Block, Better Tech Chiefs Reject AI Headcount Cuts

Block与Better技术高管拒绝因AI削减员工规模 工程研发负责人驳斥了开发者生产力提升10倍的说法，指出实际节省的时间仅停留在每周数小时，且缺乏监管的代码生成还会带来代价高昂的技术债务。 该专题探讨了领先的企业级与金融科技软件公司如何在内部员工赋能及面向客户的产品（如Block的Goose智能体和Better的Betsy）中将生成式AI落地应用。与会技术负责人强调，在会计和借贷等受严格监管的领域，深度上下文支撑、定制模型上下文协议（MCP）服务器以及严格的可审计性必不可少。他们反驳了董事会层面关于开发者生产力可提升10倍及不受约束的“氛围编码”（vibe coding）的炒作，警告称这会导致系统架构恶化，且浅层LLM套壳产品面临极高的流失率。最后，发言者结合经济现实审视了按Token消耗定价所带来的阻力，并指出在硬性物理能源限制下，推理成本将降至零的前提并不成立。 Block在全公司12,000名员工中全面部署了基于MCP的开源智能体Goose，并同步上线Cashbot与Square AI，构建起贯穿内部生产力乘数与外部客户体验的多端AI回报闭环。 (3:24) FloQast指出企业财会市场呈现严重两极分化：仅有顶尖10%的客户能够消化端到端自主智能体，其余90%的传统财务人员则亟需通过体系化的技能培训与认证来弥合技术鸿沟。 (6:43) 为杜绝高风险金融与财会场景下的非确定性幻觉，Block构建了100多个自定义MCP服务器提供基准上下文，Better.com则通过数学依赖图为承保模型加设确定性轨道以满足严苛的合规审计。 (12:28) 专家小组驳斥了AI coding带来“10倍工程师效率”的神话，警告缺乏经验的初创团队若过度“氛围编码”（vibe coding）极易在代码库中积聚低质垃圾代码，必须依靠资深工程团队筑起架构防火墙。 (29:21)

Software leaders argue that translating artificial intelligence into durable enterprise growth requires rigorous domain grounding, transparent governance, and operational integration rather than relying on thin wrappers or assumed productivity miracles. Across fintech and accounting platforms, incumbents are balancing the rush for speed against customer trust, compliance mandates, and physical compute economics. Grounding Context to Protect Compliance and Trust (12:28) In highly regulated sectors like consumer finance and accounting, incorrect model outputs pose existential risks to users. Angie Jones of Block notes that hallucinations typically occur when models are starved of context, prompting Block to build over 100 custom Model Context Protocol (MCP) servers so internal agents reference the same internal documentation as human staff. At Better.com, Keith Soura explains that mortgage systems ground reasoning models within deterministic mathematical dependency graphs monitored by supervisory models to ensure state and federal compliance. Market Education Gaps and Monetization Models (6:43) Adopting AI in enterprise workflows exposes a sharp divide between tech-forward buyers and traditional operators. Roshan Chekuri of FloQast notes that while 10% of customers readily adopt autonomous agents—such as one client replacing a $60,000 annual outsourced cash-posting contract with a 5-second agent sold at $50,000 ARR—the remaining 90% need foundational education and certifications modeled after early RPA playbooks. Chekuri also reports changing pricing dozens of times because enterprise CFOs routinely reject unpredictable token-consumption models in favor of fixed software budgets. Debunking 10x Engineering Gains and Managing Slop (26:45) Panelists push back against the widespread executive belief that generative coding tools immediately yield '10x developers.' Angie Jones observes that most software engineers only save one to five hours weekly by treating models as basic code-completion tools, arguing that true ROI requires embedding AI across CI/CD pipelines and deployment workflows. Both Jones and Keith Soura emphasize that without seasoned senior engineers acting as a quality firewall, uncritical 'vibe coding' introduces severe architectural flaws and technical debt into production codebases. The Physical Limits of Deflationary Inference (35:29) Startups operating with inverted unit economics often justify burning capital on frontier models by assuming inference costs will inevitably drop to zero. Keith Soura disputes this 'abundance mindset,' arguing that unlike pure software, model inference remains physically tethered to energy generation, grid capacity, and cooling infrastructure. With power costs rising, Soura cautions that betting on zero-cost compute without breakthroughs in abundant renewable energy or fusion is an operationally risky strategy for software vendors. The panel indicates that over the next 12 to 36 months, rising customer churn will likely expose superficial AI wrappers, separating durable SaaS applications from speculative experiments. As traditional graphical interfaces give way to ambient conversational agents, long-term defensibility will depend on structural data moats, customer trust, and sustainable unit economics rather than sheer hype.

在企业软件迈向人工智能平台转型的过程中，行业领袖指出企业级 SaaS 的核心竞争力并非来自激进的代码生成或投机型封装，而是取决于将模型深度锚定于受控业务上下文、满足严苛的可审计性要求，并直面物理算力与商业落地的真实约束。 内部赋能与外部产品的多维投资回报 (3:24) 面对董事会推动人工智能落地的明确要求，企业正从内部效率提升与外部客户产品创新两端寻找切实的投资回报。Block 全球副总裁 Angie Jones 介绍，该公司开发并开源了 AI 代理 Goose，借助开放的模型上下文协议（Model Context Protocol, MCP）向全公司 1.2 万名员工推广，使非技术人员能够跳过静态产品文档直接生成可用原型，并实现自助数据分析；在客户端，Block 部署了客服自动化代理 Cashbot 以及支持自然语言分析的 Square AI。Better.com 工程副总裁 Keith Soura 则将 AI 定位为降低贷款运营成本的“利润引擎”，其内部助手 Betsy 能够实时评估数万个财务数据点以辅助信贷审批，同时利用自然语言帮助传统金融服务覆盖不足的借款人理解信贷准入标准。 垂直行业的上下文防线与审计要求 (12:28) 在金融和财会等强监管领域，模型幻觉可能直接对用户的财务状况造成不可逆的影响。Angie Jones 强调，幻觉的根本诱因在于缺乏上下文，为此 Block 构建了 100 多个客制化 MCP 服务端，直接向模型注入人类代表所使用的业务参考手册与系统数据，并在安全团队监督下通过小范围测试逐步推进。Better.com 则在信贷业务中运用数学依赖图将模型推理限制在确定性轨道上，以符合联邦和州的严格监管标准。 FloQast 人工智能总经理 Roshan Chekuri 指出，市场对全自主代理的接受程度呈现两极分化：约 10% 的技术领先型客户能够使用自动化代理替代外包工作流（例如以每年 5 万美元的订阅收入替代此前 6 万美元的现金过账外包），但其余 90% 的传统企业财务主管根本无法信任完全自主的代理系统。轻量级的“LLM 封装”（LLM Wrappers）在处理特定会计准则时每次都给出不确定的答案，带来了巨大的合规风险。因此，FloQast 采取类似于早期 RPA 和 Alteryx 的培训认证策略，明确区分确定性计算与自主代理模块，优先确保数据的可解释性与绝对审计透明度。 破除十倍研发神话与氛围编码债务 (26:45) 针对企业董事会中“AI 辅助工具可使工程师实现十倍生产力并削减 90% 人力”的普遍预期，与会嘉宾给出了明确反驳。Angie Jones 指出，绝大多数工程师仅将 AI 视作代码自动补全工具，每周实际节省的时间仅在 1 到 5 小时左右；真正的工程收益需要深入持续集成、自动化测试与部署流程等全研发生命周期。如果缺乏成熟工程团队的把关，缺乏经验的团队在过度依赖“氛围编码”（vibe coding）后，往往会产生大量低质代码并引发严重的技术债务，导致项目延期。 Keith Soura 透露，Better.com 拥有严守代码库质量的资深主任工程师作为防御机制。其董事会并未因技术红利缩减工程编制，反而选择全员普及 ChatGPT 和 Cursor 等工具的使用权限，并通过扩大招聘来利用研发效率的提升开拓全新业务线。这一策略让产品经理和管理层直接体会到 AI 的实际局限性，认识到要发挥模型价值依然高度依赖专业领域的鉴别力。 定价模式冲突与算力物理成本现实 (35:29) 在商业化落地与基础设施层面，前沿 AI 设想与市场现实之间存在显著断层。FloQast 曾设计了基于 Token 和工具调用的复杂计费方案，但在 50 天内调整了约 50 次，原因在于企业首席财务官（CFO）极度排斥不确定的用量计费，坚持要求可预测的固定软件预算。此外，Angie Jones 指出，将底层模型与执行代理捆绑的闭环厂商容易产生利益冲突，促使用户产生昂贵的 Token 账单，而支持多模型选择的开放架构能提供更好的成本透明度。 针对前沿 AI 初创企业普遍奉行的“推理成本终将趋零”的充裕理论，Keith Soura 提出了严厉质疑。他强调，数字模型推理必须受制于物理基础设施的真实约束，包括电力供应、电网容量以及冷却设施的硬性成本。在当前能源成本上涨的环境下，在可再生能源或核聚变等技术取得无限突破之前，依赖推理成本降至零的商业假设是一种不负责任的经营赌注。 与会者预计，未来 12 至 36 个月内 SaaS 市场将迎来重要整合洗牌：依靠浅层接口封装的投机型产品将因高客户流失率被淘汰，而基于自然对话的多模态交互可能逐步重塑传统图形用户界面。在企业级软件的竞争中，最终胜出的将是那些具备真实用户信任、开放互操作架构并能有效化解物理与合规约束的深层解决方案。

## Background 该视频由 YouTube 频道 Ai4 于 2025 年 8 月 24 日发布，记录了一场围绕“利用 AI 推动 SaaS 创新与增长”展开的行业圆桌讨论。视频简介指出，本场对话聚焦于 AI 如何深度重塑 SaaS 产品形态、客户体验、运营效率与定价模式，并深入探讨了企业在融入 AI 技术时所面临的落地挑战与竞争策略。 ## TL;DR 来自 Emergence Capital、Block、Better.com 以及 FloQast 的技术高管与投资人共同探讨了 AI 如何从单纯的概念采用跨越到实质性的业务成效。嘉宾们分享了 Goose、Betsy 等自研 Agent 在全员赋能、商户数据交互与信贷核保中的实践，并详解了利用模型上下文协议（MCP）和数学依赖图驯服模型幻觉的严密风控体系。针对“氛围编程（Vibe Coding）”导致的 AI 垃圾代码堆积以及 AI 销售邮件泛滥等现象，嘉宾们指出了过度依赖自动化的组织风险与保持真实人际交互的必要性。最后，讨论辩证分析了 SaaS 定价的可预测性诉求，并从现实物理能源约束的角度反驳了“推理成本必将趋近于零”的行业假设。 ## Atmosphere & style 整场对话洋溢着一线技术操盘手特有的务实与犀利，既有对前沿生产力跃迁的兴奋探索，又充斥着对行业浮夸叙事的清醒解构。从语音信贷 Agent 陪卡车司机通宵畅聊好莱坞电影而烧光 Token 的诙谐自嘲，到对“AI 垃圾代码”和“AI SDR”的尖锐吐槽，嘉宾间的互动松弛、风趣且充满化学反应。随着话题由业务应用逐步延伸到底层电网能源与物理成本约束，全场交流在理性、严谨的商业思辨中展现出扎实的工业级工程洞见。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=4s) AI 浪潮被视作继软件从本地部署走向云端之后的又一次重大平台级变革，SaaS 投资与创新重心正全面转向垂直场景智能化与 AI 原生生态。 - [03:24](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=204s) 企业应对董事会的 AI 焦虑不能仅停留在口号，而需在全员 Agent 提效、内部业务流程重塑以及终端产品创新等多个维度形成实打实的 ROI 闭环。 - [06:43](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=403s) AI 在高复杂度金融业务中充当“利润引擎”，通过自动化繁琐计算与数据核验大幅压缩单位承保成本，解放人力去专注高价值的客户深度沟通。 - [06:43](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=403s) 当前垂直 B2B 市场存在显著的认知鸿沟：仅有头部少数客户能直接采纳端到端 Agent，多数主流客户仍需供应商提供深度的体系化培训与市场教育。 - [10:02](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=602s) 老牌 SaaS 企业与敏捷的 AI 原生初创公司正面临“迭代速度与品牌信任”的博弈，在财会与金融等高合规领域，严苛的可审计性与严谨研发是立足之本。 - [12:28](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=748s) 模型幻觉的本质多源于事实上下文匮乏；通过自研 MCP 服务器注入完整业务 SOP，并结合底层确定性数学依赖图进行双轨校验，可有效规避金融非确定性风险。 - [15:29](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=929s) 自主语音或对话 Agent 在实际部署中必须建立严密的话题边界与闲聊超时拦截机制，否则极易偏离核心业务流程并产生高昂的推理 Token 账单。 - [17:16](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=1036s) B2B 客户在采购 AI 功能时对成本的“确定性与可预测性”有着极高要求，过于复杂的按用量/Token 动态计费模型在实际企业销售中阻力重重。 - [19:38](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=1178s) 对抗市场中的概念包装与简单的 LLM 包装层产品，核心武器是保持算法透明度，明确标定哪些环节依赖 Agent 推理、哪些依托确定性规则以满足监管审计要求。 - [21:41](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=1301s) 成熟 SaaS 应对 AI 新势力竞争应主打开放互操作性与赋能策略，例如推行模型不可知的工具生态，并指导客户一线人员亲自构建与业务深度绑定的 Agent。 - [26:45](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=1605s) “研发效率提升 10 倍”在很大程度上仍是行业叙事泡沫，多数开发者目前仅借助 AI 进行片段代码补全，企业必须推动工程团队向端到端交付流水线提效跨越。 - [29:21](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=1761s) 缺乏资深架构师把关的“氛围编程（Vibe Coding）”会导致大量 AI 垃圾代码堆积并导致项目延期，深厚的工程成熟度与严格的代码审查依然不可或缺。 - [31:49](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=1909s) 销售领域过度依赖 AI 自动化（如泛滥的 AI SDR 邮件）会导致人际互动温度缺失并引发客户厌烦，真诚的人际交流在业务拓展中依然无可替代。 - [32:42](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=1962s) 未来 SaaS 交互界面正从传统的点击/触摸式图形 UI 彻底迈向自然语言对话与视频虚拟人分身，计算形态正加速向无感融入的普适计算演进。 - [35:29](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=2129s) 初创公司赌“推理成本无限趋近于零”的商业假设存在巨大漏洞，大模型算力消耗受到现实物理电网与能源成本的硬约束，毛利倒挂的模式难以长期维系。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=4s) 嘉宾介绍与 AI 浪潮的时代背景 主持人 Yazan “Yaz” El-Baba 为本场聚焦于“SaaS 业务中利用 AI 推动创新与增长”的圆桌讨论拉开序幕。Yaz 是 Emergence Capital 的合伙人，他回顾了机构二十多年前的创立背景——当时基金押注了软件领域最重大的平台迁移之一，即软件从本地部署（on-premises）走向云端，而他们开出的第一张支票就是投资了当时还是一家“小而疯狂的公司”Salesforce.com。