CFOs Retain SAP Core Ledgers as AI Agents Absorb Junior Accounting

AI智能体接管初级会计，CFO仍保留SAP核心总账 随着自主智能体接管日常流程，财务高管正暂停初级岗位替补招聘 专家小组探讨了企业级AI如何推动企业财务从滞后的事后报告向发票级细粒度的动态、持续预测转变。发言嘉宾强调，与其追求常规的人工成本套利，不如聚焦数据流转速度与分析深度，同时直面劳动力结构性挑战，例如以往用于培养未来财务总监的初级入门岗位的消失。在权衡自研还是采购的两难选择时，与会者建议在保持不可变且可审计的核心总账系统稳定的同时，在业务边缘部署敏捷、自主的AI智能体。最终，建议CFO针对特定成效驱动型瓶颈制定举措，确立专门的高管机制统筹智能体调度，并有意识地将节省下来的运营工时重新投入到战略性业务伙伴建设中。 财务组织已迅速从持怀疑态度转变为将AI智能体视为协作团队成员，重心也从孤立的概念验证转向管理规模化的运营智能体集群。 (0:04) 财务AI的首要投资回报并非立即削减人员编制，而是数据流转速度——将报告延迟从数月压缩至实时的持续预测。 (5:08) AI模型能将月末波动分析细化并自动化至单张发票级别，从而识别出人工抽样通常会遗漏的细微运营异常。 (5:08) 企业AI的应用落地划分为三个结构层级：个人生产力助手、团队级组织智能体，以及贯穿数据湖的综合企业智能。 (10:55) 耗资数百万美元的传统数据清洗往往并非LLM落地的必要前提，LLM更像人类员工，通过流程指南和最新样本数据即可快速上手，而非依赖僵化的历史数据集。 (12:35)

Finance leaders argue that the true enterprise value of AI lies in eliminating reporting lags and enabling continuous forecasting rather than aggressive headcount reduction, though scaling these systems demands new orchestration models and risks eroding traditional talent development pipelines. Compressing Reporting Lags Over Labor Arbitrage (5:08) Panelists Mayank Sharma and Yoav Naveh emphasize that finance AI generates returns primarily by increasing data velocity rather than mimicking customer support's mass labor arbitrage. Because corporate finance departments are relatively small—often consisting of 5 to 20 people managing multi-threaded responsibilities—immediate headcount cuts are difficult to realize. Instead, automating data aggregation down to individual invoices shifts finance from delayed, post-facto monthly reporting to continuous, dynamic forecasting, empowering teams to act as proactive strategic partners. Jonas Melton adds that AI agents fundamentally expand analytical depth during month-end closes. Rather than relying on manual sampling, automated models can review transaction-level variance and flux patterns at scale, surfacing subtle operational anomalies that human reviewers typically miss. The Vanishing Junior Layer and Succession Dilemma (22:15) Repetitive, transactional tasks in accounts payable and accounts receivable face direct downward pressure, as demonstrated by EnergyX choosing not to backfill an AP position after autonomous agents absorbed the workload. While human oversight remains required for final sign-offs, automating foundational workflows threatens the traditional finance talent pyramid. Historically, junior staff cultivated the institutional intuition needed to become controllers and CFOs through routine data entry and reconciliations, leaving organizations with an emerging succession dilemma regarding how to train future executives. The panel compares this shift to software engineering, where junior coding positions have sharply contracted while senior architects transition toward directing agent fleets. As technical barriers fall through no-code interfaces, finance roles will increasingly prioritize professionals who can bridge accounting logic with agent orchestration, workflow design, and governance. Preserving the Immutable Core Ledger (39:03) Addressing system architecture, Sharma and Naveh draw a strict boundary between rigid "systems of record" and adaptable "systems of work." Core ERP platforms like SAP, Oracle, and NetSuite possess decades of proven controls, regulatory compliance, and predictable failure modes that external auditors trust. Consequently, experimental AI point solutions and cross-application agents must be confined to the operational edge, where they can orchestrate tasks across fragmented tools without jeopardizing subledgers or cash movements. Melton cautions that venture-backed, AI-native ERP startups suffer from notable functional immaturity, frequently launching as basic general ledgers devoid of native AP, AR, or inventory modules. Adopting these tools requires assembling composable, best-of-breed ecosystems alongside tools like Ramp, Tabs, or Numeric, forcing finance leaders to weigh integration complexity against the deep edge-case capabilities of incumbent enterprise software. Pragmatic Data Onboarding and Agent Governance (12:35) Naveh challenges traditional enterprise IT assumptions, arguing that spending millions of dollars cleaning years of legacy data is an inefficient use of resources. Because Large Language Models operate more like human employees than rigid statistical models, teams can onboard agents using operational overviews and sample data, allowing the models to learn dynamically and request clarification when encountering ambiguities. However, deploying fleets of autonomous agents introduces significant organizational governance gaps. Just as dedicated revenue operations teams emerged alongside platforms like Salesforce, enterprises must establish formal orchestration functions—typically under innovation or transformation leadership—to evaluate speed-versus-accuracy trade-offs, measure agent ROI, and manage digital workforce training. While artificial intelligence offers finance teams a transition from historical record-keeping to continuous strategic planning, CFOs must navigate architectural fragmentation and an unresolved talent development pipeline. Long-term adoption requires diagnosing operational process bottlenecks and establishing dedicated executive ownership rather than deploying generic software tools across unstandardized workflows.

在Ai4研讨会上，企业高管与咨询顾问指出，人工智能正推动CFO办公室从以削减成本为导向的离散自动化，转向持续预测和战略决策；但企业在推进规模化落地时，仍面临基础数据认知、组织人才梯队断层以及核心账本与外围工具选型的多重制约。 战略重心从削减人力转向持续预测 (5:08) 财务团队部署AI的主要投资回报不再仅是单纯裁员，而是缩短报表延迟并提升预测频率。EnergyX首席财务官Mayank Sharma指出，传统财务受限于人工核算的高昂边际成本，通常只能提供月度或季度等事后静态报告；而引入AI在发票粒度上自动汇总数据并生成注释后，财务能够维持动态更新的持续预测，从而介入业务早期的战略决策。Reindeer联合创始人兼首席执行官Yoav Naveh补充，不同于拥有庞大集中预算且容易通过外包套利的客户支持部门，企业财务团队规模通常只有5至20人，其真正价值在于消除季度结账长达两到三个月的滞后，提供近乎实时的业务信号。RSM合伙人Jonas Melton亦指出，AI智能体可以在月末进行单笔交易级别的差异与波动分析，捕捉人工抽样容易忽略的细微异常。 数据准备与智能体编排的认知重构 (12:35) 围绕企业推进AI的先决条件，业内正对传统IT数据清洗策略产生分歧。Jonas Melton将企业AI应用划分为个人生产力工具、组织职能智能体以及基于数据湖的企业级智能系统三个层级。针对数据准备，Yoav Naveh提出异议，认为花费数百万美元清洗多年历史数据的传统做法在很大程度上是浪费资金与时间。他主张大语言模型并不等同于严守“垃圾进、垃圾出”的传统机器学习，其运作更像人类员工，企业只需提供清晰的流程概览与最新样本让其在工作中学习，而非强记陈旧历史。此外，随着组织从单个验证性概念走向部署上百个自主智能体，企业需要建立类似人力资源部门的常态化编排职能，负责数字员工的培训、绩效评估和流程调度。 基层岗位萎缩与财务人才培养困局 (22:15) AI对流程性事务的接管正在重塑财务部门的组织架构。Mayank Sharma以其实践为例指出，EnergyX在应付账款人员离职后未再补录该岗位，日常AP处理已基本由AI智能体承担，仅保留人工在环进行最终审批。这种变化对传统财务金字塔构成了结构性挑战：初级员工历来通过发票处理、对账和基础录入积累业务直觉，并逐步成长为财务总监或CFO。若底层岗位被彻底替代，企业将面临下一代高管缺乏实战经验的断层危机。这一轨迹类似于软件工程领域的现状，高级架构师开始转向指挥代码智能体，而初级岗位大幅减少，迫使资深专业人员也必须重新定义自身的核心价值。 自主开发与现成采购的边界权衡 (31:27) 随着无代码与AI交互工具的成熟，财务人员与软件工程师之间的技术鸿沟正在缩小，但自研与外购之间存在明显边界。有观点支持“让钻井工人成为宇航员”的逻辑，即与其让软件工程师学习复杂的会计规则，不如让懂业务的财务人员利用自然语言直接构建定制化仪表板与数据模型。然而，针对该趋势的反思强调，企业财务约80%的流程是高度标准化的，“预付款就是预付款，应计就是应计”，通用会计场景应优先采用成熟的商业软件。Jonas Melton指出，许多企业错误地将基础流程问题归咎为AI需求，例如有客户试图用Claude在Excel标签页间复制粘贴试算平衡表，这本质上属于流程标准化与基础RPA的范畴，盲目套用生成式AI实属误用。 核心账本的稳定性与外围工具的解耦 (39:03) 在面对原生AI初创平台与传统巨头的选型博弈时，嘉宾一致主张将“核心系统”与“工作系统”进行分层隔离。Mayank Sharma强调，凡是涉及子账本、资金转移及法定审计的交易记录，必须具备极高的稳定性与成熟控制，因此应继续依赖SAP、NetSuite等具有数十年验证历史的成熟平台，而将新兴AI工具限制在可随时开关的外围实验层。Yoav Naveh分析称，大型企业并无理由冒巨大风险推倒成熟ERP，但面临系统升级的中端市场企业构成了AI原生ERP的主要试验场。Jonas Melton提醒，当前许多风险投资支持的AI原生ERP在功能上仍不成熟，往往只是现代化总账而缺乏应收、应付及库存模块，迫使企业必须采购Ramp、Tabs等外围工具自行拼装，企业在采纳前须权衡组装复杂度与传统平台的深厚功能积累。 总体而言，财务领导者推进AI转型的核心在于紧扣业务瓶颈而非盲目跟风，同时必须设立明确的执行责任人，并将自动化节省下来的工时主动引导至企业前瞻性的战略推演中。

## Background 本内容由频道 Ai4 于 2026 年 8 月 22 日发布在 YouTube 平台上。视频记录了一场关于人工智能对首席财务官（CFO）办公室影响的圆桌讨论，探讨了企业财务领袖如何部署 AI 实现核心职能现代化、生成战略前瞻洞察，并在严谨监管下管理透明度、内部控制与合规性。 ## TL;DR 圆桌嘉宾围绕 AI 在企业财务中的实际落地路径展开深入探讨，指出 AI 的真正价值不仅在于削减基层重复工作成本，更在于通过实时数据流实现持续滚动预测与穿透式差异分析。专家们分析了团队从抗拒到高度依赖的心态转变，并针对数据清洗的必要性、AI 原生架构与传统 ERP 的权衡，以及初级岗位收缩导致的人才培养断层提出了深刻见解。最后，与会者建议 CFO 们采取结果导向策略，设立专属负责人，并将自动化释放出的人力与时间积极转化为高阶商业洞察。 ## Atmosphere & style 整场讨论呈现出务实、前瞻且充满行业思辨的专业基调。主持人与来自一线企业（EnergyX）、AI 初创公司（Reindeer）及专业咨询机构（RSM）的嘉宾形成了多维度的思维碰撞，既有对“数据清洗是否浪费时间”等前沿观点的锋芒交锋，也有对技术冲击下初级人才梯队断层危机的冷峻反思。现场对话节奏紧凑明快、案例鲜活，生动展现了当下财务管理者在拥抱 AI 浪潮与坚守内控合规之间的真实权衡与探索。 ## Key takeaways - [02:00](https://www.youtube.com/watch?v=xntbxng9yRw&t=120s) 财务团队对 AI 的态度在数月内经历了从抵触、被动接受到极度信任的演变，企业探索也正从单个应付账款（AP）Agent 的单点实验迈向百级智能体与人类共存的规模化协同。 - [05:08](https://www.youtube.com/watch?v=xntbxng9yRw&t=308s) AI 大幅降低了财务数据生成的边际成本与精力负担，推动财务职能从低频、静态的“事后报告”升级为穿透至单张发票颗粒度的持续滚动预测。 - [08:26](https://www.youtube.com/watch?v=xntbxng9yRw&t=506s) 财务领域部署 AI 的核心 ROI 并非单纯裁减少数编制，而是彻底消除传统手工流程导致的数据滞后，基于近乎实时的真实数据流提供高阶差异分析与业务洞察。 - [11:29](https://www.youtube.com/watch?v=xntbxng9yRw&t=689s) 企业 AI 应用呈现三层演进架构：从基础的个人生产力助手（Copilot/ChatGPT），递进到跨团队共享的组织级智能体，最终上升为深度挖掘底层大数据的企业级智能决策。 - [16:05](https://www.youtube.com/watch?v=xntbxng9yRw&t=965s) 大语言模型（LLM）的运作模式不同于传统机器学习，耗费巨资清洗过时的历史数据往往是不必要的，更高效的做法是将其视为“超级新员工”，基于业务框架在新鲜动态数据中实战学习。 - [17:57](https://www.youtube.com/watch?v=xntbxng9yRw&t=1077s) 随着企业内 AI Agent 规模激增，未来将崛起类似“收入运营”（RevOps）的新型专职 Agent 运营团队，统筹智能体的入职培训、计算成本与精度的动态平衡以及 ROI 评估。 - [19:44](https://www.youtube.com/watch?v=xntbxng9yRw&t=1184s) 财务 AI 落地面临人员思维定势与系统割裂的双重阻碍：财务人员习惯于程序化合规与历史核算，而核心账务、报销及文档系统在早期设计上缺乏统一的 AI 读取接口。 - [22:50](https://www.youtube.com/watch?v=xntbxng9yRw&t=1370s) 基层岗位（如 AP/AR 专员）被 AI 接管使传统“金字塔”组织架构瓦解，带来了初级训练场缺失以及未来资深财务管理人才断层的长远风险。 - [25:48](https://www.youtube.com/watch?v=xntbxng9yRw&t=1548s) 软件工程的演进预示了财务的未来路径：基础编码岗位消退，人才重心向“智能体编排”与“闭环反馈评估”迁移，高阶专家也需经历重构工作范式的转型阵痛。 - [31:32](https://www.youtube.com/watch?v=xntbxng9yRw&t=1892s) AI 颠覆了传统软件让企业削足适履的逻辑；正如电影《绝世天劫》的隐喻，赋予掌握业务隐性知识的财务专家 AI 构建能力，远比指望外部工程师理解复杂业务更有效。 - [35:14](https://www.youtube.com/watch?v=xntbxng9yRw&t=2114s) CFO 办公室有 80% 的工作属于通用标准化流程，企业应优先采购开箱即用工具并理顺流程体系，警惕将高阶大模型盲目滥用于本应由 RPA 或基础自动化解决的场景。 - [38:11](https://www.youtube.com/watch?v=xntbxng9yRw&t=2291s) 企业在 ERP 架构上应采取“稳固核心、开放边缘”策略：总账等核心记录系统（System of Record）依赖成熟传统厂商以确保审计合规，边缘工作系统（System of Work）则大胆接入灵活的 AI 原生产品。 - [45:42](https://www.youtube.com/watch?v=xntbxng9yRw&t=2742s) CFO 制定 AI 战略时必须坚持结果导向而非盲目跟风用例，设立全职负责人推进战略迭代，并提前规划将 AI 节省的时间系统性转化为高质量的商业战略洞察。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=xntbxng9yRw&t=4s) 圆桌讨论开场与嘉宾介绍 主持人（Thomson Reuters Ventures 合伙人 Joe Dormani）拉开了本场圆桌讨论的序幕，探讨人工智能（AI）与自动化在首席财务官（CFO）办公室中的应用。