didn't have a use case tied to [27:52] it. Um, or they saw their jobs being [27:55] taken away in front of their eyes. And [27:57] and so like there's there's a bunch of [27:58] those different human components that [28:00] like we can stitch together a bunch of [28:02] cool technology to your point. Um, but [28:04] if we don't have like the human [28:05] component planned into that, then you're [28:08] you're going to hit the same roadblock, [28:09] right? This is why I loved this industry [28:11] is like it's finally where all the data [28:12] and all the cool technology meets the [28:14] human component. And so it really has to [28:16] be a part of the strategy that you have [28:18] to kind of roll it out. And so for our [28:20] customers, the folks who have been [28:21] really successful, they've started up [28:23] front with a plan that was like, how am [28:25] I actually going to change these [28:27] particular jobs? How am I going to take [28:28] the productivity gains and apply that? [28:31] And it very well could be that we say, [28:33] okay, cool. Maybe our hiring forecast [28:35] for the next year is going to be reduced [28:36] by 20 30%. And that's that's a good [28:38] thing, right? that means you can get [28:40] more done with less with less hiring, [28:41] with less new talent, right? Or maybe it [28:43] means that people are going to be able [28:44] to do more. They're going to be able to [28:46] take on new tasks, uplevel them, upskill [28:48] them in some way, shape, or form. And so [28:50] that that's kind of been the interesting [28:52] thing that we've seen is like those have [28:53] been successful in deploying AI at scale [28:56] within their organizations have put a [28:59] human component into the the goals and [29:01] objectives that they've had the upfront [29:03] training or or whatever or the plans the [29:05] success metrics for them were what are [29:07] we going to do with this gained [29:09] productivity or this gained revenue or [29:11] this this reduction in loss what do we [29:13] actually do with that as an outcome [29:15] >> totally totally agree with your points [29:18] right so it's like the way I I typically [29:21] think about this in order to run [29:22] successful implementation, right? So we [29:25] need to have the clear leadership buyin [29:28] and ownership right so it's like down to [29:31] the point that the business unit leaders [29:34] need to have the KPIs on uh the degree [29:36] of penetration or degree of automation [29:39] they need to achieve with AI. Secondly [29:42] like when you implement is also like [29:45] there shall be the bottom up buying. So [29:48] people need to clearly see and [29:50] understand the value and understand the [29:53] new process what what is the new way of [29:56] doing thing it basically new operating [29:59] procedure that uh actually I need to [30:02] implement so it's like my role not [30:04] anymore like to do those things manually [30:07] my role uh is to critically think about [30:10] the outputs of AI agent and to point the [30:15] uh uh to give the feedback to point AI [30:18] agent in the right direction to improve [30:20] the output further. Then uh the third [30:23] you already mentioned also like training [30:25] right so we need to train people to [30:27] ensure that they clearly um uh [30:31] understand how to do how to use and the [30:34] fourth point is their reinforcement [30:37] mechanisms. So we need to have the [30:39] mechanisms to if we have the new process [30:42] if we have have like uh people trained [30:45] and if we have everyone aligned that [30:48] this is the way we do right so then we [30:51] need to have reinforcement mechanism to [30:54] catch like non-compliance right we need [30:57] to measure the usage we need to measure [30:59] the value capture and we need to uh [31:02] actually have the to give the feedback [31:05] when the things are not going Right. [31:09] >> Yeah. So, if I'm understanding right, [31:10] the question here is how do we balance [31:12] scaling humans with scaling AI? Is that [31:15] a is that a fair interpretation? H [31:17] >> how as a AI technology being embraced by [31:21] an organization? Uh well, how much of [31:23] the burden is on you to help [31:25] organizations think through workforce [31:27] changes? Uh and then two, just more [31:29] broadly, how do you think it's going to [31:30] evolve over time uh as your product or [31:33] other similar great technologies get [31:35] deployed in these companies? [31:36] >> Yeah. Okay, that that makes total sense. [31:38] Um, we're we're uniquely qualified to [31:41] talk about this one as we're building AI [31:42] humans, right? So, it it hits the nail [31:44] on the head a little bit. I I actually [31:45] think the framework you laid out was [31:46] really apt for this, right? There's four [31:48] areas that AI in general can help an [31:50] organization. And uh when I think about [31:52] our enterprise clients that we work [31:53] with, all the use cases fall into one of [31:55] those four buckets. And what's really [31:57] interesting though is in any of the the [32:00] deployments we've done, it's been [32:01] actually very rare to see human teams [32:03] replaced or even scaled down. Yes, [32:04] efficiency goes up. Yes, sometimes [32:06] hiring forecasts go down as a result of [32:08] that. But largely what we see is that [32:10] these applications are being put in [32:12] place where humans previously were not [32:15] uh able to be there or scalable to be [32:17] there. And I'll give an example of this. [32:19] There's a healthcare company based in [32:20] New York. Uh they have a you know [32:22] customer success team who's supposed to [32:24] be helping out all of their their users. [32:26] And what they would love to have happen [32:28] is that every user gets activated [32:29] earlier. They get upsold and they get uh [32:32] you know sort of converted at at support [32:34] tickets. But what they find is that they [32:35] aren't able to have humans online 24/7 [32:38] to serve them. So when someone's [32:39] reaching out with an issue in the middle [32:40] of the night, they oftent times cancel [32:41] their subscription instead of being able [32:43] to connect with someone. Uh often times, [32:45] you know, a human is not able to handle [32:48] all of the demand at once or a lot of [32:50] the tickets need to be automated because [32:52] they're too low value in terms of [32:53] transaction dollars where it doesn't [32:54] justify human's time. So what we're [32:56] actually seeing is a lot of these [32:57] applications are being put in place [32:59] where humans can't currently scale, but [33:01] there's still enough of a monetary value [33:03] to justify an application for it. So [33:06] it's beating around the bush a little [33:07] bit, but I think what we're actually [33:09] seeing is enterprises are becoming far [33:10] far more effective and more efficient [33:12] and startups for that as well. Uh with [33:14] the same human headcount while being [33:16] able to scale AI in conjunction to that. [33:20] Yeah, I think uh so from uh scaling [33:23] right I think uh we're living in a very [33:26] interesting phase of our lives. Um [33:28] anytime the consumer technology [33:31] uh outpaces the enterprise technology [33:35] it's a pressure on enterprise technology [33:37] teams to deliver right right now the [33:39] demand is really coming from every [33:41] single employee of tris we have 40,000 [33:45] 50,000 teammates they all have chat GP [33:47] at their hands the mobile phone if I can [33:50] do it over there why when can I do it [33:52] here internally right it's like you know [33:55] that's kind of the number one problem [33:56] that we have Right now the demand is so [33:58] high that technology inside enterprises [34:02] is not up to speed. This is where we [34:04] have to do lots of this. What this [34:06] really means is democratizing AI to [34:10] everybody in the company because they [34:13] know how to use it and that's number [34:15] one. Number two, if you look at the how [34:19] the day starts with people that have [34:23] access to the GPT like capabilities, [34:28] even though they know the answer, the [34:30] human tendency in psychology is to first [34:33] check. [34:34] I write an English sentence in response [34:37] to my boss, okay? Or some kind of RFP, [34:40] whatever. again I want to check and see [34:43] if the agents can actually do a better [34:45] job. So that almost has become a de [34:48] facto standard now that every single [34:52] employee will adopt meaning every single [34:54] employee of a company will have their [34:56] own set of agents and I think so the [34:58] human agents I think would as you said [35:00] earlier will never disappear. We I mean [35:02] the way that I classify we classify [35:04] agents here at TR is broadly into four [35:06] buckets and I think I always often joke [35:09] with my friends if you tell me the fifth [35:11] bucket I'll probably give you a million [35:12] dollars and Cosman is right here [35:14] laughing is uh so we had a conversation [35:16] yesterday so the four types of agents [35:19] are number one human agent humans we [35:22] humans still are very relevant number [35:25] two is a deterministic agents like if no [35:28] if then else logic this is all robotic [35:30] process automation right We've been [35:31] doing it for some time. That's exactly [35:33] how we write software today. Three is [35:35] where the probabilistic agents, this is [35:37] where the AI agents come into play. [35:39] That's just one of three. The fourth is [35:42] robotic. This is a physical agent like [35:45] opening mails. People still send checks [35:49] or postal mail and somebody has to open [35:52] it. We can actually put a humanoid or [35:54] some kind of a robot machine to do it. [35:56] That's the fourth. That's it. Four. But [35:58] if you want to build application you may [36:01] need either all of them or maybe three [36:03] of them two of them right but we you [36:06] know humans wherever or needed we will [36:09] have them and I think with this [36:11] capability we'll have we need more [36:13] humans that understands how to interact [36:15] with other agents that's where the [36:17] democratization of AI will happen and I [36:19] think we'll be better with our chat GPT [36:22] than without it we'll be better with [36:24] internet without it it's exactly how I [36:27] think it goes in the future It's going [36:28] to be I think you know sometimes you [36:30] know I think it's me and my CIO calls we [36:33] don't call AI is artificial we call as [36:35] augmented I think you know that's kind [36:37] of how it goes [36:38] >> there there's an interesting point in [36:40] there as well and you know AI or humans [36:43] need to learn how to interact with AI [36:44] and and scale within the or I think one [36:46] thing that we're starting to see a lot [36:47] too is people are starting to hire [36:49] generalists right there are less and [36:51] less really specialized roles being [36:53] hired whereas generalists are able to be [36:55] much more effective with the tools that [36:56] they are able to have I think one thing [36:58] that'll be really interesting to keep an [36:59] eye on is not how org headcount changes, [37:01] but actually how org structure changes [37:03] and how the roles and responsibilities [37:04] within that org evolve and and specific [37:06] people's roles evolve with that. [37:09] >> Yeah, I was actually going to build on [37:11] on exactly that is like we're we're [37:12] starting to see folks who are looking at [37:15] roles that are maybe maybe they're [37:17] generalized as they use uh AI, but the [37:19] folks that are there supporting an AI [37:21] application. So folks whose jobs even [37:23] even inside our organization, our sales [37:26] engineers are are really like they're [37:29] managing context, they're managing [37:30] scripting, they're managing prompt [37:32] engineering, etc., etc. Um, and so that [37:34] is kind of what we see as being the the [37:36] sort of evolution of some of these roles [37:37] is like learning how to use the agents [37:39] and having training and then suddenly [37:41] you are an orchestrator yourself and [37:42] you're an operator. And so, you know, we [37:44] we've been joking kind of internally [37:46] that like 2025 was the year of the [37:48] agent. 2026 is the year of the human is [37:51] the orchestrator. The human is the one [37:53] who gets to be the one in charge of many [37:55] of these different operations happening [37:56] all around them. And that's how you [37:58] create these scales of efficiencies so [38:00] that you have uh still auditability and [38:02] and visibility into it. Yeah. [38:04] >> Amazing. Well, want to say thank you to [38:06] all four of you for your time and [38:08] sharing all your insights and appreciate [38:10] it. looking forward to many more uh [38:12] discussions ahead around how AI scales [38:14] in enterprise. Thanks so much. [38:16] >> Thank you.