随着多租户架构（Multi-tenancy）的普及和 SaaS 产业的爆发，Emergence 见证了云原生时代的发展。如今，他们对新一轮 AI 平台变革同样甚至更加兴奋，并已重金布局了包括 Bolt、Bland、Together AI 等在内的 AI 原生公司。 随后各位嘉宾依次介绍了各自的业务背景与核心职责： - **Angie Jones**：Block（Square、Cash App、Tidal、Afterpay 的母公司）工程副总裁兼开发者关系负责人。她的核心职责是面向内外部推广和赋能 AI，尤其是推动 Block 内部从“AI 采纳（Adoption）”跨越到“产生实质业务影响（Impact）”，并向全球开发者普及如何利用 AI Agent、大语言模型（LLM）来自动化工作流并提升工程效率。 - **Keith Soura**：Better.com 工程副总裁。Better 主营消费金融与按揭贷款（Mortgage）业务。Keith 表示自己的本职工作是领导工程团队并保障研发生产力，但他个人最狂热的爱好依然是“亲自写代码搞工程”。 - **Roshan Chekuri**：FloQast AI 平台总经理。FloQast 成立已有 12 年，Roshan 在此工作了 8 年。FloQast 专注于财务会计数字化转型，帮助企业会计团队更快、更准确地完成月度结账。Roshan 幽默地指出，虽然公司 40% 的员工都是前会计师，但“会计工作确实非常枯燥，现在几乎没人读会计专业了”，未来 10 年内将有 75% 的注册会计师（CPA）退休。面对巨大的行业人才缺口，AI 成为绝佳的解决方案，但难点在于如何面对极度规避风险、对数据安全与审计合规要求极高的财务客户。 --- ### [03:24](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=204s) 董事会的 AI 焦虑与企业真实的 ROI 落地场景 Yaz 抛出了几乎每个董事会都在反复讨论的痛点问题：当董事会拍着桌子强调“我们必须做更多 AI”时，这到底意味着什么？究竟是通过 AI 优化运营效率，还是在产品中上线全新功能？企业在实际落地中又从哪里看到了实打实的投资回报率（ROI）？ ```mermaid flowchart TD accTitle: AI 在 SaaS 业务中的三维价值落地框架 subgraph Internal["内部运营与研发提效"] A["全员 Agent 赋能<br/>(Block Goose / MCP 协议)"] B["工程与产品提效<br/>(自动化取数、免写设计文档直接出原型)"] end subgraph Operations["业务流程与协同中枢"] C["业务中台助手<br/>(Better Betsy 电话与多维度按揭评估)"] D["端到端流程替代<br/>(FloQast 自动现金过账 Agent)"] end subgraph Product["面向客户的产品创新"] E["C 端客服与交互<br/>(Cash App Cashbot)"] F["B 端自然语言数据交互<br/>(Square AI 报表问答)"] G["普惠金融与客户咨询<br/>(按揭准入多方案匹配)"] end Internal --> Value["企业 ROI 闭环与利润引擎"] Operations --> Value Product --> Value ``` Angie Jones 率先回应称“所有这些方面我们都在做”。在 Block，ROI 体现在产品研发、产品功能整合以及全员提效等各个维度： - **内部开源 Agent 工具“Goose”**：Block 自研了名为 Goose 的开源 AI Agent，全公司 12,000 名员工（涵盖市场、销售、财务、设计、工程甚至高管助理）均在使用。该工具深度整合了模型上下文协议（Model Context Protocol，即 MCP）。非技术人员借此可以直接生成原型而无需先写冗长的设计文档；业务人员无需通过商业分析师即可自行调取所需数据。 - **产品端 AI 落地**：近几个月 Block 连续发布了两大 AI 产品功能：一是 Cash App 中的客服机器人 **Cashbot**，取得了显著成效；二是面向 Square 商户控制台的 **Square AI**。商户无需在复杂的仪表盘中层层翻找报表，只需通过自然语言提问（例如：“上个月销量前 10 的商品是什么？”、“什么产品快缺货了，能帮我补货吗？”）即可直接获取商业洞察。 --- ### [06:43](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=403s) 作为“利润引擎”的 AI 与财务领域的普及鸿沟 Keith Soura 从消费金融的角度分享了 Better.com 的落地逻辑。他将 AI 视为公司的“利润引擎（margin engine）”。由于房贷业务本质上是重人际沟通的行业，Better 的核心思路是利用 AI 解放人力，让团队有更多时间与客户深度交流。Better 的 AI 应用主要分为三大板块： 1. **面向 C 端消费者**：2024 年全美有超过 50% 的人群在传统银行服务中未得到充分服务，缺乏金融素养信息且对复杂的信贷产品感到恐惧。Better 通过自然语言交互为他们提供清晰的信贷与按揭知识。 2. **面向内部业务（Betsy 助手）**：Better 构建了名为 Betsy 的内部助手，她不仅负责接听电话，还能协助贷款专员（Loan Officers）在几秒内瞬间评估成千上万个数据点。过去专员可能会因为硬性指标直接拒贷，而现在 Betsy 能迅速匹配出“帮你成功买下这套房子的 11 种可行方案”，极大降低了专员的计算负担。 3. **面向产研团队**：消除重复性枯燥劳动，将数据监控与报表生成的权限民主化，大幅加快了产品和工程团队迭代新功能的速度。 Roshan Chekuri 则指出了财务 SaaS 领域的残酷现状与市场认知分化。当前，VC 和私募股权（PE）机构正不断向 CFO 和财务总监（Controllers）施压要求部署 AI。然而，市场上存在巨大的知识断层： - **头部 10% 的成熟市场**：已经能够采纳端到端的 Agent 工作流。例如某客户此前每年花费 60,000 美元将“现金过账（cash posting）”流程外包给共享服务中心，FloQast 为其部署了专门的 Agent，仅需 5 秒即可完成处理，并以每年约 50,000 美元的订阅费成交。 - **其余 90% 的主流市场**：对 AI 的理解仍停留在“用大模型总结邮件”。Roshan 直言：“你不可能直接把一套深度整合业务系统的端到端 Agent 工作流卖给辛辛那提大主教管区的财务总监。” - **应对策略（教育与技能提升）**：FloQast 借鉴了 Alteryx 和 RPA 时代的市场推广策略，重点不在于“用 AI 取代会计”，而是通过推出官方认证培训计划，教会会计师如何在平台上构建和部署 Agent，帮助行业人才实现技能升级。 ### [10:02](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=602s) 速度与信任的博弈：SaaS与金融科技公司的AI落地困境 在主持人 Yazan "Yaz" El-Baba 看来，当前所有企业不仅在内部部署 AI，更在逐步将其直接暴露给终端客户。这种交互有时体现为辅助内部员工的半托管体验，有时则是让客户直接与名为“Betsy”或“Goose”的自主智能体（Agent）对话。然而，眼下整个行业正面临一种尖锐的矛盾：**速度与信任的冲突（Speed vs. Trust）**。 当前 AI 原生初创公司（AI-native companies）的迭代速度令人咋舌。正如 Yaz 所调侃的，每周科技媒体头条都会冒出一家之前闻所未闻的公司，在极短时间内完成 ARR 从 0 到 100 的火箭式蹿升。与此同时，老牌 SaaS 企业虽然手握巨大的“分发渠道优势”（distribution advantage），却也背负着沉重的包袱——忠实客户多年来积累的信任容不得闪失。一旦贸然推出容易产生幻觉（hallucinations）的 AI 功能，多年建立的品牌信誉可能毁于一旦。针对如何平衡开发速度与用户信任，嘉宾们给出了各自的思考与实践。 FloQast 的 Roshan Chekuri 将当下的市场竞争格局划分为三大阵营： 1. **老牌企业级巨头**：深陷庞大的技术债务（tech debt）与经典的“创新者窘境”（Innovator's Dilemma），想要彻底掉头去重构和部署此类前沿 AI 极其困难。 2. **时机适中的成熟 SaaS**（如 FloQast）：既拥有既成的分发渠道优势，又尚未被沉重的老旧架构完全绑死，仍具备灵活创新的技术活力。 3. **Y Combinator（YC）的新生代极客**：一群毫无包袱的年轻天才，正如开车驶过旧金山海湾大桥（Bay Bridge）时所见，满大街的广告牌都在铺天盖地宣传 AI。 Roshan 强调，FloQast 服务的是对“可审计性”（auditability）要求极为苛刻的会计师群体，任何差错都是致命的。因此，团队在正式推向市场前，光是在底层 R&D 与算法研究上就扎扎实实投入了一年半的时间，确保每一步都经得起检验。 --- ### [12:28](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=748s) 金融级风控：通过充足上下文与数学依赖图驯服模型幻觉 来自 Block 的 Angie Jones 与 Better.com 的 Keith Soura 均指出，金融服务业（Financial Services）的容错率为零。“在这里你绝不能搞砸，因为赌注实在太高了。”在金融场景中，AI 绝不像帮人写学术论文那样只给出一个错误引用那么简单，一次错误的财务计算或信贷引导，可能会彻底改变一个普通人的财务命运与整个人生轨迹。 Angie Jones 深入剖析了幻觉的本质：**绝大多数 LLM 的幻觉源于“缺乏上下文”（lack of context）**。大语言模型的机制决定了它本能地“想要给出正确答案”，当缺少充分的事实输入时，它便会开始凭空捏造。因此，Block 的核心策略是尽可能为模型投喂极端详尽的上下文，确保它拥有与人工座席完全相同的业务指南、模版和标准操作流程（SOP）。为此，Block 甚至在内部为员工和产品涉及的各种 SaaS 工具自主构建了超过 100 个 MCP（Model Context Protocol）服务器，确保 Agent 在任何调用环节都有即时数据可用，无需做任何猜测。同时，信息安全（InfoSec）、合规（Compliance）与安全审计团队全程深度介入，以小规模公测（Beta）的形式谨慎灰度发布，步步为营。 ```mermaid flowchart TD accTitle: 金融级 AI 智能体上下文补充与双轨校验架构 subgraph Context_Layer["上下文供给层 (Block 实践)"] SaaS["内部 SaaS 应用"] --> MCP["100+ 自研 MCP 服务器"] MCP --> Guides["标准业务指南 / 操作模版"] Guides --> Agent["核心 LLM 业务智能体"] end subgraph Defense_Layer["实时风控与校验 (Better.com 实践)"] Agent --> DualTrack{"双轨执行与校验"} DualTrack --> Reasoning["高阶推理模型 (实时音视频/文本监控)"] DualTrack --> MathGraph["底层数学依赖图 (法定合规公式)"] Reasoning & MathGraph --> Gate["严格门控与一致性验证 (QC)"] Gate --> Output["输出精确承保/信贷决策给客户"] end ``` Better.com 的 Keith Soura 补充了他们在房贷承保（Underwriting）领域的防御机制。房贷计算大多受到联邦法律或各州法规的严格约束，团队将其抽象并转化为一套精密的“数学依赖图”（mathematical dependency graph）。在模型进行逻辑推理的同时，系统底层同步运行确定性的数学验算： - 依靠规则与确定性数学计算将推理模型牢牢锚定在“轨道”（Rails）上。 - 引入更高阶、更复杂的监控模型，对语音通话、文本对话以及推理链条进行全方位的毫秒级实时质检（QC）。 - 建立自反馈纠偏机制，让系统在受控环境下自主学习与纠错。 --- ### [15:29](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=929s) 线上“翻车”轶事：聊通宵电影八卦的语音信贷 Agent 当 Yaz 抛出一个刁钻的问题——“在 QE、QA、QC 和层层防护栏（scaffolding）之下，是否有漏网之鱼的 Bug，团队又是如何应对的？”现场气氛顿时变得轻松起来。Keith Soura 笑着看了一眼坐在观众席第一排正“怒视”自己的 AI 产品经理，随即分享了一个极具戏剧性的真实踩坑案例。 全美有超过半数的劳动力从事非传统工时的按时计酬工作，传统银行“朝九晚五”的柜台服务根本无法满足这部分人群的房贷申请需求。为此，Better.com 率先上线了合成语音 Agent——“Betsy”。在理想流程中，打进主线电话的客户即使在深夜也可以通过 Betsy 完成身份核验、填写贷款申请表，甚至直接锁定房贷利率（Rate Lock）。 然而上线初期，团队忽略了一项至关重要的人性洞察：他们以为用户听到通话开头的“本电话由 AI 提供服务”提示后，就会公事公办地办理业务。直到产品负责人 Corey 调出一通令人哭笑不得的通话录音——一名卡车司机一边开车一边与 Betsy 通话，由于路途信号断断续续，在等待系统调取信息的间隙，司机居然兴致勃勃地跟 Betsy 闲聊了起来。而 Betsy 表现得极其热情健谈，双方从体育赛事一路畅聊到好莱坞电影，司机在车里聊得不亦乐乎。最终，司机尽兴地挂掉了电话，贷款申请却压根没填完。这一通漫无目的的漫长闲聊瞬间烧掉了巨额的模型推理 Token 费用。团队在后台看着飙升的账单惊出一身冷汗，立刻连夜上线了闲聊超时拦截机制与话题引导防护栏。 --- ### [17:16](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=1036s) AI SaaS 的商业化探索：定价迷局与市场教育挑战 当讨论转向 AI 工具与功能的定价模式（Pricing Strategy）时，三位嘉宾展现了从“免费赋能”到“按量计费与可预测性冲突”的多元现状。 - **Block 的探索期**：Angie 表示 Square 的 AI 功能仍处于公测阶段，Cash App 等产品中的 AI 辅助功能目前也并未单独收费。公司当前的首要目标并非过早通过 AI 变现，而是将公测作为观察客户真实痛点与使用摩擦的窗口，这些数据将用于反哺未来的产品架构与长远商业化决策。 - **Better.com 的“利润机器”逻辑**：Keith 认为，他们将 AI 视为 SaaS 平台的“利润引擎”（margin machine）。公司的商业逻辑在于利用 AI 最大限度压缩房贷的审批与发放成本，省下来的成本空间可以直接让利给贷款人，提供更具竞争力的贷款利率，帮助更多人买得起房。因此，Better.com 衡量 AI 成功的核心指标不是软件订阅费，而是相同团队规模下承保产能的数倍提升与单位成本的下降。 - **FloQast 面对的定价现实**：Roshan 直言：“在过去的 50 天里，我们调整了大约 50 次定价策略！”现场负责定价的两位同事同样坐在台下。团队曾精心设计了基于 Token/算力消耗（consumption models）、搭配各种工具积木的优雅计费体系；然而一旦拿着这套方案去和企业 CFO 谈判，对方的第一反应永远是：“我只要**确定性与可预测性**（predictability）。” Roshan 与 Angie 一致指出，当前整个 B2B 市场存在巨大的“认知鸿沟”（knowledge gap）。在把自主 Agent 和 AI 全面推向市场之前，SaaS 厂商承担着极其繁重的市场教育工作：企业客户往往无法理解什么是基于用量的计费模型，也不懂企业为何要动态切换 LLM 或对模型做持续微调。作为供应商，不仅要交付代码，更被迫成为前沿 AI 概念的布道者与教育者。 ### [19:38](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=1178s) 客户与监管者的 AI 认知教育及对模型非确定性的透明度 在企业级销售中，向客户推销前沿的 AI Agent 往往会遭遇巨大的认知鸿沟。很多客户甚至连基础功能都还未摸透，听到复杂的 Agent 概念只会感到困惑和脱节。Angie Jones 指出，当前市场极度缺乏对 AI 工具“能做什么”以及“不能做什么”的清晰教育；即便是在逻辑上理解大语言模型（LLM）原理的软件工程师，在亲眼目睹模型给出错误答案时，依然会感到震惊——哪怕他们明知 LLM 本质上是非确定性（Non-deterministic）系统。 ```mermaid flowchart TD accTitle: AI 企业级落地中的透明度与合规流程 subgraph MarketNoise["市场现状与痛点"] A["LLM 包装层产品 (Wrapper)"] -->|非确定性输出 / 每次答案不同| B["合规与审计风险"] C["市场营销过度包装"] -->|混淆 Agent 与确定性规则| B end subgraph TrustFramework["企业级信任构建"] D["算法透明度"] --> E["明确标定 Agent 与确定性模块"] F["严密审计追踪 (Auditability)"] --> G["满足监管合规 (如核保/财会标准)"] H["客户深度赋能教育"] --> I["建立长期客户信任与粘性"] end B -->|驱动需求| TrustFramework ``` Roshan Chekuri 强调，在向客户做产品宣导时，“透明度”是穿透市场杂音的最关键武器。FloQast 会对客户保持极度坦诚：明确标示哪些功能是真正的 Agentic（智能体驱动），哪些流程出于审计合规（Auditability）要求依然运行在确定性算法上，即便市场推广中可能更倾向于贴上 Agent 标签。Keith Soura 对此深表赞同，他提到很多董事会往往把 AI 视作“银弹”（Silver Bullet），误以为只要开启 AI 开关一切问题就会迎刃而解。在 Better.com 所处的房贷与核保（Underwriting）领域，企业必须向监管机构用确凿的数据证明算法的公平性（Fairness）、权益保障（Equity）以及预测逻辑。Roshan 进一步补充，市场上许多追求快节奏的初创企业不过是简单的“LLM 包装层”（LLM Wrapper），例如一些声称能提供安永（EY）会计准则（ASC）咨询的工具，每次生成的结果都不一样，给不了解非确定性风险的会计人员带来了巨大的合规隐患。 - **非确定性风险**：通用消费级 GPT 或简单的 LLM 包装层在财会与金融合规场景中存在严重隐患，相同法规提示词多次查询会产生不一致结论。 - **监管合规考量**：在信贷核保中引入 AI，必须具备透明的数据支撑与解释能力，以满足监管对算法偏见与公平性的审查。 --- ### [21:41](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=1301s) 应对 AI 原生初创公司：开放互操作性、平台粘性与客户信任 面对市场上层出不穷、在 LinkedIn 上高调宣传且没有历史包袱的“AI 原生初创公司”，成熟 SaaS 企业正采取截然不同的差异化反制策略。针对主持人 Yazan El-Baba 关于如何与这些行动极快的初创对手竞争的提问，Angie Jones 分享了 Block 反其道而行的做法：不搞传统的厂商锁定（Vendor Lock-in），而是主打“互操作性”（Interoperability）与“用户自主选择权”。例如 Cash App 推出的 Cash Pool 拼单分账功能允许非 Cash App 平台的用户共同使用；开源 AI 编码助手 Goose 也是完全模型不可知的（LLM Agnostic）。Angie 调侃道，随着用户月底收到高达 2000 美元的“Vibe Coding”账单，大家很快就会意识到潜在的利益冲突——如果同一家供应商既按 Token 计费，又开发了负责执行任务的 Agent，“这个 Agent 会不会为了抬高账单而故意绕远路？” ```mermaid flowchart LR accTitle: 成熟 SaaS 对抗 AI 原生初创公司的三种差异化路径 subgraph BlockApproach["Block: 开放生态"] A1["模型不可知 (如 Goose)"] A2["跨平台互操作 (如 Cash Pool)"] A3["避免厂商锁定 / 消除计费利益冲突"] end subgraph FloQastApproach["FloQast: 深度赋能与粘性"] B1["3,500+ 存量品牌信任"] B2["沉浸式培训与认证体系"] B3["客户自行构建定制化 Agent"] end subgraph BetterApproach["Better.com: 客户利益优先"] C1["全财务背景分析与辅导"] C2["开放贷款市场平台"] C3["梅西百货模式: 推荐竞品建立信任"] end ``` Roshan Chekuri 指出，FloQast 依托 3500 家客户的品牌根基，通过教育与认证体系建立极高的护城河。