本次对话将从多个视角展开：财务领袖如何看待 CFO 职能中的 AI 机遇，以及最前沿的团队如何思考落地部署 AI 并培训团队使用尖端工具。随后，各位嘉宾依次进行了自我介绍。 首先发言的是 EnergyX 的首席财务官 Mayank Sharma。他介绍 EnergyX 是一家锂提取与精炼技术公司，专注于为电动汽车和锂离子电池提供关键原材料，推动清洁能源革命。接着，Reindeer 联合创始人兼首席执行官 Yoav Naveh 进行了自我介绍，他创立的公司致力于帮助企业运行 AI Agent（智能体），以自动化应付账款（AP）、应收账款（AR）、税务申报以及供应链运营等核心业务流程；他观察到行业内涌现了大量尝试，并期待探讨这些技术将在未来几年如何重塑 CFO 办公室。最后，RSM US LLP 的合伙人 Jonas Melton 作了介绍，作为全球规模最大的会计师事务所之一，他在 RSM 的核心职责正是围绕自动化与 AI，致力于将这些技术解决方案交付给广大客户。 ### [02:00](https://www.youtube.com/watch?v=xntbxng9yRw&t=120s) 过去12个月AI落地带来的最大震撼与态度转变 主持人抛出了第一个议题：鉴于当前 AI 市场的惊人变革速度，回顾过去12个月，从客户、技术供应商或内部团队的角度来看，CFO 职能中最令人惊讶的转变是什么？ Mayank Sharma 结合自身推动团队使用 AI 的亲身经历，分享了财务团队心理态度的剧烈演变。在短短数月内，团队经历了完整的周期： 1. **漠不关心与抗拒**：最初对 AI 持冷漠态度，甚至在工作流引入时产生抵触； 2. **勉强接受**：逐步在日常流程中尝试并被动接受 AI 的辅助； 3. **全面拥抱与极度信任**：最终将 AI 视为团队不可或缺的“新同事”。Mayank 特别提到，团队中甚至出现了“部分员工对 AI 的信任度超过了对人类同事的信任”的戏剧性转变，最初的质疑者纷纷转变成了 AI 的狂热支持者。 Yoav Naveh 则指出了企业部署模式的规模化跃迁。大约一年前，CFO 和财务团队通常只是做点状实验，比如“假设部署一个 Agent 帮我处理应付账款”；而如今，CFO 们开始思考：“当企业内运行着100个 Agent 解决100项不同任务，并与人类员工并肩协作时，组织形态会变成怎样？”这种从单点工具到大规模人机协作的转变带来了全新的管理挑战，引发了关于“在组织内究竟由谁来统筹和负责管理这些 Agent”的深入讨论。 Jonas Melton 补充强调了 AI 渗透的惊人速度。尽管像 Siri 这样的 AI 技术已存在多年，但自 ChatGPT 面世以来的18个月里，技术呈现爆发式增长。董事会和首席执行官（CEO）对各部门极速采纳 AI 的期望值被大幅拉高。虽然伴随着巨大挑战，但企业已经开始切实体会到 AI 带来的实际成果。 ```mermaid flowchart TD accTitle: 财务团队对AI的采纳演进路径 subgraph MindsetEvolution["团队心态演化"] A["漠然置之与抵触"] --> B["勉强接受并融入工作流"] B --> C["全面采纳并视为新同事"] end subgraph ScaleEvolution["管理规模演进"] D["单点尝试<br/>(例如：单个应付账款 Agent)"] --> E["规模化协同<br/>(100+ Agents 并行与人机共存)"] end ``` ### [05:08](https://www.youtube.com/watch?v=xntbxng9yRw&t=308s) 财务职能跃升：从降低制表成本到实现持续滚动预测 主持人引导大家探讨投资回报率（ROI）的核心命题。他指出，业界讨论 AI 时往往聚焦于“降本增效”，却忽视了 AI 带来的真正新机遇——让财务团队在组织价值链中“向上跃升”（moving up the stack），成为真正的战略顾问。 Mayank Sharma 深度剖析了这一转变。他指出，AI 带来的最大变革是**大幅降低了生成财务数据的成本和精力负担**： - **颠覆报告节奏**：过去由于人工生成报表耗时费力，财务团队往往只能维持季度或月度更新的节奏；AI 接管繁重的数字处理与说明文字生成后，EnergyX 实现了**持续更新的滚动预测**（continuously updating forecasts）。 - **穿透至单张发票的颗粒度**：借助 AI 强大的精细化分析能力，财务团队能按月与各个业务部门对齐，不仅能逐笔穿透到具体单据，还能让 AI 自动生成深度的业务分析说明。 - **重塑财务部门的核心价值**：省下来的时间不再用于做“事后报告”（post-facto reporting），而是让财务团队在决策初期就深度介入，主动提问、挖掘数据洞察，为公司战略提供前瞻性驱动力。 ### [08:26](https://www.youtube.com/watch?v=xntbxng9yRw&t=506s) 超越短期效率：实时数据流与高阶业务智能 Yoav Naveh 对 Mayank 的观点表示高度认同。他指出，虽然 AI 在客服领域能通过替代人工直接削减数百万美元成本，但财务部门的工作极为分散：可能是10个人负责这项、20个人负责那项，或是少数人兼顾多重职责。因此，如果单纯寄希望于裁减几个 HC（Headcount）来衡量短期效益，财务部门的 AI ROI 并不明显。 Yoav 强调，AI 真正的战略价值在于**彻底消除了手工流程造成的数据滞后**。以往很多企业的季度报告往往在季度结束两到三个月后才能出炉；而当60%至70%的流程由 Agent 串联打通后，财务数据能够即时生成，使整套预测模型建立在近乎实时的真实数据之上，为企业战略决策提供决定性优势。 Jonas Melton 则从“高阶商业智能”的角度进行了补充： - **穿透式差异分析（Flux / Variance Analysis）**：在月末结账进行变动分析时，AI Agent 或算法模型能够深入交易底层，捕捉到人工排查极易遗漏的异常与深层洞察。 - **AI 部署的双轨制与个性化落地**：Yoav 补充指出，财务领域的 AI 应当分为两条轨道：一条是负责处理高重复性工作、提高速度与准确性的**任务型 Agent**；另一条则是**智能助手（Copilot）**。他分享了一个行业见闻：许多 CFO 在被问及 AI 进展时往往只回答“我们给全员买了 Copilot”，但当被问到具体怎么用时却茫然不知。Yoav 强调，必须将 AI 助手与企业系统及核心数据打通，并针对具体角色深度定制——财务总监（Controller）需要的智能助手与金融犯罪调查部门（Financial Crimes）所使用的助手截然不同。 ### [11:29](https://www.youtube.com/watch?v=xntbxng9yRw&t=689s) 企业 AI 应用的三层架构模型 关于企业如何落地人工智能，业内通常将其划分为复杂度递增的三个主要层级。第一层是“个人生产力”（Personal Productivity），也就是目前最为普及的场景——员工拥有独立的 Copilot 或 ChatGPT 账号，用来起草日常邮件、搜索整理个人收件箱或处理基础文本。第二层是“企业级/组织级智能体”（Organizational/Enterprise-wide Agent），表现为部门内部或跨团队共享的 Chatbot 与特定业务技能，能够协助多个团队成员协同处理标准化流程。 最复杂但也最具变革意义的是第三层——“企业级智能”（Enterprise Intelligence）。在这一层级中，企业会在底层庞大的数据库或数据湖之上运行特定的智能体与前沿大模型，深度挖掘业务底层规律，直接为管理层提取具备前瞻性的商业洞察。随着应用层级的攀升，企业面临的技术复杂度和组织协调要求也呈指数级增长。 ```mermaid flowchart TD accTitle: 企业AI应用的三层递进模型 subgraph L3["第三层：企业级智能 (Enterprise Intelligence)"] D3["跨系统大数据分析 / 商业战略洞察"] end subgraph L2["第二层：组织级智能体 (Enterprise-wide Agents)"] D2["团队共享技能 / 业务流程Chatbot"] end subgraph L1["第一层：个人生产力 (Personal Productivity)"] D1["ChatGPT / Copilot (邮件草拟、个人搜索)"] end L1 --> L2 L2 --> L3 ``` --- ### [12:35](https://www.youtube.com/watch?v=xntbxng9yRw&t=755s) AI 就绪度之争：数据整理与组织变革 在探讨企业如何迈向更高阶的 AI 应用时，“AI 就绪度”（AI Readiness）成为了核心议题。企业在推进智能化时普遍面临两大主要障碍：其一是“混乱的数据”（Messy Data），即数据分散在各个孤立系统中、缺乏统一规范，难以清洗合并以满足分析和报告需求；其二是“变革管理”（Change Management），即如何引导员工改变既有的工作习惯、重构数字处理流程并主动拥抱新技术。 从专业咨询与客户服务实践来看，RSM 提出了名为“RSM Foundations”的评估框架，强调就绪度必须“基于具体用例”（Use-case Specific）来考量： - **轻量级用例的数据准备**：如果企业只是为了内部系统上线一个问答机器人，只需将产品文档和用户手册导入知识库即可。例如为一个新实施的系统构建智能助手，员工只需向 Chatbot 提问“如何添加新用户”或“如何更改填报人与审核人”，AI 即可精准检索文档并作答，这类项目所需的数据清洗成本极低。 - **高阶用例的数据准备**：若要实现企业级智能洞察，则必须打通包含多个业务系统的数据仓库或数据湖，涉及繁重的数据治理工作。 - **组织与技术层面的评估**：除了数据，变革管理的核心在于人才结构的梯队建设，确保团队能够跟上节奏；同时在技术选型上，企业需要审视自身是应当充分挖掘现有软件生态的内置 AI 能力，还是选择定制化自建（Build）或对外采购（Buy）。 --- ### [16:05](https://www.youtube.com/watch?v=xntbxng9yRw&t=965s) 颠覆传统观念：为什么数据清洗可能在浪费时间 面对“数据清洗到底该由数据工程师负责还是由财务部门主导”的提问，Yoav Naveh 提出了极具颠覆性的观点：“在我看来，在数据就绪度（Data Readiness）上投入任何资金都是在浪费时间。”他指出，企业如果回头去看过去两三年的历史数据，会发现其中很大一部分早已随着业务模式的快速迭代而失效；耗费数百万美元雇佣数据工程师去清理陈旧数据，往往是出于对传统机器学习模型的认知惯性。 传统 AI 遵循严格的“垃圾进，垃圾出”（Garbage In, Garbage Out）定律，高度依赖大规模、清洗完毕的样本库进行训练；但现代大语言模型（LLM）的运作逻辑截然不同。 > “训练大模型运行工作流的最佳方式，就像你在组织中培训一名新员工一样。新员工入职时，绝不会去死记硬背过去三年的历史数据，也不会去背诵 200 页的操作手册。他们只需要了解业务流程的整体运作框架，看几个具体案例，然后在实际工作中边干边学。” 如果把 LLM 视为一位具备 100 倍常人产能的“超级新员工”，培养它的合理方式也是提供流程框架与即时反馈：遇到明确任务直接执行，遇到模糊问题主动提问。企业每天都在生成最新的动态数据，让大模型直接基于新鲜数据在实战中学习，不仅落地速度更快，还能彻底免去跨部门协调数据工程团队、层层审批的冗长与挫败感。 --- ### [17:57](https://www.youtube.com/watch?v=xntbxng9yRw&t=1077s) AI 智能体的组织架构：未来的“Agent 运营团队” 当企业内部的 AI 智能体数量从单个扩展到成百上千个时，谁来负责智能体的日常管理与训练？目前绝大多数企业尚未设立此类专职岗位。然而，就像组织需要人力资源团队负责员工的招聘、培训、绩效考核与研讨会一样，管理规模化智能体的专门职能部门必将应运而生。 - **新型职能部门的崛起**：该角色类似于当年 Salesforce 普及后迅速崛起的“收入运营”（RevOps）部门。未来企业中将会出现专门的智能体运营团队，有时隶属于首席信息官（CIO），有时归属于创新或数字化转型负责人（Head of Innovation/Transformation）。 - **智能体管理的核心职责**： - 衡量智能体团队的投资回报率（ROI）； - 动态权衡计算成本、响应速度与输出精度之间的平衡（Trade-off between cost, speed, and accuracy）； - 负责智能体的入职培训（Onboarding）、教学微调以及全员使用培训。 - **战略紧迫性**：AI 已成为不可逆转的最高战略之一，企业无论是在财务团队内部设立专人，还是在全公司层面设立统筹部门，“如果今天还不指定专人全身心投入解决这一管理难题，就无法最大化转型的成功概率”。 --- ### [19:44](https://www.youtube.com/watch?v=xntbxng9yRw&t=1184s) 财务人员的心态转变与底层系统的割裂挑战 从实体企业一线运营的角度来看，“AI 就绪度”的内涵远比单纯讨论数据干净与否更为广泛，它直接受制于“人员思维”与“系统架构”两座大山。 首先是**财务人员的心态转变**（Mindset Shift）。财务总监、会计师与风控人员长期以来受到的职业训练，核心在于遵循公认会计准则（GAAP）、严格执行合规协议以及维持严谨的月度/季度结账周期。这种高度程序化、注重合规与历史准确性的思维惯性，往往不擅长以灵活、探索性的方式去主动查询数据或捕捉异常指标背后的前瞻性趋势。 ```mermaid flowchart LR accTitle: 财务AI落地中的系统孤岛与连接挑战 subgraph Systems["分散的底层系统"] S1["核心核算<br/>(QuickBooks / Sage)"] S2["支出与报销<br/>(Ramp)"] S3["文档与制度<br/>(SharePoint)"] S4["项目管理软件"] end subgraph Bridges["集成与交互层"] B1["官方/付费连接器<br/>(如 DataRails / Claude)"] B2["定制化手动数据桥梁<br/>(Manual Bridges)"] end subgraph AI["AI 智能体应用"] Agent["财务前瞻洞察 / 异常检测 Agent"] end Systems --> Bridges Bridges --> AI ``` 其次是**系统生态的严重碎片化**。任何一家企业的 IT 架构都是在不同时期由不同团队选型的产物： - 记账核算可能采用 QuickBooks 或 Sage； - 账单支付、采购订单（PO）生成和员工报销可能运行在 Ramp 之上； - 内部制度和财务文件存储在 SharePoint； - 业务进度则散落在各类项目管理软件中。 这些系统在选型之初从未设想过未来有一天需要让 AI 智能体无缝读取其全量数据。当前无论是接入支持 Claude 的专用连接器（通常需要额外的采买成本与集成周期），还是依靠人工搭建数据提取桥梁，打通系统底座都需要耗费大量时间。企业必须先解决人员思维定势与系统割裂的复杂现实，才能进一步探讨数据本身的质量。这种对“效率提升”与“专家训练 AI”的讨论，也在财务团队内部悄然引发了关于职业安全感与角色转变的深层焦虑。 ### [22:50](https://www.youtube.com/watch?v=xntbxng9yRw&t=1370s) 金融团队的“倒金字塔”危机与初级岗位的消失 随着 AI 逐渐接管日常工作，企业财务部门的组织架构与人员编制正在经历实质性的冲击。主持人 Joe Dormani 抛出疑问：AI 带来的效率提升究竟会转化为纯粹的裁员，还是会重塑未来的财务团队？EnergyX 首席财务官 Mayank Sharma 分享了切身经历：当团队的一名应付账款（AP）专员离职时，公司决定不再填补该空缺，因为 AI Agent 完全能够胜任绝大多数与 AP 相关的日常流程。尽管人类员工仍需进行最终的复核与审批以确保合规，但这一案例证明，AI 对基层编制确实带来了向下的收缩压力。 更深层的影响在于人才梯队的断层。Sharma 指出，传统财务团队呈现清晰的“金字塔结构”：底层由大量初级员工构成，负责对账、付款、发票录入等琐碎事务，向上为资深高管提供决策数据；资深人员则在此基础上注入战略思考。如今底层的初级岗位正被 AI 大规模取代。这带来了一个棘手的长期问题——未来的财务主管从何而来？“当今精明的财务总监（Controller）和会计师并非一日炼成，他们历经多年基层磨砺才积累出敏锐的商业直觉与专业经验。”如果彻底抽空基层的训练场，未来的高管将失去成长的摇篮；若是委婉地将此称为“工作增强（Job Augmentation）”，但在现实中，这确实构成了残酷的“岗位替代（Job Replacement）”。 - **应付账款（AP）不补缺口**：EnergyX 在 AP 员工离职后直接由 AI Agent 接管大部分流程，仅保留人类把关和签署（Human-in-the-loop）。 - **金字塔结构的瓦解**：底层繁琐的对账与发票处理自动化，使传统初级入行通道大幅收紧。 - **人才培养断层风险**：初级训练场的缺失，导致未来具备深厚经验的资深财务领袖面临培养真空。 ```mermaid flowchart TD accTitle: 传统财务金字塔与 AI 驱动组织对比 subgraph Traditional["传统财务团队架构（金字塔型）"] T_Senior["资深高管与财务总监<br>战略决策与商业洞察"] T_Mid["中层管理人员<br>流程审查与分析报告"] T_Junior["庞大的初级员工层<br>发票录入、对账、AP/AR 处理"] T_Junior --> T_Mid --> T_Senior end subgraph AI_Driven["AI 时代新型架构（纺锤/沙漏型）"] A_Senior["高阶编排者与商业战略家<br>监督 Agent 输出、闭环治理"] A_Bridge["跨界复合型角色<br>理解业务流程并编排智能体"] A_Agent["AI Agent 智能体集群<br>全天候处理高重复度流程"] A_Agent --> A_Bridge --> A_Senior end ``` --- ### [25:48](https://www.youtube.com/watch?v=xntbxng9yRw&t=1548s) 软件研发的先例：工程师的“身份危机”与智能体编排 当被问及初级人员该何去何从时，Yoav Naveh 认为过早下定论无异于猜测，但他建议观察 AI 渗透最快的前沿阵地——软件工程部门。工程领域发生的变化令人震撼，其真正的转折点发生在去年 12 月。今天优秀的资深工程师甚至“不再直接打开或查看底层代码”，而是调度和管理一组编程助手与智能体（Coding Assistants & Agents）来完成代码编写。在这一模式下，初级工程师的传统职能已几乎彻底消失且难以逆转，取而代之的是两类新能力：一是具备编排智能体能力的人，二是能够搭建闭环反馈机制、评估 Agent 决策优劣的基础设施构建者。 然而，这种技术跃迁并未让资深人员彻底高枕无忧。Naveh 直言，AI 正持续向价值链上游攀升，有能力承接越来越多高阶决策，甚至威胁到资深工程师的生存空间。在转折点之前，人类是按代码块精确指挥 AI 并逐行审查；而随着大模型能力的跃升，AI 已能自主接管整套系统实现。许多从业二十年的资深工程师遭遇了剧烈的“身份危机”（Identity crisis）：“我花了整整二十年时间成为业内顶尖高手，如今脚下的地毯却被瞬间抽走，不得不彻底颠覆自己的工作方式。”这种技术阵痛虽率先在拥抱新技术更敏捷的技术圈爆发，但很快便会波及财务与运营领域。 - **转折时间节点**：去年 12 月大模型与编程智能体能力的质跃，重构了软件工程的工作范式。 - **初级工程师职能的消退**：单纯写基础代码的初级岗位难以为继，岗位需求向“Agent 编排”与“闭环评估基础设施建设”迁移。 - **资深专家的职业阵痛**：二十年深厚积累的专业技能在自动化面前贬值，倒逼高阶人才重构自身定位。 --- ### [28:31](https://www.youtube.com/watch?v=xntbxng9yRw&t=1711s) 财务职能拓展与“技术-业务”鸿沟的弥合 针对上述技术冲击，Joe 提出反思：财务职能天然涵盖广泛的商业版图，若能从繁重的报表制作、建模和预测中抽身，团队是否能深入业务一线，对 GTM（走向市场）策略或客户获客效率展开更具商业价值的深度分析？Jonas Melton 赞同初级岗位不会瞬间归零，毕竟随着老一代管理者退休，企业必须找到为中高层输送血液的升级路径。他强调，裁员主要集中在 AP、AR 处理等具备极高“可扩展性（Scalability）”与重复性的业务模块，甚至部分基础 FP&A 也正在被压缩；而在月结关账（Financial Close）等复杂环节，AI 虽能执行特定任务，但仍高度依赖人工复核与治理机制。 Melton 进一步指出，当前市场上最显著的痛点在于“技术与财务之间的沟通断层”——精通算法与工具的工程师往往不懂财务内控与核算逻辑，而深谙业务流的财务人员又缺乏工程构建能力。所幸这一鸿沟正在迅速收窄，低代码/无代码工具与大模型交互界面的进化，使财务人员如今能够亲手搭建三年前根本无法想象的自动化流程。未来的企业绝非由智能体独立运转的冰冷机器，人类依然是核心支柱，但其角色将重塑为 Agent 的管理者、维护者与业务编排师。 - **裁员分水岭**：标准化、高扩展性的流水线流程（AP/AR、基础数据录入）首当其冲；非标准化、强治理属性的流程（关账、内控）仍需“人在回路”。 - **市场核心瓶颈**：懂构建 AI 的技术工程师与懂财务流程的业务专家之间存在认知鸿沟。 - **工具平民化趋势**：随着自然语言与智能体门槛降低，业务人员自主搭建系统的能力空前提高。 --- ### [31:32](https://www.youtube.com/watch?v=xntbxng9yRw&t=1892s) 自研还是外购？“绝世天劫”隐喻下的工具构建权 随着软件与定制工作流的开发门槛被彻底打破，Joe 将焦点引向“自研还是外购”（Build vs. Buy）的抉择难题。目前像汤森路透（Thomson Reuters）这样的成熟机构内部，各类业务与财务团队都在跃跃欲试地自建定制应用与工作流，但这种遍地开花的实验是否属于资金的最优配置？对此，Yoav Naveh 提出了极具颠覆性的视角：传统企业软件的模式是工程师试图理解你的业务后交付一套僵化系统，强迫企业调整自身流程去迎合软件；而生成式 AI 时代的核心变革在于“让软件适应既有业务流程”。 ```mermaid flowchart LR accTitle: 电影《绝世天劫》隐喻对比软件开发逻辑 subgraph Analogy["电影《绝世天劫》的抉择"] A1["宇航员学打洞 ❌<br>(技术人员强行理解财务)"] A2["钻井工学当宇航员 ✔️<br>(赋予财务专家 AI 工具)"] end subgraph Reality["AI 时代自研软件路线"] R1["传统外包/IT主导<br>工程师理解偏差导致交付脱节"] R2["财务专家亲手搭建<br>结合深厚隐性知识与专属逻辑"] end A1 -.-> R1 A2 -.-> R2 ``` 为了具象化这一观点，Yoav Naveh 引用了经典灾难电影《绝世天劫》（*Armageddon*）的故事情节：当小行星即将撞击地球时，决策层面临两难——究竟是花时间教宇航员学会深度钻井去安放核弹，还是把一群脾气火爆但经验老到的钻井工人训练成宇航员？电影的答案是后者，这与当下财务软件的逻辑如出一辙。与其指望外部工程师去理解复杂幽微的财务逻辑与企业专有的制度知识（这极易导致系统水土不服），不如直接把强大的 AI 逻辑构建工具交到财务人员手中。“最懂数据与业务细节的人就在财务部门内部，将构建权交给业务专家，才能避免技术团队交付出脱离实际的错误方案。” - **软件交互范式逆转**：从“企业削足适履适应商业软件”转向“软件弹性适应企业独有流程”。 - **《绝世天劫》隐喻**：让懂业务的“钻井工人”（财务骨干）掌握“航天技术”（AI 编程工具），优于让外行工程师凭空猜测业务规则。 - **业务沉淀与敏捷赋能**：将系统逻辑的编排权归还给掌控隐性知识（Institutional Knowledge）的业务一线，能有效减少试错成本与沟通损耗。 ### [033:45] 赋能初级财务人员：从依赖工程师到氛围编程（Vibe Coding） Yoav Naveh 阐述了 Reindeer 的核心理念：虽然目前仍需构建连接不同系统的集成通道，但长远目标（未来一至三年内）是彻底赋能业务人员，让他们能够借助自身的业务沉淀与业务知识独立管理并配置 AI Agent，而非每次需求变更都依赖软件工程师。 ```mermaid flowchart LR accTitle: 财务赋能与工具分工演进 subgraph Core["核心记录系统 System of Record"] direction TB ERP["传统 ERP (SAP / NetSuite)"] Ledger["总账与明细账 (严格合规/低容错)"] end subgraph Edge["边缘工作系统 System of Work"] direction TB Agent["AI Agent / 跨系统编排"] Dash["自动化报表与分析看板"] VC["业务人员 Vibe Coding"] end Core -->|稳定数据底座| Edge ``` Mayank Sharma 对此深表赞同。他指出，财务部门的本职工作从来不是开发复杂的软件，而是数据运算、趋势分析以及构建可视化数据看板。在实际团队管理中，Mayank 观察到许多具备技术思维的初级财务人员已经开始积极拥抱“氛围编程”（Vibe Coding），利用 AI 极为简便的生成能力自主搭建看板、分析模型与报告洞察。这种自主性大幅减少了引入外部工程师维护系统所带来的沟通摩擦。Mayank 强调，团队绝不会轻易放弃初级财务专业人员，AI 工具反而正在放大他们的价值与独立性。 --- ### [035:14] 财务职能的 80% 规律与“先梳理流程、后叠加 AI” Jonas Melton 从“用户画像”（Persona）的视角对财务任务进行了拆解。他认为，CFO 办公室处理的事务有 80% 属于高度重复的标准化工作——无论处于哪个行业，预付款（Prepaid）和应计费用（Accrual）的会计逻辑都是相通的。因此，去自研定制化工具毫无意义，直接采购开箱即用（Out-of-the-box）的成熟产品才是最优解；相比之下，COO 办公室或产品研发团队代表了企业的独特业务与核心 IP，更适合定制化开发。 针对 Yoav 提到“一家大型 CPG（消费品）客户内部 5 个部门处理应付账款（AP）流程各不相同”的现状，Jonas 认为应当优先从**流程与数据架构**着手进行标准化，而非盲目引入 AI： - **警惕滥用 AI**：Jonas 分享了一个真实案例，曾有客户提出想要调用 Claude 大模型去复制粘贴 Excel 资产负债表中的试算平衡表（Trial Balance）数据，以联动各个账户的工作表。Jonas 当场指出“这根本不是 AI 的使用场景”，这类需求只需要简单的 RPA（机器人流程自动化）或现成的开箱即用对账工具即可搞定。 - **技术落地层次**：合理的实施路径必须是“先建立正确流程 $\rightarrow$ 再打通数据集成 $\rightarrow$ 然后应用 RPA 等基础自动化 $\rightarrow$ 最终在顶层叠加 AI 赋能高阶智能”。 --- ### [038:11] 核心与边缘：AI 原生 ERP 初创与传统巨头（SAP / NetSuite）的选择权衡 主持人 Joe Dormani 将话题引向厂商选型（Vendor Selection），提出当前财务领导者普遍面临的痛点：面对不断涌现的 AI 原生初创公司与承诺未来会推出 AI Agent 的传统巨头（如 SAP、Oracle、NetSuite、QuickBooks），企业该如何进行权衡与取舍？ ```mermaid flowchart TD accTitle: 核心记录系统与边缘工作系统架构 subgraph SystemOfRecord["核心记录系统 (System of Record)"] A["传统 ERP (SAP / NetSuite)"] B["审计监管透明 & 故障模式已知"] C["强安全性、刚性防错设计"] end subgraph SystemOfWork["边缘工作系统 (System of Work)"] D["AI 原生初创产品 (Reindeer 等)"] E["跨系统灵活工作流与 Agent"] F["可随时启闭的智能洞察工具"] end SystemOfRecord ---|"底层资产沉淀"| SystemOfWork ``` Mayank Sharma 明确划分了“核心底座”与“业务边缘”的界限： 1. **核心记录系统（System of Record）保持绝对稳定**：涉及总账（General Ledger）、子账本（Subledger）以及资金划转的关键核心系统，企业必须保持极高的审慎度。SAP、NetSuite 等传统巨头拥有数十年行业背书，其系统运行机制与故障模式（Failure Modes）被审计师和监管机构透彻理解，具有不可替代的可靠性。 2. **边缘系统（System of Work）大胆拥抱创新**：AI 工具非常适合部署在边缘层，用于生成各类业务洞察与情报，因为这些工具企业可以随时“开启或关闭”，即便出现异常也不会危及底层核心账目。 Yoav Naveh 补充道，核心 ERP 系统在设计上就必须保持“刚性与防错”（Rigid by design），以确保数据绝对安全，因此指望刚性的传统核心系统自身演化出高度灵活的跨系统 AI Agent 是不切实际的。现实中大量的业务工作流天然跨越多个软件孤岛，这恰恰为灵活的 AI 原生边缘层产品提供了巨大发展空间。 --- ### [042:20] 风险承受度与市场定位：中端市场机会与 AI 原生工具的成熟度现状 针对数百亿美元风投资本涌入 AI 原生 ERP 的现象，与会嘉宾探讨了不同体量企业的接受度差异： - **更换系统的先决条件**：Yoav 认为，已经稳定运行 SAP 的大型银行绝不会仅为了“迎接 AI 革命”而耗费两三年去重新实施系统。AI 原生 ERP 的真正目标客群在于那些业务规模快速扩张、本来就正处于“必须在传统 ERP 与新一代系统之间做选型抉择”的中端市场（Mid-market）企业。 - **AI 原生工具的成熟度瓶颈**：作为负责技术实施的咨询专家，Jonas Melton 直言，其团队在持续评测各类 AI 原生产品时发现，当前许多 AI 原生方案的**整体成熟度依然不足**。 - **落地切入点**：这并不意味着 AI 原生产品没有市场价值，而是表明它们目前更适合在受限、垂直且具体的利基场景（Narrow Use Cases）中率先落地，企业在选型时需要对其应用边界保持清晰认知。 ### [44:42](https://www.youtube.com/watch?v=xntbxng9yRw&t=2682s) 打造 AI 原生财务生态与工具成熟度评估 在探讨现代财务系统架构时，嘉宾们深入分析了新型 ERP 与专项工具集成的现状。许多新兴的系统本质上仅仅是一个总账（General Ledger, GL），甚至不具备库存管理、应收账款（AR）或应付账款（AP）等核心功能。这种局限性迫使企业必须与其他专业解决方案进行整合。 目前行业正在探索一种将核心 ERP 与 AI 原生工具（例如 Ramp、Tabs、Numeric 等）打包集成的策略，以此构建一个互联互通的 AI 原生解决方案生态系统。然而，嘉宾指出这背后存在重大的战略权衡：企业需要明确整体战略并评估自身对新技术的接受度。尽管这些 AI 原生工具具备前瞻性，但与那些深耕行业 10 到 20 年的传统遗留系统相比，其整体成熟度仍显不足。主持人 Joe Dormani 随即风趣地补充提到，Tabs 正好是 Thomson Reuters Ventures 的投资组合公司，听到它被提及感到非常高兴。 ```mermaid flowchart TD accTitle: AI 原生财务生态系统构建 subgraph AI_Ecosystem["AI 原生财务生态系统"] direction TB GL["核心总账 (GL)"] Ramp["Ramp (费用管理)"] Tabs["Tabs (应收/计费)"] Numeric["Numeric (月结对账)"] GL <--> Ramp GL <--> Tabs GL <--> Numeric end AI_Ecosystem --> StrategicFit{"战略评估：成熟度 vs. 传统遗留系统"} ``` - **生态整合需求**：纯 GL 架构的新型系统缺乏 AR、AP 和库存等功能，高度依赖第三方工具生态。 - **成熟度对比**：AI 原生工具在创新性上有优势，但在稳定性和功能完备度上与发展数十年的传统财务软件仍有差距。 --- ### [45:42](https://www.youtube.com/watch?v=xntbxng9yRw&t=2742s) 财务掌舵者的 AI 指南：目标导向、专属负责人与洞察转化 在圆桌会议的最后环节，主持人邀请各位嘉宾为财务管理者（CFO）提供一条关于制定 AI 战略行动手册（AI Playbook）的核心建议，嘉宾们从不同维度给出了务实的落地指导。 Jonas Melton 强调必须“以业务结果为起点”（Start with outcomes）。他指出许多客户常陷入盲目追问“别人在做哪些用例”的误区，但其他企业的痛点未必是自身的痛点。企业应首先锚定要达成的业务目标与痛点，再判断解决该问题究竟需要优化流程、清洗数据，还是真正引入 AI。 Yoav Naveh 提出明确权责是战略成功的关键。探索和制定 AI 战略本身就是一项“全职工作”（full-time job），在推进过程中会不断遇到预料之外的新问题。因此，企业必须尽早指定专人负责（Find an owner），由其负责战略的持续迭代、调整与落地执行。 Mayank Sharma 则告诫大家绝不能忽视 AI 带来的行业变革。他建议从应付账款（AP）等小规模、高重复性的任务切入作为试验田。最重要的是，企业绝不能仅仅“把节省下来的时间存起来就完事了”（don't just bank the savings of time），而必须提前规划好如何利用这些释放出来的人力与时间，将其转化为更深层次的战略商业洞察。主持人随后宣布会议圆满结束并致谢，现场报以热烈掌声。 - **结果导向而非用例驱动**：厘清核心业务挑战，再反推是需要流程优化、数据治理还是 AI 技术介入。 - **设立专属负责人**：明确 AI 战略制定与实施的责任人，确保策略能够敏捷迭代。 - **时间节约转向价值创造**：从小而重复的环节（如 AP）起步验证价值，并将释放出的时间系统性地转化为高质量的战略决策洞察。 ## Notable quotes > [16:05](https://www.youtube.com/watch?v=xntbxng9yRw&t=965s) 训练大模型运行工作流的最佳方式，就像你在组织中培训一名新员工一样。新员工入职时，绝不会去死记硬背过去三年的历史数据，也不会去背诵 200 页的操作手册。他们只需要了解业务流程的整体运作框架，看几个具体案例，然后在实际工作中边干边学。 — Yoav Naveh > [16:05](https://www.youtube.com/watch?v=xntbxng9yRw&t=965s) 在我看来，在数据就绪度（Data Readiness）上投入任何资金都是在浪费时间。 — Yoav Naveh > [22:50](https://www.youtube.com/watch?v=xntbxng9yRw&t=1370s) 当今精明的财务总监（Controller）和会计师并非一日炼成，他们历经多年基层磨砺才积累出敏锐的商业直觉与专业经验。 — Mayank Sharma > [25:48](https://www.youtube.com/watch?v=xntbxng9yRw&t=1548s) 我花了整整二十年时间成为业内顶尖高手，如今脚下的地毯却被瞬间抽走，不得不彻底颠覆自己的工作方式。 — Yoav Naveh > [31:32](https://www.youtube.com/watch?v=xntbxng9yRw&t=1892s) 最懂数据与业务细节的人就在财务部门内部，将构建权交给业务专家，才能避免技术团队交付出脱离实际的错误方案。 — Yoav Naveh > [35:14](https://www.youtube.com/watch?v=xntbxng9yRw&t=2114s) 这根本不是 AI 的使用场景，这类需求只需要简单的 RPA 或现成的开箱即用对账工具即可搞定。 — Jonas Melton > [35:14](https://www.youtube.com/watch?v=xntbxng9yRw&t=2114s) 先建立正确流程，再打通数据集成，然后应用 RPA 等基础自动化，最终在顶层叠加 AI 赋能高阶智能。 — Jonas Melton > [45:42](https://www.youtube.com/watch?v=xntbxng9yRw&t=2742s) 企业绝不能仅仅把节省下来的时间存起来就完事了，而必须提前规划好如何利用这些释放出来的人力与时间，将其转化为更深层次的战略商业洞察。 — Mayank Sharma ## Mentioned Joe Dormani, Mayank Sharma, Yoav Naveh, Jonas Melton, Thomson Reuters Ventures, EnergyX, Reindeer, RSM US LLP, ChatGPT, Siri, Claude, Salesforce, QuickBooks, Sage, Ramp, SharePoint, DataRails, SAP, NetSuite, Oracle, Tabs, Numeric, 《绝世天劫》（Armageddon） ## Video description Joe Dormani - Partner, Thomson Reuters Ventures Mayank Sharma - CFO, EnergyX Yoav Naveh - CEO, Reindeer Jonas Melton - Principal, RSM US LLP Corporate finance leaders are shifting from reactive reporting to proactive insight—and AI is the engine. In this session, finance executives will unpack how they're deploying AI to modernize core functions and generate strategic foresight. Learn how AI is helping organizations identify risks earlier, forecast more accurately, and reduce costs through smart automation. Panelists will also discuss how they’re managing AI transparency, internal controls, and compliance amid evolving regulatory landscapes. [Source](https://www.youtube.com/watch?v=xntbxng9yRw)

[16:05] Yoav Naveh: 训练大模型运行工作流的最佳方式，就像你在组织中培训一名新员工一样。新员工入职时，绝不会去死记硬背过去三年的历史数据，也不会去背诵 200 页的操作手册。他们只需要了解业务流程的整体运作框架，看几个具体案例，然后在实际工作中边干边学。 [16:05] Yoav Naveh: 在我看来，在数据就绪度（Data Readiness）上投入任何资金都是在浪费时间。 [22:50] Mayank Sharma: 当今精明的财务总监（Controller）和会计师并非一日炼成，他们历经多年基层磨砺才积累出敏锐的商业直觉与专业经验。 [25:48] Yoav Naveh: 我花了整整二十年时间成为业内顶尖高手，如今脚下的地毯却被瞬间抽走，不得不彻底颠覆自己的工作方式。 [31:32] Yoav Naveh: 最懂数据与业务细节的人就在财务部门内部，将构建权交给业务专家，才能避免技术团队交付出脱离实际的错误方案。 [35:14] Jonas Melton: 这根本不是 AI 的使用场景，这类需求只需要简单的 RPA 或现成的开箱即用对账工具即可搞定。 [35:14] Jonas Melton: 先建立正确流程，再打通数据集成，然后应用 RPA 等基础自动化，最终在顶层叠加 AI 赋能高阶智能。 [45:42] Mayank Sharma: 企业绝不能仅仅把节省下来的时间存起来就完事了，而必须提前规划好如何利用这些释放出来的人力与时间，将其转化为更深层次的战略商业洞察。

[0:04] Mayank Sharma: yet another co-worker in the team [5:08] Mayank Sharma: continuously updating forecasts [10:55] Yoav Naveh: Oh, we got Copilot to everyone in the organization [12:35] Yoav Naveh: waste of time and waste of money [12:35] Yoav Naveh: garbage in, garbage out [22:15] identity crisis [31:27] Teach Drillers to Be Astronauts [31:27] A prepaid is a prepaid and an accrual is an accrual. [37:32] Joe Dormani: forward-deployed engineer [41:12] Yoav Naveh: system of work [41:12] Yoav Naveh: prepare for the AI revolution. [45:35] Jonas Melton: Start with outcomes. [45:35] Mayank Sharma: Don't ignore the change because it's happening for real.

