Kognitos Launches English-Based AI Engine for Enterprise Finance Audits

Kognitos推出面向企业财务审计的英语AI引擎 当概率性大语言模型缺乏特定企业的业务上下文时，企业财务AI项目往往会失败；这类项目需要确定性执行与可重放的流程谱系，方能经受住审计审查。 针对企业财务AI项目的高失败率，Kognitos通过其“英语即代码”（English as Code）神经符号平台，提供了一种替代概率性大语言模型的方案。通过将确定性符号解释器与捕获人工异常处理的上下文图谱相结合，该系统能够在没有Token浪费或幻觉的情况下执行财务工作流。平台组件（如Exception Workbench、动态Agent Builder和Evidence Vault）确保了免触碰日记账分录与结账等流程的完全可审计性。演讲最后为财务主管提供了一个四支柱框架，以及用于评估潜在AI供应商的五问试金石。 企业财务需要立足于公司特定运营上下文的AI，而非基于互联网训练的智能，以实现审计级的自主运行。 (0:04) Gartner研究预测，由于Token成本、缺乏基于流程的对齐以及风控问题，当前40%的AI项目将在2027年底前失败。 (1:53) 将标准大语言模型直接应用于财务领域会使企业面临四大核心失效风险：幻觉、模型漂移、高复杂度Token成本以及不透明的黑盒操作。 (3:33) Kognitos提出了“英语即代码”，使纯英文标准操作程序（SOP）能够通过专有解释器直接作为代码执行。 (6:07) 该神经符号解释器以零Token确定性地处理标准工作流，同时在应对边缘情况时调用神经引擎和人工操作员。 (6:07)

According to Kognitos executives Binny Gill and Amit Gupta, deploying generic AI models in enterprise finance creates severe audit and operational risks because probabilistic systems lack company-specific context and lineage traceability. They argue that achieving audit-proof autonomy requires a neuro-symbolic architecture that executes plain-English standard operating procedures deterministically while capturing institutional knowledge in a context graph. Pitfalls of Generic LLMs in Financial Workflows (1:53) Citing Gartner projections that 40% of enterprise AI projects will fail by the end of 2027, Gill identifies uncontrolled token costs, poor alignment with finance workflows, and unprovable data lineage as primary operational bottlenecks. In enterprise accounting, where entries face strict auditor scrutiny, relying on unverified probabilistic outputs risks regulatory non-compliance and job loss for finance leaders. Deploying standard large language models directly into financial operations exposes firms to four specific failure points: hallucinations, model drift that alters reasoning behavior over time, escalating token expenses caused by circular reasoning loops on complex queries, and opaque black-box outputs that cannot be reconciled against human-readable Standard Operating Procedures (SOPs). Deterministic Neuro-Symbolic Execution via English as Code (6:07) To address the fragility of custom scripts and black-box AI, Kognitos advocates for an "English as Code" methodology that executes plain-English SOPs through a proprietary interpreter rather than directly prompting an LLM. Standard, unambiguous workflows follow a symbolic execution path that runs deterministically like Python, eliminating hallucinations and unnecessary token usage during routine processing. When ambiguities or edge cases arise, the system engages a neural engine and routes the exception to a human-in-the-loop operator. The operator's resolution is recorded as tribal knowledge and integrated into the system, steadily broadening automated capabilities over time while maintaining an Evidence Vault for total audit replayability. Preserving Institutional Knowledge Through Context Graphs (7:37) Gill contrasts traditional process mining tools, which map out rigid workflows that degrade as corporate rules evolve, with dynamic Context Graphs. Connecting directly to enterprise ERPs and communication channels such as SAP, Oracle, NetSuite, Outlook, and Excel, the graph maps operational entities and boundaries within days of deployment. By executing baseline happy-path volume automatically and routing remaining edge cases to staff, the system institutionalizes operational memory. This structure ensures that critical accounting knowledge remains preserved within the organization even when key personnel depart, supported by versioned historical records to verify past reporting methodologies. Operational Governance and Vendor Auditability Criteria (14:45) Gupta outlines four mandatory principles for corporate finance leaders adopting AI: zero tolerance for fabricated data, strict preservation of specific business context, reproducible audit receipts showing identical outputs for identical inputs, and measurable financial return. Decoupling workflow logic into English SOPs supported by an underlying graph creates fungible operational capacity, reducing organizational dependence on rigid IT scripts or external service providers. To separate viable technology from unproven claims, Gupta recommends a five-point evaluation for AI vendors: whether auditors can replay every step of every run, whether identical inputs consistently yield identical outputs, who owns the underlying process definitions upon contract termination, whether system autonomy expands past the initial deployment, and whether the vendor will place fees at risk against validated savings. While Kognitos positions neuro-symbolic execution and Context Graphs as the solution to enterprise audit requirements, the model relies on consistent human-in-the-loop intervention to resolve edge cases and train the system. Long-term operational success depends on whether finance teams can maintain disciplined governance over continuously evolving English-language rules without introducing new operational errors into core ERP ledgers.