其团队中的 John Woodard（被戏称为“JohnBot”）负责沉浸式入职赋能项目，指导客户中的会计师成长为“构建者”（Builders），在平台内亲自搭建适配自身业务流程的专用 Agent。一旦 Agent 与客户日常使用的全套工作流深度集成，产品粘性便无可撼动。Keith Soura 则表示，Better.com 借鉴了历史上“梅西百货”（Macy's）的经典服务哲学——永远以帮助客户为第一目标： > “哪怕我们现有的房贷产品不合适，我们会告诉客户如果花三四个月调整几项财务指标就能符合条件，并提供辅导；或者通过我们自研的开放贷款平台，直接把客户引流到其他有更合适产品的银行。你永远不对客户说不，他们因为信任你而最终还会回来。” - **开源与模型不可知**：Block 的 Goose 允许开发者自由接入任意模型，规避了模型商既卖算力又做 Agent 带来的经济利益冲突。 - **赋能型粘性**：FloQast 让客户一线会计人员自己上手“捏”Agent，将个人工作流固化在平台内。 - **开放市场策略**：Better.com 在过去 8 个月内力推自研贷款市场平台，不强行推销自营房贷，通过跨机构推荐建立深层用户信任。 --- ### [26:45](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=1605s) 内部研发采用与“10倍工程师”神话的现实差距 谈及企业内部对 AI 的应用与实际效能时，主持人 Yaz 提到，外界和董事会常常被“AI 编码工具或 Vibe Coding 让研发效率提升 10 倍”的案例研究所吸引，甚至反问管理层是否因此可以削减 90% 的工程团队编制。针对这一现象，嘉宾们深入剖析了期望与落地现状之间的巨大鸿沟。 Angie Jones 直言，所谓的“研发效率提升 10 倍”在很大程度上是行业虚构的叙事。尽管工具确实带来了一定改善，有时也会让人产生效率暴增的错觉，但放眼全球绝大多数工程师的实际应用，大家普遍还停留在非常初级的“代码生成”（Code Generation）阶段。许多工程师虽然知道高级功能的存在，但缺乏有效的使用场景认知，最终 AI 仅能在日常工作中每周帮他们节省 1 到 5 个小时，这根本谈不上 10 倍的飞跃。面对大量投入巨资却未见明显回报而焦虑的创始人与董事会，当前技术领导者的核心任务是推动研发团队“从浅层采用转向实际业务成效”（From Adoption to Impact），将 AI 创造性地嵌入到代码编写之外的更深层流程中，例如借助 AI 加速整个系统的持续部署（Deployment）与端到端交付流水线。 - **效率提升现状**：绝大部分一线开发者仅利用 AI 工具进行片段代码补全，平均每周节省时间仅约 1 至 5 小时。 - **工作重点转移**：AI 研发效能建设必须超越纯代码生成层面，向全流程自动化、智能部署工作流以及衡量真实 ROI 迈进。 ### [29:21](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=1761s) “氛围编程”与 AI 垃圾代码的隐患 主持人 Yazan “Yaz” El-Baba 分享了他在最近一次董事会上的奇特见闻：某家积极拥抱 AI 辅助工具的被投企业意外延误了产品交付截止日期。公司在汇报幻灯片上给出的延误理由让 Yaz 大开眼界——“我们的工程师氛围编程（vibe coded）过头了，以至于我们不得不花大量时间去清理生成的 AI 垃圾代码（AI slop）”。Yaz 由此向嘉宾抛出疑问：随着企业对这些 AI 辅助工具产生过度依赖，工程师是否正在放弃专业判断与深度思考，进而引发广泛的组织性风险？ ```mermaid flowchart TD accTitle: AI辅助开发中的工程质量把控 subgraph ToolAdoption["AI 编程普及"] A["工具全员分发<br/>(Cursor / ChatGPT)"] --> B["快速生成原型 / Vibe Coding"] end subgraph OrgMaturity["工程成熟度过滤"] B --> C{"工程组织成熟度"} C -->|"缺乏资深把关 (初创公司)"| D["AI 垃圾代码堆积<br/>错过交付节点"] C -->|"资深/Staff 工程师守护"| E["严格 Code Review<br/>阻止不合规代码合并"] end subgraph BusinessOutcome["业务成效"] E --> F["产能提升 10 倍<br/>拓展新业务线"] end ``` Angie Jones 率先回应，指出这种现象高度取决于工程团队的成熟度。在她所在的企业，团队中拥有一批优秀的资深工程师，他们会毫不妥协地坚守底线，明确表态“绝不把这种垃圾代码发布上线”，在风险蔓延至生产环境之前将其掐灭。然而，很多新兴的“AI 原生（AI-native）”初创公司可能会抱有不切实际的幻想，认为只要掌握了“氛围编程”的技巧，就不再需要经验丰富的资深开发者。这种团队往往严重匮乏基础工程素养，甚至连识别和捕获 AI 垃圾代码的能力都不具备。 Keith Soura 对此深表赞同，提到自己团队中的 Staff 级别工程师同样抱有强烈的代码所有权意识，绝不允许质量欠佳的代码混入代码库。但他也指出，其董事会采取了完全不同的开明策略：与其用 AI 缩减人员编制，董事会反而认为“既然工具能带来 10 倍的生产力提升，那就招募更多人员、打造更多产品并开拓新业务”。在实际操作中，公司将 Cursor、ChatGPT 等工具的许可证发放给了全公司所有人——不仅是工程师，还包括产品经理、运营人员乃至所有高管。甚至连公司的代码库也对全员开放，打破了传统的技术壁垒。Keith 戏谑道：“连我们 CEO 现在都拿到了 Cursor 许可证，上帝保佑我吧。”全员上手的体验让每个人都直观看到了 AI 的上限与局限：当你听说它有多神奇而亲自尝试时，才会意识到如果没有深厚、敏锐的专业辨别力，根本无法理解底层究竟发生了什么。 - Yaz 首次在董事会报告中接触到“vibe coded too hard”（氛围编程过猛）与“AI slop”（AI 垃圾代码）的行业新说法。 - Angie 警告缺乏资深工程师把控的年轻初创企业容易丧失代码审阅的基础工程能力。 - Keith 透露其公司向高管、产品、运营全员发放 Cursor/ChatGPT 权限，并开放代码库让大家认清技术真实边界。 --- ### [31:49](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=1909s) 销售场景下的 AI 泡沫与“反 AI SDR”立场 Roshan Chekuri 从销售与业务视角切入探讨。他坦言在内部应用 AI 提升效率方面，他们团队推进得相对审慎缓慢，而在听完前面关于代码失控的故事后，他更庆幸这种克制。Roshan 强调，在销售领域过度剥离“人性化元素（human element）”是非常危险的。他以自己的日常体验为例，指出目前的收件箱每天都被各种显而易见的“AI 生成的 SDR（销售开发代表）开发信”塞满。如果全行业都盲目狂奔向纯自动化，最终会导致用户对毫无温度的机械互动产生极度厌恶，反而无比怀念真正的人际沟通与销售温度。 Yaz 随即调侃道：“所以你现在是‘反 AI SDR’（Anti-AI SDR）阵营的？”Roshan 毫不犹豫地明确承认了这一身份。面对 Yaz 提到的其投资组合中某些正朝此方向发展的企业（如 Unify 等），Roshan 澄清优秀的销售赋能产品应该致力于改进流程，而不是简单粗暴地用 AI 取代真诚的人际交互。 - Roshan 直言对 AI 销售邮件充斥收件箱深感反感，明确表明“反 AI SDR”态度。 - 销售领域过度追求自动化可能导致丧失最具核心价值的真实人际互动体验。 --- ### [32:42](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=1962s) 展望未来：从无处不在的计算到开发者的付费转向 在小组讨论的尾声，Yaz 请三位嘉宾开启“水晶球预测”，探讨未来 12 到 36 个月内最有可能重塑 SaaS 行业的 AI 能力或市场趋势。Keith 回忆起幼年参加 IBM“带孩子上班日”的经历（他的多位长辈乃至祖母都在 IBM 工作）。当时 IBM 提出的核心愿景是“普适计算（ubiquitous computing）”——未来的计算将不再局限于捧在手里的笔记本电脑或智能手机，设备会隐形融入物理环境，计算与交互将变得无所不在且自然无感。过去几十年这一演进似乎遭遇了瓶颈，人们出差仍得扛着笨重的手机和三台笔记本。但生成式 AI 的出现正在打破基于点击（click）与触摸（tap）的传统图形界面，转向以语言、文字、视频为核心的全方位交互。Keith 举例说明，他们现在已经推出了持有执照的房贷审批专员（Loan Officer）专属视频分身（video avatar），能直接为客户办理抵押贷款业务，“计算正在主动迎合用户所处的场景”。 Roshan 则认为未来一到两年将是“辨别真金与骗局（grift vs. real）”的关键清洗期。过去一年，许多风险投资机构都在追踪企业引入 AI 后的第一年流失率（attrition metrics）。他非常期待看到哪些 AI 产品最终会被证明是“AI 界的 NFT 投机泡沫”，哪些才是真正具有商业壁垒的生成式 AI；好在目前投身该领域的顶尖智力资源远比当年的 NFT 和加密圈更扎实。 Angie Jones 提出了一个让技术界极为意外的逆转现象：在过往的软件生态中，开发者历来是最不愿意为软件工具付费的群体；而现在，大量开发者却毫不眨眼地每个月掏出数千美元购买 LLM Token。Angie 对这种高昂支出的持久性表示怀疑，并好奇开源大模型（Open-Source LLMs）以及 AI Agent（智能体）的崛起，最终会对传统 SaaS 商业模式造成怎样的颠覆与洗牌。 - Keith 预言界面将从点击式 UI 彻底迈向自然对话与视频交互（如具备展业资质的虚拟人房贷专员）。 - Roshan 强调未来 12 至 24 个月市场将刺破 AI 炒作泡沫，根据产品真实留存率甄别实质价值。 - Angie 指出开发者消费行为发生巨变（月掷千金购买 Token），但质疑其可持续性，并聚焦开源 LLM 与 Agent 对软件业的冲击。 --- ### [35:29](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=2129s) 推理成本下行假设与现实能源壁垒的冲突 针对未来商业模型的可持续性，Yaz 抛出了最后一个核心命题：当前许多早期初创公司的商业模式十分脆弱——“向客户收取 100 美元，转头就付给 OpenAI 或 Anthropic 120 美元”。这些公司维系业务的核心假设在于：大模型推理成本（inference costs）未来必然趋近于零。它们信奉“充裕思维（abundance mindset）”，试图在技术前沿不计成本地做极限尝试，赌未来模型账单大幅缩水后，自己能凭借最佳产品占领最牢固的客户粘性。 Keith 明确反驳了这种“推理成本归零”的盲目乐观。他分析指出，任何技术成熟都会带来单位成本下降，但大模型推理本质上与物理世界的实体资源深度绑定——尤其是电力和能源消耗。而在当前全球现实中，能源成本绝非处于下降通道。Keith 直言：“在一个依然需要为物理电网支付高昂电费的世界里，我绝不会把注押在‘推理成本会趋近于零’上。”除非人类迎来太阳能彻底普及或可控核聚变商业化的无限能源未来，否则计算的物理代价始终是悬在商业模式之上的现实硬约束。全场讨论在对未来新能源与基础设施投资的思考中圆满收尾。 | 发展路线与核心假设 | 优势与推进策略 | 潜在风险与现实瓶颈 | |:--- |:--- |:--- | | **前沿充裕模型（初创公司）** | 不计成本追求极致体验，抢占先发优势与核心客户留存率 | 毛利长期倒挂（收 $100 倒贴 $120），极端依赖推理单价暴跌 | | **物理资源约束视角（技术工程派）** | 务实评估算力与能源开销，追求兼顾真实 ROI 的架构设计 | 受到现实世界电力与能源成本硬约束，推理成本无法无限归零 | - 早期初创团队赌推理成本归零以换取高粘性，当前阶段普遍承受负毛利运转。 - Keith 认为电力消耗与电网账单构成物理世界硬壁垒，推翻了算力无限廉价的空中楼阁。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=4s) 会计工作确实非常枯燥，现在几乎没人读会计专业了 — Roshan Chekuri > [06:43](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=403s) 你不可能直接把一套深度整合业务系统的端到端 Agent 工作流卖给辛辛那提大主教管区的财务总监。 — Roshan Chekuri > [12:28](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=748s) 在这里你绝不能搞砸，因为赌注实在太高了。 — Keith Soura > [17:16](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=1036s) 在过去的 50 天里，我们调整了大约 50 次定价策略！ — Roshan Chekuri > [21:41](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=1301s) 哪怕我们现有的房贷产品不合适，我们会告诉客户如果花三四个月调整几项财务指标就能符合条件，并提供辅导；或者通过我们自研的开放贷款平台，直接把客户引流到其他有更合适产品的银行。你永远不对客户说不，他们因为信任你而最终还会回来。 — Keith Soura > [29:21](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=1761s) 我们的工程师氛围编程（vibe coded）过头了，以至于我们不得不花大量时间去清理生成的 AI 垃圾代码（AI slop） — Yazan "Yaz" El-Baba > [29:21](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=1761s) 绝不把这种垃圾代码发布上线 — Angie Jones > [29:21](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=1761s) 连我们 CEO 现在都拿到了 Cursor 许可证，上帝保佑我吧。 — Keith Soura > [35:29](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=2129s) 向客户收取 100 美元，转头就付给 OpenAI 或 Anthropic 120 美元 — Yazan "Yaz" El-Baba > [35:29](https://www.youtube.com/watch?v=QhlsvI-tLGU&t=2129s) 在一个依然需要为物理电网支付高昂电费的世界里，我绝不会把注押在‘推理成本会趋近于零’上。 — Keith Soura ## Mentioned Yazan "Yaz" El-Baba, Angie Jones, Keith Soura, Roshan Chekuri, Emergence Capital, Salesforce.com, Bolt, Bland, Together AI, Block, Inc., Square, Cash App, Tidal, Afterpay, Better.com, FloQast, Goose, Model Context Protocol (MCP), Cashbot, Square AI, Betsy, Alteryx, Y Combinator (YC), Cash Pool, John Woodard, Macy's, Cursor, ChatGPT, Unify, IBM, OpenAI, Anthropic ## Video description Keith Soura, VP Engineering, Better.com Roshan Chekuri, GM, AI, FloQast Angie Jones, Global VP, Developer Relations, Block, Inc. Yazan "Yaz" El-Baba, Partner, Emergence Capital Abstracts This session will explore how AI is transforming the SaaS industry by enhancing product offerings, improving customer experiences, driving operational efficiency, and changing SaaS pricing models. Panelists will discuss how AI is being used to power advanced analytics, automate customer interactions, and personalize services within SaaS platforms. Learn how AI-driven tools are enabling SaaS companies to innovate faster, optimize user retention, and scale more effectively. The panel will also address the challenges of integrating AI into SaaS products and the strategies for leveraging AI to create competitive advantages in a rapidly evolving market. [Source](https://www.youtube.com/watch?v=QhlsvI-tLGU)

[0:04] Roshan Chekuri: 会计工作确实非常枯燥，现在几乎没人读会计专业了 [6:43] Roshan Chekuri: 你不可能直接把一套深度整合业务系统的端到端 Agent 工作流卖给辛辛那提大主教管区的财务总监。 [12:28] Keith Soura: 在这里你绝不能搞砸，因为赌注实在太高了。 [17:16] Roshan Chekuri: 在过去的 50 天里，我们调整了大约 50 次定价策略！ [21:41] Keith Soura: 哪怕我们现有的房贷产品不合适，我们会告诉客户如果花三四个月调整几项财务指标就能符合条件，并提供辅导；或者通过我们自研的开放贷款平台，直接把客户引流到其他有更合适产品的银行。你永远不对客户说不，他们因为信任你而最终还会回来。 [29:21] Yazan "Yaz" El-Baba: 我们的工程师氛围编程（vibe coded）过头了，以至于我们不得不花大量时间去清理生成的 AI 垃圾代码（AI slop） [29:21] Angie Jones: 绝不把这种垃圾代码发布上线 [29:21] Keith Soura: 连我们 CEO 现在都拿到了 Cursor 许可证，上帝保佑我吧。 [35:29] Yazan "Yaz" El-Baba: 向客户收取 100 美元，转头就付给 OpenAI 或 Anthropic 120 美元 [35:29] Keith Soura: 在一个依然需要为物理电网支付高昂电费的世界里，我绝不会把注押在‘推理成本会趋近于零’上。