[0:04] very much looking forward to the conversation  today. As was just mentioned, we're going to be [0:08] talking about uh automation and AI in the office  of the CFO. And we're going to be talking about [0:14] it from a number of different lenses. what  these leaders see as the opportunity to use [0:19] AI in the office of the CFO and how some of  the most forward-looking teams are thinking [0:24] about deploying AI and using uh in training their  teams how to use the most cutting edge AI. So, [0:31] uh we're looking forward we'll we'll have a  very informal and general discussion. Um maybe [0:36] what makes the most sense is we'll start with  a quick round of introductions. Hello everyone, [0:42] my name is May Sharma. I'm the chief financial  officer for Energy X. Energy X is a lithium [0:49] technology company. Um, if you think about  electric vehicles and lithium-ion batteries, [0:55] lithium is a common raw material that powers  that. And so, we're in the business of producing [1:00] lithium to enable the clean energy revolution  to move forward. Hi everyone, I'm Yav Nave. I'm [1:07] the co-founder and CEO of Reindeer. We work with  enterprises and help them run agents focused on [1:12] automating core operation workflows from account  payable, account receivable, tax preparation, [1:19] uh supply chain operations. And I think what we're  seeing mostly uh in in that space, we're seeing [1:25] a lot of [snorts] attempts, a lot of uh people  trying and looking forward to talk a little bit [1:31] about kind of how that potentially changes the  office of the CFO in the next few years. And I'm [1:36] Jonas Melton. I'm a partner at RSM. We are uh one  of the largest public accounting firms in in the [1:44] world. Uh my role is really around automation and  AI and how we deliver those technology solutions [1:51] to our clients. Fantastic. And so one of one of  the things that we all are feeling I think in the [2:00] market is just the pace of change in in the market  when it comes to AI. And so in preparing for this [2:06] uh we thought it might be good just think about  the last 12 months. what what are one of the [2:11] things that you've been most surprised about or  one of the things that you've observed in the way [2:16] that a CFO function has been operating maybe from  a client perspective from a vendor perspective or [2:22] from an internal organization perspective you  know what are what one of the things that you [2:26] know over the last 12 months has surprised you  go one by one absolutely so we uh so I'm myself [2:33] a CFO and I've been pushing for the use of AI  within my teams and the thing that surprised [2:39] me the most is over the last several months the  attitude that the team has had towards AI. So it [2:45] went from largely the team being indifferent  to AI and almost resisting the implementation [2:52] of AI going from that to begrudgingly uh  accepting the use of AI into our workflows [3:00] uh towards a more full-on uh adoption of AI. In  fact, some of our colleagues now consider AI to [3:07] be yet another co-worker in the team. Uh a and  on the flip side, some of our co-workers actually [3:13] trust AI more than other co-workers in the team.  So for me, it's been this full full cycle adoption [3:21] uh and which has turned many critics into avid  supporters of AI. I think we're we're seeing [3:29] a couple things from the perspective. one is I  think a year ago we've mostly seen um CFOs or or [3:36] uh financial teams experimenting with agents and  there's something very simple say okay I imagine I [3:42] have an agent that helps me with account payable  or I imagine I have an agent that helps me with [3:45] another task but now we're seeing CFOs starting to  think okay how does the world look when I have a [3:50] 100 agents solving a 100 different tasks and  we have agents operating alongside people in [3:55] the organization and and that's a little bit of  a mental shift because it's not like a a one-off [4:00] operation It's something you actually have to  manage at a different scale and I think there [4:03] are a lot of questions about who's responsible  of doing that in the organization and I would say [4:09] um just the pace that this has actually rolled  out right if you think about we've been seeing [4:16] AI out in the world right Siri on your iPhone  or whatever it's been out there for a long time [4:23] quite a few years but really over the past 18  months when kind of chat GPT hit the market it [4:29] just exploded and [clears throat] the the pace at  which board of directors or CEOs are expecting now [4:39] uh organizations to adopt AI uh very quickly um  you know and just I guess there's always expected [4:47] to be challenges I think with that that type  of pace um but seeing now like we're actually [4:52] starting to see some real results um which  is which is pretty cool. Yeah, I feel like [4:56] that's a really good synopsis. increased trust,  AI being able to do more and no no stopping the [5:03] pace of progression. Um I think you know probably  many people in the room would would say they have [5:08] similar experiences. Um so I I think one of the  things that um when folks are talking about AI [5:21] ROI, one of the things that are uh talked about a  lot is efficiency. We talk about efficiency, but I [5:27] think what gets talked about less is actually the  new opportunities that AI opens up and you know [5:34] like in the financial world a lot of that goes  around moving up the stack to be more strategic [5:41] advisors. Maybe we can talk a little bit about  some of the efficiency opportunities we see and [5:48] some of the other side of the coin which is what  are the strategic opportunities that having AI do [5:53] more of the repetitive tasks opens up for the CFO  and the off you know the folks that work in the [6:00] office of the CFO. You want to start with Yeah,  absolutely. So the way I look at it is the biggest [6:07] difference that AI um is having is to lower the  cost and effort of generating numbers. So whereas [6:15] in the past finance teams would be expending  a lot of effort in producing numbers and that [6:21] uh level of effort may have only been justified  to produce once a quarter updates. I think AI [6:28] completely takes away the burden of actually  generating the numbers. And so what we find [6:34] within my organization now is instead of having  a quarterly or monthly cadence of updating our [6:42] forecast, we actually have a continuously updating  forecasts. And uh furthermore, uh AI does give you [6:49] a lot of power in terms of looking at information  in a granular way in that the analysis of the [6:55] granularity can be left to AI. rather than people.  And so one of the other things that it is allowing [7:00] us to do is uh within my company, we discuss with  each and every department on a monthly basis and [7:08] we really look at uh a a set of numbers that  can be drilled down to the level of individual [7:14] invoices. And so on a running basis from these  monthly conversations with different departments, [7:20] we are able to prepare a very granular picture of  what's going on. But a lot of the heavy lifting [7:26] for generating commentary can also be left to AI.  And so what this is allowing us to do is to have a [7:33] constantly updating forecast for the company. And  so that really is an important switch in terms of [7:40] a the timing that you have saved in generating  the numbers and b what you use that time saving [7:46] towards. And so for us as a company and for the  CFO office that time is better used in asking [7:53] questions uh and analyzing the data and really  being involved as a finance function early on [8:01] in the decision-m process rather than reporting a  number on a postfactor basis. Right? So it's that [8:09] uh a lowering of the effort required to  generate numbers which gives us time savings. [8:14] But the real win in here is to use that time  saving and use it to really provide strategic [8:21] uh drivers and insights into the business. I mean  I really agree with that. I think that what the [8:26] biggest take away from me for what you said is how  do we use all this to do better forecasting and to [8:31] streamline process that we have and I think when  we think about finance organizations in the end [8:35] of the day what they do is they look at the past  the past data and try and build models that would [8:40] help make better decisions for the business make  better strategic decisions and I think often we [8:44] get caught up with efficiency there's definitely  efficiency but if you look at the waves that [8:48] happen with AI the biggest one okay there's coding  agents that's a whole different topic that works [8:54] really well but In business, it's mostly customer  support. Customer support is one big cost that [9:01] some organizations pay million dollars a year that  if you bring AI, you can save a lot of money. With [9:05] finance, the problem is that I'm not sure there's  many individual workflows that are that big. You [9:12] have a 10 people doing this and 20 people doing  that, maybe three people doing uh this. Sometime [9:16] it's 20 people doing multiple things. It's very  hard to see short-term efficiency gains that are [9:21] that meaningful. five headcount that's not that  meaningful to me the biggest advantage is once [9:26] you do enough of these things it's actually okay  you'll see the efficiency but it's really now we [9:30] have data readily available all the manual process  that create delay in getting data ready that [9:37] you get quarterly reports that are 2 months or 3  months at best uh after the quarter ended imagine [9:42] you're starting to get all these agents to kind  of bring you that data immediately so you're much [9:46] closer to real time and your entire forecasting  is not based on delay data my man that's the [9:51] real value of what you're able to achieve with  improving u um like operations but generally [9:58] like efficiency will be there but sometimes we  just get caught around that and looking for that [10:01] short-term gain because it is work to get to a  point that you have 60 70% of uh of your process [10:08] that are streamlined with agents and I would  just kind of add to that the one of the things [10:15] beyond efficiency that I've seen is just better  intelligence right if you think about you know now [10:21] you've at some type of model or agent that talk  about a flux analysis um or variance analysis at [10:27] month end, right? If you think about how deep the  agent can go digging into transactions and and the [10:34] insights that it can pull out versus someone  manually sitting there doing that, they're [10:38] not going to catch everything. An AI agent or a  model is going to catch way more things or produce [10:44] different insights uh and probably more insights  than what a person ever could get to. And so I [10:50] think that there's a lot of value in that as well.  