Kognitos 创始人 Binny Gill 与副总裁 Amit Gupta 指出，通用大语言模型因缺乏组织特有上下文与确定性验证机制，难以应对严苛的企业财务审计；必须通过“以英语为代码”的神经符号架构与上下文图谱，才能在保证可审计性与数据血统的前提下实现端到端自动化。 通用大模型在财务审计下的失效根源 (1:53) Binny Gill 引用 Gartner 的预测指出，到 2027 年底将有 40% 的企业 AI 项目面临失败。在财务领域，阻碍落地的核心瓶颈集中在不受控的 Token 成本、缺乏业务流程对齐的“业余试验”，以及缺乏数据血统和风险控制。在严格的审计监管下，财务高管无法仅凭“结果来自大模型”作为账目依据，每一笔分录和决策都必须在数月后依然具备可溯源的推导链条。 演讲者进一步归纳了标准大模型在审计环境中的四重结构性失效：首先是数据幻觉，财务核算对数字编造持零容忍态度；其次是模型漂移，供应商全年的模型升级会导致推理行为发生不可控变化，等同于人员流动带来的失范；再次是复杂推理陷入循环而导致的成本与错误“双重惩罚”；最后是黑盒机制无法提供符合人类可读标准的英文标准作业程序（SOP），无法通过外部审计师的抽样核查。 以英语为代码与神经符号执行机制 (6:07) 为替代维护成本高昂的 Python 自定义脚本与不透明黑盒模型，Kognitos 提出了“以英语为代码”（English as Code）的解决方案，利用专用解释器而非大模型直接解析执行以自然语言书写的标准操作规程。该系统采用确定性的符号执行层，在常规顺畅流程（Happy Path）中以零 Token 消耗、零幻觉的方式严格按照预设步骤运行，确保计算过程与结果的绝对可复现。 当系统遇到业务歧义或边缘异常情况时，架构会自动调度神经推理引擎并接入人工介入（Human-in-the-Loop）机制。业务人员在例外处理工作台做出的决策与操作会被系统实时捕获，沉淀为组织内部的隐性知识并更新至流程逻辑中，从而在避免重复犯错的同时持续扩充系统的自主执行边界。 上下文图谱与组织隐性知识沉淀 (7:37) 不同于传统流程挖掘（如 Celonis）仅映射复杂混乱的固定流程切片并导致系统随业务规则演化而僵化，Kognitos 通过直接对接 SAP、Oracle、NetSuite 等 ERP 系统以及 Outlook 和 Excel 构建企业上下文图谱。在系统上线初期，确定性引擎直接自动化约 80% 的常规业务，其余 20% 的边缘长尾案例交由人工处理，系统通过观察异常处理过程将新规则反哺回上下文图谱中，逐步提升自动化覆盖率。 该技术框架依托五大核心支柱：执行自然语言规程的“以英语为代码”、兼顾精准度与灵活性的神经符号执行、防止关键人员离职导致知识流失的上下文图谱与隐性知识学习、为审计提供完整推演证据链的证据凭证库（Evidence Vault），以及记录历史流程定义的版本化流程机制。该平台还提供超过 200 个预置代理模板，并支持直接将遗留的 RPA 工作流（如 UiPath、Blue Prism）或视频会议记录解析为可审计的自然语言代码。 财务合规底线与采购技术的检验标准 (14:45) 特许会计师 Amit Gupta 提出，企业首席财务官（CFO）与财务主计长对 AI 技术的引入存在四项不可妥协的底线：杜绝任何形式的数据编造、保留企业独特的商业上下文、保留完整的复核凭证（确保相同输入必然产生相同输出），以及确保投资带来可量化的实际回报。通过将流程逻辑与底层 IT 实现解耦，财务部门可获得不依赖特定外包商或 IT 团队的可替代运营产能（Fungible Capacity）。 为帮助财务决策者甄别外部方案，Gupta 提出了评估 AI 供应商的五项试金石问题：审计师能否逐一重放每次运行的每一步骤；完全相同的输入是否每次都产生完全相同的输出；终止合作后业务流程定义的知识产权归属何方；系统自主性是否能随时间自主演进还是依赖外包开发人员维护；以及供应商是否愿意根据经核实的成本节约承担基于效果的收费风险。 本场演讲强调了将业务上下文与确定性执行融入企业级 AI 的必要性，主张通过符号逻辑与审计凭证库化解大模型的幻觉与漂移风险；但演讲内容主要基于 Kognitos 自身的产品架构与商业论述，其在极端非标异构数据环境下的实际人机协作成本及长期维护难度仍有待不同企业的实证检验。