[3:24] Yazan "Yaz" El-Baba: pound on the table and say we need to do more AI [0:04] Angie Jones: AI adoption to actually AI impact [6:43] Keith Soura: There are like 11 ways that we can get you into this home. [6:43] Roshan Chekuri: You can't sell an end-to-end agentic workflow that integrates your platforms to a controller at the Archdiocese of Cincinnati. [12:28] Angie Jones: really change the direction of someone's financial life [15:29] Keith Soura: Suddenly you're burning massive amounts of money on this call and this person is just having the time of their life. Finishes the call, hangs up, doesn't even finish the application. And we were all sort of like, we're going to need to put in some guards for this like real quick or the bill's going to be crazy. [17:16] Roshan Chekuri: changed pricing like 50 times in the last 50 days [19:38] Angie Jones: You're like, yo, I'm still on a [basic step]... what are you talking about? [21:41] Angie Jones: When you get that $2,000 bill at the end of the month for your vibe coding hobby... and you realize like, wait a minute, the same vendor that makes the LLM that's charging me per token is also the vendor that has the agent doing the task. Could it be going the long way on purpose? You know, maybe to run the bill up. [29:21] Yazan "Yaz" El-Baba: our engineers vibe coded too hard and we had to clean up a lot of AI slop [29:21] Angie Jones: Fortunately we have a lot of brilliant senior engineers who say no, I'm not shipping the slop. [29:21] Keith Soura: absolutely not, that's not going in my codebase [31:49] Roshan Chekuri: I'm anti-AI SDR here. [31:49] Yazan "Yaz" El-Baba: Oh, we need to work on our marketing because they do a great job. [35:29] Keith Soura: I'm not going to bet my money on inference costs going to zero while the computational world still has to pay a power bill.

[0:04] Hi everyone. Thank you for joining us [0:05] for our panel around leveraging AI [0:08] within SAS businesses. I am very lucky [0:11] to be the the moderator for this this [0:13] awesome panel that we have in front of [0:15] us. Um so I'm going to go around and do [0:17] introductions. I will start and I'll [0:19] pass it to my left where uh folks will [0:22] share who they are, where they work, and [0:24] what they do at that company. And then [0:26] we'll dive into the meat. So for me, I'm [0:28] Yaz Albaba. I'm one of the partners at [0:31] Emergence Capital. We're a B2B software [0:33] fund that started [0:35] over two decades ago at this point. And [0:38] and that's a it's a pretty salient point [0:40] because the firm started on one of the [0:43] biggest platform shifts that we've seen [0:44] in software, which is that the firm's [0:47] founders bet on the fact that software [0:49] would move from on premises to the [0:51] cloud. And the first check that [0:53] emergence capital ever wrote was in a [0:55] small crazy company at the time called [0:57] Salesforce.com. [0:59] We've obviously since seen multi-tenency [1:01] emerge and all the companies that have [1:03] come to bear from that. And so now we're [1:06] equally as excited, if not more excited [1:08] about this next big platform shift and [1:10] have been investing pretty heavily into [1:12] this next AI wave and companies like [1:14] Bolt, Bland, Klein, Together AI and a [1:17] number of others. really excited to be [1:19] discussing all of this with you both in [1:21] terms of how our SAS companies are [1:23] adopting AI in addition to more AI [1:26] native companies and what their [1:27] advantages are. [1:29] >> Hi, I'm Angie Jones and I am VP of [1:32] engineering specifically focused on AI [1:35] tools and enablement. Um, I work at [1:38] Block, which is the parent company, The [1:40] Square, Cash App, Title, Afterpay. Uh, [1:44] and my job is to teach everyone about AI [1:49] both outside and inside. So, a lot of, [1:51] uh, my role is AI enablement internally [1:55] for all of our employees. uh right now [1:58] I'm really focused on engineering and [2:01] moving that team from uh AI adoption to [2:05] actually AI impact and then I also lead [2:09] the developer relations team. So we [2:11] teach uh engineers all over the world [2:14] how to utilize things like AI agents uh [2:17] LLMs and things of that nature to [2:20] automate their workflows and improve [2:22] efficiencies. [2:24] >> Hi, I am Keith Sora. I'm from a company [2:26] called Better which works in consumer [2:28] finance primarily on mortgages and I'm [2:29] the VP of engineering. Uh my actual job [2:32] is to lead engineering and make sure [2:34] that everyone keeps building things [2:35] productively. Um but my my passion is [2:37] actually doing the engineering myself. [2:41] >> What's up guys? Uh my name is Roshan [2:42] Shakuri. I'm the GM of uh our AI [2:44] platform at a company called Flowcast. [2:46] We've been around for about 12 years. [2:48] I've been at Flowcast for eight. Uh [2:50] we're accounting transformation platform [2:51] that helps accounting teams close their [2:53] books faster, more accurately. Uh we're [2:55] in an interesting thing where we're [2:56] selling to a bunch of riskaverse [2:58] accountants where financial data and [3:00] security and auditability is super [3:02] important. Uh and uh it's funny I can [3:04] say this in here because like 40% of our [3:06] our company our former accounts but [3:08] accounting is super boring. No one's [3:09] studying it anymore. Uh 75% of people uh [3:12] of CPAs are retiring in the next 10 [3:14] years. And so how do you fill that [3:15] talent gap? AI has been served on a [3:17] platter very nicely. Uh but how do you [3:19] do it in a way that's secure, auditable? [3:20] Um and so we can talk a little bit about [3:22] that. Nice to meet you all. [3:24] Great. So, one of the first things that [3:27] I'd love to talk about is the topic that [3:29] has been in every boardroom [3:32] conversation, which is every board [3:34] pounds on the table and says we need to [3:36] do more AI. [3:38] What does that mean? [3:40] Is it running your business, your [3:42] operations more effectively? Is it [3:43] launching new features? [3:46] What does AI actually mean? And where [3:48] have you seen the ROI from deploying AI [3:50] within your businesses from your [3:51] experiences so far? [3:54] I would say yes to all of that. Um, we [3:58] need to [4:00] use AI to build our products. We need to [4:04] incorporate AI into our products. We [4:07] need our teams and employees utilizing [4:11] AI to move faster uh and develop more [4:14] innovative features. And so we've seen [4:17] uh at Block, we've seen ROI on all of [4:20] those fronts. Uh we have developed our [4:24] own open-source AI agent. Its name is [4:27] goose. And uh really cute, right? And uh [4:31] so so goose is used uh you know a lot [4:35] externally but internally all 12,000 of [4:38] our employees use goose no matter what [4:41] function they they work in. So [4:43] marketing, sales, finance, design, [4:46] engineering, uh executive assistants, [4:49] everyone uses this AI agent. And we've [4:52] uh embraced uh this open protocol, MCP, [4:56] which stands for model context protocol. [4:58] I see some heads nodding. Uh so we've [5:01] embraced that very heavily and we [5:04] utilize that with our agent to give our [5:07] employees all of the functionality that [5:09] they need to be able to use these [5:11] agents. So that's been very beneficial. [5:14] We've seen, you know, non-engineers be [5:17] able to stand up prototypes instead of [5:19] writing design documents, for example. [5:22] Uh we've seen people be able to get data [5:25] they need without needing to go to like [5:27] a business analyst for that data. And so [5:29] everyone is just able to unlock [5:31] themselves and move faster. We've also [5:34] uh integrated AI. We've released two AI [5:37] features uh in the last couple of [5:40] months. one in Cash App which is called [5:42] Cashbot. So Cashbot uh does our customer [5:45] service. We're seeing uh significant [5:48] success there. And then also in uh [5:51] Square. So Square is for business owners [5:53] spec specifically owners who sell [5:56] things. So uh these sellers have a [5:59] dashboard and we've embedded uh AI in [6:02] there called Square AI which allows them [6:05] to ask these questions in natural [6:07] language. We all know how complicated [6:09] any dashboard is for this sort of [6:11] functionality. So instead of digging [6:13] through things and trying to generate [6:15] reports, you can just say things like, [6:17] "Who are my top 10 sellers last month?" [6:20] You know, for example, or what's what's [6:22] my bestselling product? What am I about [6:25] to run low on? Can you order more? Like [6:27] those sorts of things just in natural [6:29] language. So we're seeing a lot of [6:30] return on investment on both fronts. [6:33] engineering uh employee [6:36] contributions and all of that as well as [6:39] in the products themselves. [6:43] >> I think uh for us we've kind of looked [6:45] at it as a margin engine and I think a [6:47] lot of the things that we've heard from [6:49] the board generally speaking is that it [6:51] can be used to automate a lot of things [6:52] that are traditionally very [6:54] humanpowered. Uh we we've chosen a kind [6:56] of interesting approach and one that I [6:57] like a lot which is that we actually [6:59] have all of the right people. They just [7:00] might not be in the right places. And so [7:02] what we like to do because it's mostly [7:04] mortgage and consumer finance which is a [7:06] very human um business is try to get [7:09] more time for those people to speak with [7:12] other people. That's the basis of our [7:13] business. And so a lot of our tools work [7:15] in I would say