I think there's a separate track like I always [10:55] think there's agents that kind of take repetitive  work and try and streamline it, automate it, [10:59] make it faster, more accurate. But then there's  the assistant route. In my mind, a lot of I was [11:04] at another CFO event and the most common theme  when I asked CFOs was what are you doing with [11:09] AI? Said oh we got co-pilot to everyone in the  organization. She's like okay great but what [11:13] do they do with it? We don't really know. I think  there's so many layers about connecting it to the [11:18] relevant system in data and personalizing it. So  the the assistant that the controller is using is [11:24] different than the assistant that someone in uh  I know the financial crimes department is using [11:29] and giving people the ability to use that and and  be able to make those smarter decisions. I think [11:32] that's definitely a super interesting track that  hasn't been cracked yet. Yeah, we we typically [11:39] are bucketing like AI into three three things.  personal productivity, right? Which is kind [11:45] of what you just talked about with uh you know,  everybody's got a co-pilot or a chat GPT license [11:51] and they're using it for their own to draft an  email to search their inbox. The second one would [11:56] be more of an organizational or enterprisewide  type of agent. So maybe a chatbot um or maybe a [12:02] a skill or an agent that multiple team members  could use. And then there's this third layer [12:08] which is probably more complex but this layer of  enterprise intelligence right where you've now got [12:14] a large database that you're running uh uh some  type of agent over some type of model to pull out [12:22] um real insights for the business. Um so again  kind of increasing in complexity but um kind of [12:27] how we view the different layers of of AI. Great.  Um, one of the things I think we want to tackle [12:35] is AI readiness. I think there are two lenses that  we kind of framed it up around, one is messy data, [12:42] right? Very common problem in many finance  functions. You have data in disparate systems. [12:47] There's no commonality in how to get data into one  place to slice and dice it and prepare it for the [12:54] reporting needs of of the organization. And then  the other one is change management. to like more [13:00] of the people aspect of, you know, getting people  to change the way that they think about doing [13:05] their work and how they pull numbers together and  actually produce the things that they need to do [13:10] to push the business forward. So, so maybe um you  can get us started from a client perspective like [13:16] what are you seeing in terms of uh your clients  ability to address that sort of data wrangling [13:23] issue and you know change management aspect which  I'm sure you encounter a fair fair amount. Yeah. [13:28] So we we call this what RSM foundations right and  it's really this idea of kind of readiness across [13:36] the board and it's it encompasses a few different  things right so you mentioned the data um and [13:42] it's really very like use case specific uh to a  certain extent because every model or every agent [13:49] that we're building there is an underlying data  source or knowledge database right so when I if [13:56] I go back to the previous example example I gave  around enterprise intelligence or you know you're [14:01] talking about maybe entire data warehouse or data  lake uh with information from a lot of different [14:06] systems that is needed to pull that information  and that data and provide those insights whereas [14:12] if you're just creating a one-off agent you can  drop a you know I've created an agent for a client [14:18] once that we put it was for a system that we had  implemented but we put in product documentation [14:27] and user guides into a simple knowledge database  and created a chatbot. So any of their users [14:32] could drill back into um those user guides in  that like how to add a new user to the system, [14:39] how to you know change a prepare and a reviewer to  a new person like all of those things they could [14:44] do through a simple chatbot. So that's a much  different data exercise. Now you're talking about [14:50] pulling a few documents together versus wrangling  a whole a whole data set. And so that's a big you [14:56] know the use case really drives kind of how much  data work is required um from a change management [15:02] perspective you know talent and and readiness  in the organization is also huge right I think [15:08] um my you mentioned like the team kind of got  on board and and uh you know came a long way [15:14] over the past year um but that's something that  we constantly are looking at as well when we go [15:18] in and do a readiness type project for a client is  to make sure that they've got the right structure [15:25] in place to enable and bring bring people along.  Um the other piece that we look at a lot is is the [15:32] technology that's in place. So what solutions are  we trying to roll out? Um so you know we I think [15:40] we'll get into some talks around the different  types of solutions, but you know have we taken [15:46] advantage of the tools that we already own or are  we going out and trying to build or buy something [15:52] custom? Um so again there's a lot of angles I  think to readiness. Um but yeah data and and [15:58] change are two of the bigger ones. Maybe just  to to add to the question who who owns the data [16:05] wrangling issue. Is that a data engineer question?  Is do we need data engineers in the finance [16:12] function? Like who who actually functionally owns  getting the data in the right state for it to be [16:18] AI ready? I say from my perspective any investment  in data readiness is a waste of time. I think that [16:27] when you look at the data in your organization  if you look at the last two years worth of data [16:31] a lot of it is no longer relevant. Your business  changes all the time. I see organizations spending [16:36] millions and trying to get data engineers to clean  up the data. I think a lot of that comes because [16:41] traditional AI models require like the garbage in  garbage out required a very large sample of data. [16:48] LM don't work that way. Uh I think the best  way to think about how to train LLMs to run [16:53] your workflows is the same way you train a new  employee. A new employee doesn't go and memorizes [16:58] three years worth of data or go and reads 200page  manual about how the work gets done and memorizes [17:03] that. They get explanation about the how the  general process works. They get some samples [17:07] and then they learn on the job. They do what's  clear to them and whenever something is not clear, [17:12] they ask questions. ALMs are built for that  type of process and it's much faster to train [17:18] it. Imagine you have a super new employee that is  able to work at a capacity of 100 others. You'd [17:24] still teach them in the same way. You'd still  tell them start with some samples when you don't [17:28] know something. Keep training and they would be  much better on boarded because they will do that [17:31] on fresh data. Your organization is generating  new data every day. If you're able to run that [17:37] through that process, it's going to be so much  faster than trying to look at a year's two years, [17:41] three years worth of data and trying to clean it  up involving data engineers. That that's always [17:46] fun jumping through the engineering departments  trying to get their approval to do things. So [17:50] like to me it's like always such a frustrating  conversation when we meet an enterprise that tries [17:54] to do that because I feel it's a waste of time and  waste of money for them. Yeah, that that's great. [17:57] So then who owns the AI training process? Is it  the subject matter experts? Is it the you know [18:06] always on transformation office? So I I think the  reality is that person in most companies doesn't [18:14] exist yet. I think again I talked about in the  beginning in the sense that it's very easy for [18:19] me to think about I'm implementing one agent in  my organization and if I think about I'm going [18:24] to have a hundred imagining that works without  someone orchestrating that is to me that sounds [18:31] insane. The same way you have a team that helps  you hire people and train them and give them [18:38] performance reviews and um run learning seminars  and like that I think something would have to [18:43] emerge for for agents. We're seeing that pick up  in some enterprises. Sometimes it's the under the [18:48] CIO, sometimes the head of innovation, head of  transformation, but in my mind it's similar to a [18:54] lot of other departments that have kind of risen  like you know revenue operations wasn't a such a [18:59] big thing uh before Salesforce became such a big  thing. I think we'll be years from now, we're [19:04] going to have a team that will be responsible of  measuring ROI for agents and understanding the the [19:10] trade-off between cost and and speed and uh and  accuracy and being able to understand how do we [19:16] teach agents and how do we on board them um and  how do we train the organization to use them. So, [19:22] so I think the time is now and really trying  to figure out who's going to do that. It can [19:26] happen within the finance organization, someone to  own that. It can happen companywide. But I think [19:32] if we all agree as superi strategic, not pointing  someone to obsess about that problem today is not [19:38] maximizing your probability of succeeding. Maybe  if I can add to that, I think the AI readiness [19:44] question is broader than just a question on clean  or messy data, right? Because there is obviously [19:51] people involved and systems involved. On the  people's side, there needs to be a mindset change. [19:58] uh in finance function you know finance people,  accountants, controllers, they have grown up with [20:04] a mentality of you know producing um uh compliance  statements following protocols uh gap procedures [20:14] and having a uh monthly cadence or a quarterly  cadence of reporting which is a very disciplined [20:21] art that they have learned over time but it  doesn't easily lend itself to looking at the data [20:27] uh flexibly and querying it flexibly looking at  the anomalies and going to try to understand them. [20:34] So there's definitely a mindset mindset shift  that's required within the finance function [20:40] for people who are looking at the data and trying  to detect the forward-looking trends therein. The [20:46] other issue as well is on uh in terms of systems.  