three categories. One uh [7:18] external to the consumer. Um over 50% of [7:21] America in 2024 was underserved in [7:24] banking. They didn't have a great access [7:26] to financial literacy information. uh [7:28] they didn't know about complex credit um [7:31] products and so applying for a mortgage [7:33] can be very scary and we like to give [7:35] them all of that access in a natural [7:37] language way which I think is really [7:38] interesting. The second is internally [7:40] very similar to what what you were just [7:41] saying. We have an internal um assistant [7:44] who we call Betsy who also happens to be [7:45] the person that answers our phone lines. [7:47] She does a lot. She's all over the [7:48] place. Yeah. uh and and Betsy helps our [7:52] loss spend less time thinking about [7:54] computation and more time actually [7:56] talking with the people who are applying [7:57] for mortgages and we're able to assess [8:00] thousands or tens of thousands of data [8:02] points instantaneously to help find [8:03] products that match someone's financial [8:05] situation. So before you might have been [8:06] like you know what you don't qualify for [8:08] this mortgage now we can be like there [8:10] are like 11 ways that we can get you [8:11] into this home which is a huge important [8:13] part of building wealth in the country. [8:16] Uh and then the third way is just for [8:17] our engineering and product teams. I [8:19] mean, I'm sure you guys have experienced [8:20] this, but the ability to take away some [8:23] of the wrote work and democratize access [8:25] to reporting, for example, or even [8:28] observation data um has been incredible [8:30] for us. It just has really sped up um [8:32] the engineering and product teams and [8:33] their ability to iterate on new new [8:35] features for the business. [8:37] >> Yeah, we're selling to uh CFOs and [8:39] controllers and we're seeing quite a bit [8:40] the VCs and private equity people are [8:42] breathing down the neck saying, "Hey, go [8:44] deploy AI solutions." If you think about [8:46] accounting, it's just a thousands of [8:47] bespoke journal entries that accounts [8:49] have to do and it's a bunch of data [8:50] manipulation and spreadsheets and like I [8:52] said, boring stuff. Um 10% of the [8:54] market, I'd say, is pretty educated on [8:56] that. There's a knowledge gap uh quite [8:58] clearly around AI in general. So for [9:00] like 10% of the market, we've deployed [9:02] endto-end agents that are, you know, uh [9:04] completely eliminating a process like [9:07] cash posting. Uh we had a company that [9:09] uh outsourced that to a shared service [9:11] center that spent $60,000 on it per year [9:13] and uh we sold them one agent that does [9:15] it in 5 seconds uh for for like $50,000 [9:19] recurring. But then the reality is like [9:21] 90% of the market uh doesn't understand [9:24] and and they just u use LLM to summarize [9:26] emails like you can't sell an endto-end [9:29] agentic workflow that integrates your [9:31] platforms to a controller at the arch [9:32] dascese of Cincinnati. And so we're [9:35] trying to figure out ways to leverage [9:37] some of like the Altterx type model and [9:39] the RPA model where uh we want to [9:41] educate the market and give accountants [9:43] an opportunity to upskill. Um this is [9:46] not taking an accountant's job. This is [9:47] um accountants need to learn how to [9:49] upskill to deploy things like this and [9:51] leverage AI and the future of accounting [9:53] is going to be that. So we have [9:54] certification programs that now we're [9:56] we're we're putting out there to learn [9:57] how to build agents within our platform [9:59] as well as other solutions. Um and so [10:02] that's that's kind of where we're seeing [10:03] the ROI. kind of the full endto-end [10:05] automation, but then this education [10:07] piece. [10:09] >> So, it sounds like each of your [10:11] businesses are not only running AI [10:14] internally, but kind of exposing that to [10:17] the end customer. Sometimes through a [10:18] handheld experience where it's assisting [10:20] your employees, other times where the [10:22] where the customers are interacting with [10:23] Goose or Betsy or whatever your agent is [10:26] called, if it has a name directly. One [10:29] of the things that we we often feel is [10:32] at odds now is speed and trust. [10:36] And the challenge is that we're seeing [10:38] companies AI native companies moving [10:40] faster than we've ever seen before. I [10:41] feel like every week there's a new [10:42] headline of some company I've never [10:44] heard of before going zero to 100 in [10:46] ARR, whatever that means. We can debate [10:48] it later. Faster than ever. And then you [10:52] also see these companies that have them [10:53] around also deploying AI, but there's [10:56] this risk now, right? Like everyone [10:58] talks about the distribution advantage [11:00] that you have as a larger company, but [11:02] you also have the other side of it, [11:04] which is loyal customers trusted you for [11:07] years and now you're rolling something [11:08] out that is prone to hallucinations. [11:12] How do you think about balancing [11:15] speed with the trust that you've [11:18] developed with your end users and [11:19] customers over the years of building [11:21] your businesses? [11:25] >> Who wants to go? Me. Um yeah, so I would [11:29] see there's three buckets in this. Um [11:31] one is like the enterprise solutions [11:33] that have been around for a long time [11:34] that are in huge tech debt and [11:36] innovators dilemma. It's really hard for [11:38] them to go deploy these types of things [11:39] and build these types of things. Uh the [11:41] second is you know unfortunately this is [11:43] a lot in a lot of cases a timing thing [11:45] like if you were in market at the right [11:46] time you have the benefit of [11:48] distribution but then you're also kind [11:50] of out of not in innovative dilemma yet [11:52] and you can still build. Then you got [11:54] all these like YC kids that are geniuses [11:55] that are just building stuff and you see [11:58] it when you drive across the Bay Bridge, [11:59] you live in San Francisco, it's like [12:00] everyone in their mother has AI on their [12:02] billboard. Um, so there's those three [12:04] categories. Um, for us, we spent a year [12:07] and a half in the research piece of the [12:09] R&D around this type of stuff because [12:11] again, we're selling to to accountants [12:13] that care about auditability and so we [12:14] we had to make sure that we were doing [12:16] it the right way. Um but yeah, it's [12:18] going to be interesting to see how it [12:19] unfolds over the next two, three years [12:20] around companies that have tech debt, uh [12:23] companies that are YC and companies that [12:25] you know are deploying this pretty fast. [12:28] >> Yeah. Um I think we both work in like [12:32] you know financial services. You can't [12:34] mess this up, right? [12:35] >> Um the stakes are too high and so we [12:39] have to be extremely careful in how we [12:41] roll this out. Um we do our due [12:45] diligence. We make sure like a lot of [12:46] the hallucinations [12:48] are because of lack of context. [12:51] Essentially, if uh if a LLM doesn't have [12:55] the context, it's going to make it up, [12:56] right? It wants to be right. And so, um [13:00] our job is to feed it as much context as [13:03] possible. And so, uh all of the products [13:06] that we've rolled out, we make sure that [13:08] it has like the same guides and [13:10] templates that like an employee who [13:13] might be answering that call would have. [13:15] given this information, it's really able [13:17] to uh to to give accurate uh results. [13:22] Same with uh the the square side, right? [13:25] And so we're making sure we're if we [13:27] need to develop MCP servers ourselves to [13:31] provide that context internally. We've [13:33] built over a hundred MCP servers for [13:37] various SAS applications that our [13:39] employees use that we need for our [13:41] products um to ensure that these agents [13:44] do not have to guess anything. They have [13:47] the information readily available to [13:49] them. With that, we also uh heavily [13:53] involved our infosc folks, our [13:56] compliance folks, our security folks to [13:59] make sure that we're not missing [14:01] anything. And then as fast as this all [14:03] is moving, we know better um than to [14:06] just race to it. And so we're we're [14:08] moving we're moving quickly quicker than [14:11] we have since we've been a little [14:13] startup. But we also are being very [14:16] cautious and kind of rolling it out like [14:18] a beta, watching it very carefully, [14:21] tweaking what we need to before uh [14:24] larger rollouts. [14:27] >> I have a handheld bank now. It's very [14:28] exciting. [14:30] >> Doing a concert. I I couldn't agree [14:31] more. I think the um the challenge [14:33] especially in the financial services [14:34] industry is that you're not just giving [14:36] someone some bad citation for like a [14:38] research paper, right? you could really [14:39] change the direction of someone's [14:41] financial life effectively by giving the [14:44] wrong information. And so we've taken [14:45] the same approach I think that you've [14:46] been describing which is a lot of care [14:48] around gating and monitoring what the [14:51] system is doing. Um there's a lot of of [14:53] QC that goes into it and we even have [14:56] other models like more sophisticated [14:58] models monitoring um right on a [15:00] real-time basis both voice and the text [15:02] and the reasoning things that we're [15:04] we're processing. But I think the core [15:05] of what we've done is most of our [15:07] computations because we're an [15:08] underwriting which I think is a little [15:09] bit easier than probably what you do is [15:12] really federally mandated or [15:13] statemandated and we're able to describe [15:15] that in terms of like a mathematical [15:17] dependency graph. And so we can keep the [15:19] reasoning model on Rails by making sure [15:22] we're also computing the math and [15:23] checking the results and then kind of [15:25] allowing it to teach itself. That's been [15:27] a large part of it. [15:29] So, so I'm going to ask maybe a slight [15:31] curveball question to the group, which [15:34] is we know these systems don't always [15:37] work as intended. You always have QE, [15:39] QA, QC, all the scaffolding, [15:41] >> all the cues, [15:42] >> all the cues to catch quality issues. [15:46] Does anyone have a story of a time when [15:49] a bug didn't get caught and how the team [15:52] was able to respond quickly to [15:54] addressing it to make sure it doesn't [15:56] impact your end users or your employees [15:58] if it was an internal bug? [16:02] I have a funny one, I guess, and our [16:04] product manager for AI is sitting right [16:06] in the front row glaring at me, but um [16:09] but I think my favorite one uh was when [16:13] at at the beginning of this when we were [16:14] first starting to handle voice calls [16:16] because a lot of people again like over [16:18] half of America works um non-traditional [16:21] hourly jobs and so it's really important [16:22] to be able to help them with mortgages [16:24] not during 9 toive banking hours. So one [16:26] of the first things that the team rolled [16:28] out was the voice agent. So if you call [16:29] our main line today, you will get a [16:31] synthetic person. Betsy answers the [16:32] phone and she can take you all the way [16:33] through locking a rate. She can take [16:35] your application, you can call back in [16:36] and authenticate with her. Um, at the [16:39] beginning, we didn't think that people [16:42] having been fully told that this is not [16:43] a real person at the beginning of the [16:45] call would engage the AI bot and just [16:47] idle conversation while they were [16:49] looking for information. Um, and there's [16:51] a great recording that uh I've heard [16:53] from from Corey uh where someone is in [16:55] their car and they keep going in and out [16:57] of signal. And so Betsy is just making [16:58] small talk about sports and movies and [17:01] all and suddenly you're burning, you [17:02] know, just massive amounts of money on [17:04] this this call and this person is just [17:05] having the time of their life. Finishes [17:07] the call, hangs up, doesn't even finish [17:09] the application. And we were all sort of [17:11] like, we're going to need to put in some [17:12] guards for this like real quick or the [17:14] bill's going to be crazy. [17:16] >> That's great. Pricing these tools, what [17:18] what does it look like to the end [17:20] customer when you ultimately end up uh [17:23] distributing these features? [17:26] for us. Um, we haven't gotten there yet, [17:30] honestly. Uh, we are learning with our [17:33] customers. Uh, our Square AI feature is [17:36] in beta, so we're, you know, watching [17:38] closely. We're learning, we're [17:40] observing. Uh, our cash buy also not [17:44] necessarily something you would charge [17:45] for. So, these give us a great [17:48] opportunity to see what our customers [17:50] actually need, what friction points they [17:53] might run into, and then iron it out. [17:55] And all of this is data that will [17:57] influence like future product and [17:59] pricing decisions for new things that we [18:02] might release. Um, but for now it's all [18:05] about uh experimenting. [18:09] >> Yeah, I I would agree. Most of our items [18:11] that we're building are part of our our [18:13] kind of SAS platform and are just things [18:15] that we think of as a margin machine, [18:17] going back to my earlier comment. And so [18:19] the better we can be at efficiency, the [18:22] less cost there is in originating a loan [18:24] or a mortgage, the better the pricing we [18:26] can offer to the consumer, the more [18:27] people we can get into a home basically. [18:30] And so we're kind of measuring success [18:32] on on that decreasing costs and increase [18:34] in efficiency for the end consumer and [18:36] eventually the the throughput that we're [18:38] able to originate with the same size [18:40] team. [18:41] >> We've changed pricing like 50 times in [18:43] the last 50 days. [18:45] Uh it's funny because the two people [18:46] working on pricing are sitting right [18:48] here and and they've built like the best [18:50] pricing strategy, consumption models, [18:52] you know, uh with tools and building [18:54] blocks and all this type of stuff, but [18:55] then you talk to a CFO and they're like, [18:57] I want predictability. Um and so we've [19:00] we've made a bunch of adjustments. We've [19:01] made adjustments to go to market [19:02] strategy. So we're still we're still [19:03] learning about it. Um I think again I [19:05] want to go back to this theme of [19:07] education in the market. Like I think [19:08] the before actually deploying agents and [19:10] AI in the market, it's there's this huge [19:13] knowledge gap around education of what a [19:15] consumption model means. How are you [19:17] picking LLM? Why are you iterating on [19:18] LLM? Like [19:19] >> all that type of stuff is very complex. [19:21] And so I think it it's upon the seller [19:23] to also educate which is new. [19:26] >> Yeah, that's a big part of my role too [19:29] is just educating uh the market, right? [19:32] Um, and I agree with you, even folks who [19:36] use these tools, you know, they're only [19:38] scratching the surface of what they can [19:40] do. And so to try to sell them like, oh, [19:43] I have this advanced agent that can do [19:45] X, Y, and Z. You're like, yo, I'm still [19:47] on a like, what are you talking about? [19:49] You know what I mean? So, I can imagine [19:50] that being a hard sale. So, there is a [19:52] lot of education that's needed around [19:55] what these tools can do, but also what [19:57] they can't do as well, right? Even [19:59] engineers who understand logically how [20:02] LLMs work are still shocked that you [20:06] know it got it wrong [20:08] even though they know these are not [20:10] deterministic systems you know. [20:12] >> Yeah. Yeah. I think transparency too [20:13] when you're educating is the most [20:15] important thing when we've been [20:16] transparent with our customers around [20:18] this is agentic. This is not aic why why [20:20] we're pushing this is potentially [20:21] marketing it as an agent when it's still [20:23] deterministic because we have to do [20:24] auditability like that stuff that stuff [20:26] works really well in the because they [20:27] just want transparency. They got to sift [20:29] through the noise. There's so many [20:29] brokers out there. [20:31] >> Yeah. [20:31] >> I think there's so much about the first [20:33] question um that that was asked around [20:36] what the boards are asking for and [20:38] there's a kind of a feeling that perhaps [20:40] this is a silver bullet. You can just [20:41] turn on AI and everything will magically [20:43] be okay. And I to your to your point, [20:45] what can it do? What can't it do? And [20:47] how do you demonstrate that numerically [20:48] to your stakeholders or in our case like [20:50] regulators, right, who have very real [20:51] questions about how it's making [20:53] predictions or how it might be involved [20:54] in the underwriting process and ensuring [20:56] fairness and equity and things like [20:57] that. So I think education is probably [20:59] one of the most critical components. I [21:01] agree with you. [21:01] >> Yeah. Cuz a lot a lot of accounts that [21:03] we sell to are asking if they can do [21:05] some of this stuff in GPT just like [21:07] consumer GPT. And then a lot of the [21:09] younger companies that are coding really [21:11] fast are just trying to get something [21:13] quick to market and a lot of times [21:15] they're an LLM rapper. Um, we have a [21:17] competitor down market and they've done [21:20] an incredible job marketing and making [21:21] the market feel like they're leading in [21:23] AI, but a lot of their customers are [21:25] relying on accounting guidance through [21:27] LLM wrapper. And if it's [21:29] nondeterministic, even if you're [21:30] filtering it by, hey, what's EY's [21:32] guidance for ASC, whatever, it's it's a [21:35] different