So in any given company uh the accounting system [20:53] may be something like a QuickBooks or Sage.  Uh in our company we also use for example [20:59] RAMP for bill payments uh and uh uh purchase  order generation and reimbursements. There may [21:06] be other policies and documents which are on  systems like SharePoint. There may be project [21:13] management software etc. All of these systems  would have been chosen by different teams at [21:20] different times for the right reasons, but not uh  envisaging that one day you will need to point an [21:28] AI agent at it and need to be able to seamlessly  read all the information therein. And for example, [21:35] I'm not even sure if there's a claude connector  for QuickBooks. I know that there was one for [21:40] uh for data rails but there was a payment involved  and you have to go through the payment and effort [21:46] of implementing that connector. Um I I think  that is a process that will take time. It is [21:53] not as easy as pointing an AI agent. I think  it's a mix of having connectors for some system [21:58] or building a manual bridge to uh extracting data  from other systems. So one has to tackle both the [22:08] people and the system level complexities first  before you can even start asking yourself the [22:15] question is the data clean or messy. Yeah,  that's great. I I want to double click into [22:20] the human aspect of this because in in many  respects that is the hard part and I know that [22:26] we probably all feel passionate about the art  of compliant forward-looking statements here. [22:31] Um but but I feel like a lot of what I would see  or what my portfolio companies see or my internal [22:38] teams a lot of what you encounter sometimes like  we throw around words like efficiency. We talk [22:44] about subject matter experts training AI. I think  there's an underlying current of anxiety around [22:50] people worrying about AI replacing them in in  the workforce and you know doing what they used [22:55] to do. And you know we talk about efficiency, we  talked about the opportunity to be more strategic. [23:01] What do you think the real actual impact is on the  finance organization from a headcount standpoint [23:07] and from an organizational structure standpoint?  Do you think it the opportunity is to reduce the [23:13] headcount in in finance or you know what is what  is the actual organization of the future kind of [23:19] look like? I I'll use uh my own team as an  example because we are seeing uh AI having [23:26] an impact. So in this specific instance when we  lost our accounts payable person we didn't feel [23:33] the need to backfill that role. We found that the  AI agents are perfectly capable of taking on the [23:41] uh you know a vast majority of the work required  for AI AP related work. Now it hasn't taken off [23:49] the burden entirely from other people because  you still have to oversee the work produced by [23:55] the agent make sure you as human agree to the  output that is produced and ultimately sign off [24:00] on it. But what that has demonstrated for us is a  yes it is having a downward impact on the number [24:09] of people you will need in the organization but  b also it is changing the type of people that are [24:16] therefore coming in because uh a lot of the junior  entry-level positions are the ones I believe are [24:23] most exposed. So if you would have thought about  a traditional finance team being a pyramid where [24:28] the juniors were the base of the pyramid maybe  larger in number doing reconciliations, payments, [24:34] invoice processing etc and then generating  the output for the seniors to have uh a bit of [24:40] strategic input into. Well now that junior layer  is being replaced and it's not just a a problem [24:47] of the new entrance coming into that junior layer.  It's also uh a problem for the future senior folks [24:55] because the savvy controllers today, the savvy  accountants today didn't become savvy overnight. [25:02] They had many years of uh training at the junior  levels to develop that sense and experience. And [25:08] so if you're going to be replacing the junior  layer away, then you have a problem of how the [25:14] uh seniors of the future will indeed get the  training. And then is the other question of [25:19] where do the junior people come in? What sort  of uh criteria and skills they should have [25:25] uh that allows them to thrive in this new era  where I think if you're trying to be kind, [25:31] you may say it's not job replacement, it's  job augmentation, but in some cases it is [25:36] job replacement. So how do those junior folks  equip themselves with the knowledge that you [25:42] know they're able to use AI and really go for  that augmented output? So what what's the answer [25:48] to that question? Time will tell. [laughter] I I  think so early it would be guessing. What I like [25:55] to do in that case is try and find where is it  already happening. The place that is most advanced [26:00] in AI adoption today is engineering team. software  development has changed drastically by the way not [26:06] that long ago I think the moment the the pivot  was December uh last year and today all the good [26:13] engineers that I know they don't even open they  don't even see the code anymore they manage set [26:18] of coding assistants and agents that do the work  for them and their work dramatically transitions I [26:25] think the senior people are still very coveted  they have a lot of opportunities by the way a [26:31] lot of the opportunities is building and adapting  uh solutions for other enterprises. Um but their [26:37] role changed dramatically and the ones that were  able to adapt that. I think the role of junior [26:41] engineers disappeared and I don't think that's  coming back. I think if you're going into that [26:47] space there's going to be potentially a new role  that is coming up but I don't think there's a need [26:50] for junior engineers uh anymore. I think there's a  need for people who are able to orchestrate agents [26:55] or a need for people who build infrastructure of  feeding closing the loop of feedback of whether [27:01] the are the decisions that the coding agents made  are good or bad and I think that's going to happen [27:06] very similar across every department. So I think  the the senior people would like become people who [27:13] manage and orchestrate agents and I think the  junior ones that those roles are going to get [27:18] disappeared. But does that mean that there's not a  need for junior engineers? Or does that mean that [27:22] the job that junior engineers used to do is no  longer needed? Therefore, we need to create space [27:29] for and think about roles that are more aligned  to what the senior engineers are doing for an [27:36] entry-level constituency. So, so again, this would  guess what the way things are progressing, I think [27:42] the real risk is for the role of senior engineers.  I think that AI just keep going up and and to a [27:48] point that it's going to be able to take more and  more decisions. Again, I mentioned December. What [27:53] happened before is not that AI wasn't writing  code, but people were instructing what code [27:59] in specific blocks and specific segments and were  looking at all the codes I was writing and there's [28:03] a point that there was enough knowledge that a  new model shifted and just took over that. Sure, [28:07] there was a little bit of time of I've seen  some of my engineering friends had like a [28:12] identity crisis. I spent 20 years of my career  learning to be really really good. I'm one of [28:17] the top in my game and doing something and that  the rug got swept under me and I need to change [28:21] completely what I'm doing. So it it it is scary.  I think engineers are probably a little bit more [28:26] uh fast adopters than than you know finance people  operations people but I think that's coming. [28:31] Do do you think this maps well to the finance  function? Because I mean there's a very broad [28:37] set of things that could potentially fall into  the purview of a finance function. Like there's [28:42] much more you can do on the commercial side.  Like if if you don't have to spend all of your [28:46] time creating reports and building models and you  know working on forecasts you know you can spend [28:51] more time maybe doing more in-depth analytics  of you know the way the go to market function is [28:55] working or you know the way customer acquisition  is working to help you know accelerate other parts [29:00] of the organization. I mean finance has a very  broad mandate in in a lot of organizations. So, [29:06] I'm just just wondering if do you think there this  is purely like we're going to replace the junior [29:13] layer or is there an opportunity to broaden the  opportunities? I would say that it I don't think [29:21] it's going to completely replace the junior layer  only from the standpoint longer term somehow we've [29:28] got to get more of the middle layer upper layer. I  mean, I'm not working forever. So, at some point, [29:35] people are going to retire or move on. And  we've got to figure out how do we get the next [29:40] layer upskilled enough that they're doing the  higher level tasks. Uh because I do think that [29:47] agents will replace some of those lower level  tasks, especially in the areas that May mentioned, [29:53] the areas that are very scalable, right? very  repetitive like the AP processing, AR processing, [29:59] even some of the FPNA work that's that's happening  today. So, I think we're already seeing that stuff [30:05] where it's very scalable. That's where we're  seeing significant headcount reductions. Um, [30:11] in the other areas more in whether it's financial  close or some of that, again, those aren't as [30:17] scalable. I still think there's going to be  and there already are agents that can go and [30:23] do certain of those functions and certain of those  tasks, but there's still that in the loop element [30:28] of reviewing. Um there's still going to be some  governance. You have mentioned earlier there's [30:33] this idea that there might be new roles that we  don't even know about today. I think that's very [30:38] true. whether it's building or maintaining agents,  um whether it's the orchestration side of it and [30:44] and kind of who's setting that up and maintaining  that. The biggest gap that we're seeing in the [30:50] market is this bridge between the engineers, the  real technical people who can actually build some [30:57] of these things and the finance individuals who  understand the processes and how do those people [31:03] work together and collaborate to create something  that's that's effective. Right? By the way, [31:09] that gap is decreasing as well because today as a  finance person, you can build a lot of things that [31:13] three years ago there were not even an option.  I don't think I don't think it's going to be a [31:18] company run just by agents. I think people are  still going to be a major part of companies. I [31:22] just think it's going to be new roles. So, it's  it's a great segue and I'm glad we double click [31:27] into that because I think it's a question on a lot  of folks mind what happens with with talent. Um, [31:32] but it's a great segue because I think we  wanted to dive into, you know, this question [31:36] of Yeah. like now that it's much easier to build  software, custom applications, custom workflows, [31:43] um, and the barrier between, you know, getting  a software engineer to develop those workflows [31:48] and just maybe having someone with some technical  aptitude on the team being able to spin those up [31:53] is is obviously uh, breaking down. How do you guys  think about where and when it makes sense for you [32:01] know someone in the finance function to build an  application versus just working with you know an [32:08] existing vendor that you know has has a you know  customizable application off off the shelf because [32:14] I think I think we we're in a phase right now I  I see it a lot at at Thompson Reuters where a lot [32:20] all the different internal teams in finance and  across the organization are essentially building [32:25] their own custom some applications. There's a  lot of experimentation. I'm not convinced that [32:31] it's the best use of dollars. So, you know,  maybe Jo you you start us off. Yeah. I mean, [32:39] I think that the way software has been built  historically is the engineers came in try to [32:44] understand or the product people try to understand  your process and give you a solution and then the [32:48] company had to adapt to use a software to run  the the process certain way. I think one of the [32:52] biggest changes now is that software should adapt  to your process. And and you heard me say this, [32:58] I think it always reminds me of the the Armageddon  movie where, you know, there's an asteroid coming [33:04] uh to to blow up the Earth. There's a question,  are we going to take the astronauts and teach them [33:09] how to drill a big hole so they can put uh the  bomb inside it that would blow up the the asteroid [33:16] or are we going to teach the mind drillers on  how to become astronauts? which are obviously [33:20] what they choose to do and to my mind it it's kind  of the same the answer of the engineers that oh [33:24] they they know better like no let's give tools to  finance people to actually build the logic that [33:30] they need um and implement that because honestly  like most of the data is here there's so much [33:36] institutional knowledge trying to give engineers  to build that gives us the wrong solution in my [33:41] mind that doesn't mean that they need to build  runtime environment that doesn't mean that they [33:45] need to build integrations to different systems  but I think our our goal when we think about it [33:50] is how do we empower those people to be able to be  independent maybe not today but in a year and two [33:56] years and three years and how they manage their  agents and use their institutional knowledge to [34:00] do that instead of relying on engineers  every time they want to make a change. [34:05] I think I'll agree with you up uh because it's  not as if uh the in the finance function uh [34:12] we're tasked with building complex software uh  it's it's a lot about uh crunching the numbers [34:19] uh trying to build uh trends uh try trying to  see what the trends are and ultimately maybe [34:24] an output could be building a dashboard. uh  a lot of this in my team have started to see [34:30] uh the technically minded juniors have taken  this uh in their stride and you know come up [34:35] with dashboards they are big on vibe coding uh  ultimately that is what you need uh because the [34:43] uh requirements of the finance function can change  over time and by relying on external engineers to [34:49] ma you know maintain a system can in itself bring  some friction into the process and and the tools [34:55] to develop dashboard boards, analysis, reporting,  insights are really that simple with AI that the [35:02] teams can be, you know, themselves empowered to  develop, right? So, we're not giving up on our [35:07] junior finance professionals just yet. We're not.  I I look at this very much from a persona lens. [35:14] So when I think about what we do in the  office of the CFO, I think 80% of it is [35:22] fairly repetitive. No matter what industry you're  in or what organization you're in, a prepaid is [35:27] a prepaid and acrruel is an acrruel, right?  And so to me to go and build something custom [35:34] when there are tools that are out of the box or  pre-built, uh it doesn't make a lot of sense. [35:39] Why build something that's already there? When  you think about other parts of the organization, [35:43] maybe it's the office of the COO or your product  team, that is your secret sauce, right? That's [35:49] your unique IP that you're selling out in the  market. And so that is not as consistent with [35:54] other organizations out in the market. And so in  that case, I'm looking at where they're actually [36:00] building more custom type solutions. But in  the in the office of accounting and finance, [36:05] um I'm seeing a lot more organizations trying to  leverage what they have in their tech stack today [36:11] or acquire a new technology to deploy AI within  is that we're working with a very large CPG. We do [36:17] account payable across five different departments  there. Every department has a different process [36:22] and and I think having something pre-built is a  good start, but they all have their own tweaks [36:28] of how they do things. Why and how we got here,  I don't know. But that's what we're see. Yeah. [36:33] And that's fair. And and so when we're looking at  this, we're actually looking at it not just from [36:38] an AI lens, but also from a process lens. So my  question would be, why do they have five different [36:44] processes? Maybe they need to standard standardize  across those different entities or locations. [36:50] We're also looking at the data. We're also looking  at what automation capabilities are out there of [36:56] like when I say automation, RPA, right? robotic  process automation. So I had a client come to me [37:02] last week and they they were doing an account  reconciliation. They had all their a different [37:06] tab for every in Excel of every account and then  they they were pasting in a copying and pasting a [37:12] trial balance into the first tab and that fed the  ending balance on each tab. And he came to me this [37:18] guy and he's like I want to use a claude I want  to use claude to get that trial balance and paste [37:25] it in. I'm like, that's not an AI use case, right?  