answer every time. So, I think [21:36] accountants and other people are just [21:38] are not aware that, you know, these are [21:39] these are kind of dangerous things. [21:41] >> Yeah. which which is a great segue into [21:44] one of my favorite topics to ask [21:46] founders, which is how are you any [21:48] different from that AI native company [21:51] that just showed up last year or earlier [21:53] this year? They're taking the market by [21:56] storm. They're all over your LinkedIn [21:58] feed. [21:59] >> Yeah, [22:00] >> they are doing things light years ahead [22:02] of the slower, larger company in the [22:05] space. How do you counterposition [22:08] against that? the the competitive [22:09] environment is becoming increasingly [22:11] more difficult with startups now pushing [22:13] more aggressively not caring about some [22:15] of the repercussions because they don't [22:17] have a base that they could anger. So [22:19] when thinking about competitive [22:20] differentiation what comes to mind how [22:22] well how are you all playing in this [22:25] like ever changing landscape? [22:28] You know, this might be a little [22:29] controversial in in in a in a SAS panel, [22:33] but uh [22:34] >> I see [22:35] >> we we actually [22:39] are doing the opposite of what most [22:41] companies are doing, which is uh aiming [22:44] for vendor lock in. Instead, we're [22:47] offering interoperability, [22:50] user choice. Um and these seem to be [22:53] working really well in our favor. So, we [22:55] just uh launched Cash Pool for Cash App [22:59] uh last week. And this is where, you [23:01] know, you can kind of share a build or [23:04] bill money together with a group of [23:06] friends. But guess what? They don't all [23:08] have to be on Cash App. They can be on [23:10] any platform. Those sorts of things [23:13] people really like. Like, you're not [23:15] trying to lock me into this. Same with [23:17] Goose. Goose is open source, but it's [23:19] also LLM agnostic. Bring whatever model [23:22] you want. you know, we're not making the [23:24] model and the agent. You know, if you [23:27] think about this, um, sooner or later [23:31] people are going to start figuring out, [23:33] I'm not pointing any fingers, I'm just [23:34] talking, right? Um, theoretically, uh, [23:38] but when you get that $2,000 bill at the [23:42] end of the month for your vibe coding [23:44] hobby, right? And you realize like, wait [23:47] a minute, the same vendor that makes the [23:50] LLM that's charging me per token is also [23:54] the vendor that has the agent doing the [23:56] task. Could it be going the long way on [24:00] purpose? You know, maybe to run the bill [24:02] up. So like these sorts of questions I [24:05] think are going to come to the surface. [24:07] And so for us, we offer products that [24:09] allow people to opt into what they want, [24:12] bring other features that they might [24:14] want from other vendors and like [24:17] basically design your own custom [24:20] product, you know. So that's how we're [24:23] we're standing out. [24:24] >> Some would call it building blocks. [24:25] >> Yeah. [24:31] >> Uh yeah. So um for me, I'll I'll look at [24:35] it from both directions. So for us, we [24:39] we're built on our brand. We have 3,500 [24:41] customers, but yeah, we have a bunch of [24:42] ankle biters that are coming and [24:43] building really cool things really fast. [24:45] Um, and so I I would just put brand at [24:48] the forefront and make sure that you're [24:50] executing for those customers. What's [24:51] exciting for us is this whole education [24:53] thing again, uh, and certifications. Um, [24:56] John John Woodard here on my team, I [24:58] call him JohnBot. He's like the smartest [25:00] person I've ever met. uh he runs um [25:02] these immersive onboarding programs to [25:04] uplevel our uh our customer base on how [25:06] to build agents and at the end of it um [25:09] each of the accountants that we've [25:10] identified to become builders build an [25:12] agent within the platform. So then when [25:15] you have integrations with the with the [25:16] entire platform that they're using and [25:18] they've built agents that they've [25:20] touched within their training program [25:21] that are their processes, it gets really [25:23] sticky. Um, but if I'm a if I'm a down [25:26] market competitor that's coming up fast, [25:28] I would just target one best and breed [25:30] thing heavily. Uh, keep it simple around [25:33] AI application, make sure the audibility [25:35] is there. Um, and just drive it that [25:37] that works really well. [25:39] >> I think on our side, it's a little bit [25:40] of a combination between the education [25:42] and and the building blocks. Um, we've [25:44] definitely taken the approach of unlike, [25:45] you know, going to a more traditional [25:47] like top 10 bank where they're going to [25:49] sell you whatever product they can to [25:50] keep you inside the bank, we focus very [25:52] much on consumer education. So, we're [25:54] looking at someone's entire financial [25:56] history by the time that they're halfway [25:57] through the application, and we can very [25:58] easily say things like, you know what, [26:00] none of the products that we're offering [26:01] from our one of our mortgage banks is [26:03] going to work. However, we happen to [26:05] know that if you do a couple of these [26:06] things in about 3 or 4 months, they will [26:07] for you. So, if you're interested in [26:08] that, we can coach you through it [26:10] education wise. Or secondarily, we we [26:12] have a a marketplace and like kind of a [26:14] platform that we've been offering now [26:15] for the last 8 months or so on our [26:17] homebuilt software that allows us to [26:19] sell loans from other banks. So if other [26:22] banks want to be a part of this, we can [26:23] just say, you know what, actually we [26:24] know this other bank has a better [26:26] product for you. We can tell already. [26:27] Why don't you go over there and get the [26:28] thing that you need? Because that sort [26:30] of uh what I like to call if anyone [26:31] remembers the history of Macy's. That's [26:33] the history of Macy's, right? You never [26:35] said no. You always help the person get [26:36] what they needed even if it wasn't where [26:38] uh you were and they would come back to [26:40] you because they trust you. And that's [26:41] very much kind of the approach I think [26:42] that we've been taking. [26:45] So, so I want to shift gears a little [26:47] bit uh more away from external and now [26:50] into internal which is the internal [26:53] usage and adoption of AI. [26:55] Often we're discussing or seeing case [26:58] studies where it says developer [27:00] productivity has 10xed with AI coding [27:03] tools or vibe coding. And so naturally [27:07] the board comes back and says we need [27:09] onetenth the number of of engineers [27:12] because everyone's a super engineer now [27:14] thanks to the help of claude code or [27:17] client or cursor or insert AI coding [27:20] tool here. I'm curious to get your [27:22] reactions maybe within coding as a [27:25] specific example, but if you don't have [27:27] an example for coding more broadly where [27:29] you're seeing the promise of 10x 100x [27:33] improvements, where is it actually today [27:35] with the current state of models and [27:37] scaffolding? [27:39] >> So, this is where I live. Um, I can tell [27:43] you that a lot of that is fictional um [27:47] across the industry. There's definitely [27:50] some improvements. I wouldn't say 10x [27:52] like it feels that way at some points [27:54] but then you realize okay not. Um [27:58] uh but I I I think that [28:02] a lot of engineers and this is not just [28:04] at my company. I I go all over the world [28:07] talking to engineers and teaching them [28:09] and about AI. A lot of them are very [28:12] much so stuck at the code generation [28:14] phase again with the education piece, [28:17] right? even though I know that these fun [28:19] these these uh capabilities are [28:22] available to me uh I don't know I just [28:24] don't have a need for them is what they [28:26] think right and so you're using these [28:28] tools um at a very kind of beginner [28:32] level uh and then you're not seeing the [28:35] benefits so then you say oh yeah it [28:36] saves me about an hour five hours maybe [28:41] uh tops a week that's not 10 times you [28:44] know what I mean and so uh part of my [28:47] role now is again moving us from [28:50] adoption to impact. So how do we [28:53] integrate this further? How do we get [28:55] beyond the code? How do we like embed [28:59] this into developer workflows if speed [29:02] up like our deployment processes with [29:05] AI? Like how do we think more creatively [29:08] about what we're doing? because a lot of [29:10] these founders and CEOs and board [29:12] members are thinking, "Hey, I invested a [29:15] ton of money into uh AI and I'm not [29:19] necessarily seeing the return on [29:21] investment yet, you know, so how do we [29:23] get there?" And so maybe one follow-up [29:26] question to that, which is I I just had [29:29] one of my most recent board