That is a a simple automation or integration tool [37:32] or thing that they should be using. And there's  tools out of the box that did all that do all the [37:37] account reconciliation activities that they  were trying to do. So again, I think I look [37:42] at it more from is the right process there? Are  we integrating the right data? Are we automating [37:48] where we can? And then let's layer AI on top  to drive even further intelligence. Yeah. A [37:55] lot of that makes sense especially in the age  of the forward deployed engineer when you know [38:00] many of the companies out here can do a lot of  customization and and build some pretty bespoke [38:05] highv value stuff. Um, I I want to double click  into vendor selection a little bit here. And [38:11] I think we're at a point in time where the, you  know, there there more and more AI startups and [38:17] and vendors in the space. And obviously, you know,  there are the SAPs, the oracles of the world who [38:22] have all been around financial software for a  number of years, uh, QuickBooks, Netswuite. Um, [38:28] how do you guys, what would you recommend in  thinking about evaluating when it makes sense [38:33] to maybe work with an AI native startup, a newer  company versus, you know, doubling down maybe with [38:39] an existing relationship with a incumbent  or legacy software provider who, you know, [38:45] AI's on the road map of course, but you know, it's  coming in six months and maybe that becomes eight [38:50] months and, you know, there are a lot of different  trade-offs there. So any advice that you'd give in [38:55] how to evaluate that type of situation and when it  makes sense to go in either direction? Sure. Um I [39:03] mean I'm all for AI but there are use cases for AI  where it's justified. Um but there are other areas [39:11] where you have to be a bit more reticent. Right.  So anything that uh for example you mentioned [39:17] Netswuite SAP for anything as critical as running  an ERP uh system any system that's going to touch [39:24] our ledger we have to be very careful about  uh the selection of the product and the vendor [39:32] because I'm sure there are lots of uh fancy AI  ERP startups out there and some of the stuff [39:39] that they're doing is really cool and interesting  but really their pitch to view is um predicated [39:46] on their own staying power as a company. A lot of  these are newer companies and uh for their AI to [39:53] be working and you contrast that with the decades  of uh track record that people like SAP and [40:00] uh Netswuite have um the fact that they're the  these are systems that are well understood by [40:07] auditors as well as regulators. the failure modes  of these software is well understood. I think [40:14] uh lends itself to you know a really high bar  because it's fine at the edge to be using AI to [40:21] generate all sorts of nice uh insights and intel  uh but only in you know in terms of tools that you [40:29] can switch on and off whereas for something that  is going to commit uh transactions that's going [40:35] to move money that is going to touch my subleddger  uh I'm sorry but at this point uh I have to place [40:42] reliance on an industry tried and tested uh legacy  system like a Netswuite or um or SAP or others of [40:51] that ilk. Um and and so really uh for me it is  down to the fact that you know at you want your [41:01] core to be stable at the edges of the core you  can be as experimental as you want. Uh but so [41:06] far for ERP it's not something that we're looking  to be experimental about. Yeah, I think we have [41:12] a similar framework to this. I think the system  of record, I don't see anyone rushing to replace [41:17] their ERP. Need to have a really good reason to  do that. I don't think it's there. These systems [41:21] need to keep your data secure. They need to be  rigid by design. They need to be complicated [41:26] to kind of break things and and them and what  you call the edges like think about that the [41:31] layer of work or the system of work on top of  those systems need to be super flexible. And [41:36] then you're in a really like this is what's moving  quickly. There are a lot of different opportunity. [41:40] The market is so noisy. You have startup like us  that tell you we're in a horizontal solution that [41:44] can grow with. You have point solutions and you  have the net suites and you're like they tell you [41:48] we're going to build agents for you. And I think  how you decide that strategy. How do you select [41:53] a vendor for that new piece is complicated. Um  my take is that most workflows list leave across [42:00] different systems. So probably your system of  record can really give you being rigid can really [42:04] give you an agent that's going to be flexible to  solve uh that. there aren't really legacy players [42:10] in that because this everything here is entirely  new. Um, and I think you have to move, you have to [42:15] try something. So I think building that strategy  I think is is the most complicated question that [42:20] you guys have to answer. So So there's billions  and billions and billions of dollars, venture [42:25] capital dollars going into AI native ERPs and all  types of financial software. Is there I mean it [42:32] sounds like you guys generally is there another  side of midmarket midmarket. So what what's the [42:38] other side of the argument then? What does what  does mid-market see that maybe they're willing to [42:45] go a little bit further out on the risk curve that  maybe an enterprise would would not? First of all, [42:50] I think that some of them are in a stage that  they have to make the change anyway. I you have to [42:55] have have a reason to make that change. So nobody  implementation is not a good enough reason. Sorry, [43:00] two-year implementation cycle is not good. But  but you're already implemented, right? So I mean [43:04] if you're now standing on the point is my company  has reached a certain amount I have to figure out [43:08] do I put SAP now or do I put this new system  then then I think you have a valid decision to [43:13] make I think if you're running a large bank and  you're like okay I'm already with SAP why would [43:19] I make a change and you're saying oh because  to prepare me for the AI revolution that's the [43:23] wrong the wrong way of looking at it need to  have a reason to make the change if you have [43:27] to make the change okay look at them and let's  see where they are and let's see if you believe [43:30] they'll survive in two years um and I think  that's the the market the companies are at that [43:34] point of making the change are the opportunity  for these AI ERPs. This one's a little near and [43:40] dear to my heart because I oversee a team that  implements accounting and finance tools and and [43:44] technologies. And so we're constantly evaluating  who should we partner with as an organization. [43:50] Who should we refer to our customers and bring  to you guys as as organizations to, you know, [43:56] find success in this area. So when we're looking  at technologies, obviously we've got a lot of the [44:02] triedand-true kind of major players already in our  in our partner ecosystems, we are starting to look [44:08] at these AI native tools and figure out what is  the right fit. And as we've started evaluating a [44:13] lot of these tools, what we're finding is the  maturity is not yet there in a lot of the AI [44:19] native stuff. So depending on what you're looking  at, that's not to say it's not ready for market [44:26] and and it can't be useful, but it might be in  a more restricted or or very narrow um kind of [44:32] use case of how you're using it. So I'll give you  an example. I'm not going to name the name, but [44:36] we were looking at one of the the AI native ERPs  that's now out in the market, and it's basically [44:42] a GL. All they have is a general ledger. Like it  it doesn't even have inventory capabilities. It [44:48] doesn't have AR and AP. So you're basically having  to partner with other solutions. So we're taking [44:54] a look at is there an kind of an AI native bundle  with this ERP and then maybe a a ramp or a tabs or [45:04] a numeric or some of these other tools that are  out there and can we bundle all these together [45:09] and kind of create this ecosystem of AI native  solutions. So again there is a huge strategic [45:18] play here and understanding what is the strategy  of the organization overall and how are we how are [45:23] we how comfortable are we with these tools but um  what I'm seeing is the maturity isn't quite there [45:29] compared to some of the legacy tools who have been  building this stuff for 10 or 20 years fantastic [45:35] tabs TR ventures portfolio company so always  happy to hear that um I think we're we're at [45:42] time maybe we'll just one sort of quickfire last  piece of advice for folks in the CFO uh role on [45:53] you know one word of advice around what their AI  playbook could be. We'll start with you. Yeah, my [45:59] big thing is start with outcomes. I I've seen so  many clients come to me and say what are all the [46:04] use cases? What what what's other companies doing?  what are the and just because one company has a [46:09] use case and they've you know it's a painoint for  them that doesn't mean it's a painoint for you. [46:14] So start with the outcome and understand you know  what elements are going to solve that painoint or [46:21] that challenge for us. Again it could be a process  thing. It could be a data thing. We may need AI [46:25] plugged in to to actually solve this. But start  with the outcome and then work work from there. [46:31] It's a full-time job to figure out a strategy  here. Find an owner. You're going to learn so [46:36] many things that you haven't even thought of.  You need someone to be able to keep reiterating [46:41] on that plan and change it. for a moment decide  who owns the strategy, who owns implementing it [46:46] and start there. I would say don't ignore the  change because it's happening for real. Uh a [46:52] a good and easy way to target is something small  and repetitive like accounts payable was for us. [46:58] There's lots of small and repetitive task. But I  want to end by saying don't just bank the savings [47:03] of time and and leave it at that because you also  need to have a strategy about what to do with the [47:09] save time in terms of increasing the level of  strategic insight that you generate. Amazing [47:15] conversation guys. Thank you again for the time.  How about a round of applause for our panelists.