meetings [29:31] with a company that has actively [29:33] embraced these tools within their [29:35] engineering org, and they missed a [29:37] product deadline. And the reason they [29:40] put on the slide, I was the first time [29:42] I've heard this expression used before, [29:45] uh, it was our engineers vibe coded too [29:47] hard and we had to clean up a lot of AI [29:50] slop. [29:52] >> Yeah. [29:52] >> Is like I'm I'm curious. Is that [29:54] starting to show up within these [29:56] organizations where there's too much [29:58] dependency on on these tools and people [30:01] no longer insert their own judgment and [30:04] and work into [30:05] >> you know I think that really depends on [30:06] the maturity of the engineering org. So [30:09] fortunately we have a lot of brilliant [30:11] senior engineers who say no [30:16] I'm not shipping the slop you know and [30:18] they stop before it gets to that. But I [30:20] can imagine some of these like younger [30:22] AI native companies that are like, [30:24] "Yeah, yeah, we don't need experienced [30:26] engineers. Do you know how to vibe code? [30:29] Come on." Um, then they're missing like [30:31] those those foundational skills to be [30:34] able to even catch the slops. [30:38] I'll let somebody else talk. [30:39] >> I I completely agree with you. I mean, [30:41] we have we luckily have some really [30:42] great staff engineers who are exactly [30:44] like absolutely not. That's not going in [30:46] my codebase, right? You know, they take [30:47] a lot of ownership of it. But I think um [30:50] one of the lucky things that that we [30:52] have is that our board took the opposite [30:55] approach which was oh it can make you 10 [30:56] times more productive hire more people [30:58] and make more products like go make [31:00] other businesses which is not what I was [31:02] expecting and was awesome uh probably [31:04] unusual but was pretty fantastic and I [31:07] think that combined with the fact that [31:09] everyone in the organization has been [31:11] given licenses for these tools not just [31:13] engineers our product managers everyone [31:15] in ops every executive has licenses or [31:18] chatgpt or cursor or whatever. Our [31:20] codebase is open to everyone that works [31:22] there. It's not this like siloed magical [31:24] thing that only certain engineers can [31:25] touch. People can see what the the AI [31:29] can help with very easily. Uh and they [31:30] can also see where it doesn't help them [31:31] at all. I mean, even even our CEO has a [31:34] cursor license now. God help me. And um [31:37] yeah and and it's a but it's a really [31:39] good learning tool because you think [31:40] that you can do something or you hear [31:41] that it's so easy and then you go and [31:42] you try to do it for yourself and [31:43] suddenly you realize you do actually [31:45] need some sort of discriminating [31:46] knowledge to to understand what happens. [31:49] >> Yeah, it's probably more of your guys' [31:51] domain coming from your side. Uh mine is [31:53] more from a sales perspective. Um we're [31:56] frankly pretty slow when it comes to uh [31:58] applied AI efficiency internally which [32:00] I'm happy about because of these kind of [32:02] stories. [32:04] thematically like I I'm very scared from [32:06] a sales perspective of getting away from [32:08] the human element. Like I've just [32:10] noticed that my inbox is packed with SDR [32:12] emails that I know are AI. Um and so I [32:15] feel like if people adopt that way too [32:17] quickly, it's going to get to a point [32:18] where people are going to miss the human [32:20] element of an interaction and sale. [32:22] >> So you're anti- AISDR. [32:25] >> I'm anti-ASISDR here. [32:27] >> Oh wow. [32:28] >> Yeah. Yeah. I know one of your portfolio [32:29] companies is going potentially towards [32:31] that, but uh Unifi is a great company. [32:34] It's not they're not AISDR. They're more [32:35] around impro what is it called? [32:38] >> Oh, we need to work on our marketing [32:39] because [32:41] >> um [32:41] >> they do a great job. [32:42] >> So So I'll ask my my last question which [32:45] is the looking into the crystal ball. If [32:49] you were to look ahead, what AI powered [32:51] capability or market trend uh do you [32:55] believe is going to reshape our industry [32:57] the most? [32:59] And timelines are hard because things [33:01] are are moving at such a blistering [33:03] pace. But within the next 12 months, [33:08] >> I I think when I see what's happening in [33:12] the next 12 months, next 24 months or 36 [33:14] months, I think back to when I was [33:15] younger, uh I have a long history of [33:17] family members that worked for for IBM [33:20] all the way up through my grandma. And I [33:22] remember going when I was young to a [33:24] take your kids to work day and one of [33:25] the things that they talked about was [33:27] ubiquitous computing. the idea that [33:29] having a laptop or a phone or anything [33:31] else was eventually going to kind of [33:32] vanish and you simply would have [33:34] computing and everything around you. It [33:35] would become natural to interact with [33:37] the internet or whatever else. And I [33:40] think that for a while we were kind of [33:41] heading in that direction. We kind of [33:43] hit a stopping point. Uh we're still [33:45] toing around iPhones. They're still kind [33:46] of heavy in my opinion. Um and laptops [33:49] and all that sort of stuff. I brought [33:50] three with me, you know, for compliance [33:52] reasons for the different things that [33:53] we're doing. Um, but I think that what [33:56] I'm seeing with AI is this sort of idea [33:59] that the the UI and the UX is getting [34:01] removed from having to be click- based [34:03] or tap based. We're moving towards that [34:04] ubiquitous environment. You can speak to [34:06] it. You can text it. You can video chat [34:08] with it. We even have a a video avatar [34:10] of a licensed LO that can originate a [34:13] mortgage for you, for example. So, this [34:15] idea that computing meets you where you [34:16] are, I think is sort of where we're [34:18] headed with this technology. [34:22] I think it's going to be exciting the [34:23] next 12 to 24 months to see um what's [34:26] like what's grift and what's real in [34:28] terms of what people deploy. I think a [34:30] lot of VCs have metrics around in the [34:32] first year of deploying AI how much [34:34] attrition there's been with some of [34:35] those products. So I'm excited to see [34:37] that like what is the what is the NFT of [34:39] the world and what is what is real [34:40] generative AI. I think we have smarter [34:41] people working on this than the NFT and [34:43] crypto world. Um so hopefully it goes [34:45] well. [34:48] Um, you know what's what what's [34:50] fascinating to me is um [34:54] so you know I I live in the the [34:57] engineering world and developers don't [34:59] want to pay for software. They never [35:00] have wanted to pay for software and now [35:04] they are willing to pay thousands of [35:06] dollars a month on like LLM tokens. I [35:10] find that absolutely fascinating, but [35:12] also I'm not sure how long this is going [35:14] to last, right? And so I wonder about [35:17] things like um open-source LLMs uh as [35:21] well as agents, sorry. Uh like like what [35:25] what will that do for the the SAS [35:28] market? [35:29] >> Yeah. And and maybe one follow-up [35:31] question to round things out. One of the [35:34] arguments you hear from the earlier [35:36] stage companies that are, you know, pay [35:39] me $100 and I'll pay anthropic or open [35:41] AI $120 as their business model. Their [35:44] argument for sustaining those margins [35:46] over time is the fact that they believe [35:48] inference costs are going to zero over [35:51] time. And so their their thought process [35:53] is let's think in an abundance mindset. [35:57] Let's throw as much as we can at the [35:58] frontier, see what's even possible to [36:00] begin with [36:01] >> and then over time our bill will go down [36:03] and we'll have the stickiest and most [36:04] retentive customer base. Curious to get [36:07] your reactions to both of those points. [36:09] Cost of inference in addition to if you [36:12] have the best product and you're pushing [36:13] the frontier, you will have a lead and [36:15] maintain that lead over time. [36:19] I think um inference costs will go down [36:22] like any sort of technology, but they're [36:24] deeply tied to real actual world [36:28] resources. In this particular case, [36:29] energy and energy costs are not going [36:32] down right now. Um and so I think, you [36:34] know, perhaps in the future, uh in a in [36:36] a like a wonderful future in which [36:38] everything is powered by solar and [36:40] infinitely renewable sources, it'll be [36:41] great or fusion or some other other [36:43] thing. But for now, uh, I I would just [36:45] say that like I'm I'm not going to bet [36:47] my money on in first class going to zero [36:49] while the computational world still has [36:52] to pay a power bill. [36:54] >> Maybe this will finally get people to [36:56] invest in these sorts of things.