## Background 本内容由频道 Ai4 于 2026 年 8 月 20 日发布在 YouTube 平台上。视频记录了 Kognitos 首席执行官 Binny Gill 与副总裁 Amit Gupta 的演讲，探讨了自主财务 AI 面临的审计与信任挑战，并介绍了基于确定性执行与上下文图谱（Context Graph）构建审计级财务智能体的实战方案。 ## TL;DR 演讲指出财务工作的本质是复杂的“异常处理”，而单纯依赖概率性大语言模型将导致不可控的合规风险与审计失败。为此，Kognitos 提出了以神经符号架构为底座的“以英语为代码”（English as Code）范式，通过上下文图谱沉淀企业的隐性业务知识，并确保相同输入下每一次执行的确定性。拥有 23 年经验的资深注册会计师 Amit Gupta 进一步强调了 CFO 面对突击审计时不可妥协的底线，并为财务决策者提供了评估 AI 供应商技术硬实力与商业契约的五大核心质询标准。 ## Atmosphere & style 整场演讲兼具技术创新的前瞻锐气与财务合规的极度克制，开场以系统卡顿切换备用机的小插曲轻松破题，瞬间拉近了与观众的距离。Binny Gill 的技术叙事严密犀利，直击行业“代币失控”与“黑盒模型”的虚浮痛点；而资深注册会计师 Amit Gupta 的登台则带来了极强的行业沉淀与业务代入感，“别让我去坐牢”等金句直白有力，让整场分享在务实、专业且充满警示意味的交织中直击企业落地要害。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=4s) 财务工作的本质是异常处理，人类依赖海量业务上下文做出决策，缺乏上下文的 AI 智能在财务场景中极度危险。 - [01:53](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=113s) Gartner 预警到 2027 年将有 40% 的 AI 项目走向失败，核心诱因在于 Token 成本失控、业务价值模糊以及无法满足审计追溯要求。 - [03:33](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=213s) 财务部署大模型面临四大致命隐患：幻觉与审计信任危机、处理复杂任务时的代币与错误双重打击、不可预测的荒谬异常以及不可解释的黑盒输出。 - [06:07](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=367s) Kognitos 提出“以英语为代码（English as Code）”新范式，利用专利解释器以确定性方式执行英文 SOP，并借助神经引擎构建沉淀隐性知识的“上下文图谱”。 - [08:17](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=497s) 传统硬编码自动化会随业务变化导致效率逐月衰减，而 Kognitos 采用上线先自动化 80% 常规流程、再从人工处理 20% 异常中学习的自适应进化机制。 - [09:25](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=565s) Kognitos 依托五大支柱构建审计级神经符号架构：英语即代码、常识推理与确定性符号执行分离、上下文图谱、证据保险库及业务流程版本化控制。 - [11:27](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=687s) 产品演示展示了上下文图谱的自动连接构建、200+ 预置与 SOP 动态生成的智能体、人机协同工作台以及企业级治理驾驶舱。 - [14:45](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=885s) 从 23 年资深 CPA 视角出发，CFO 对 AI 落地有四大不可妥协的底线：绝不捏造、绝不丢失业务上下文、保留所有凭证（相同输入必出相同输出）以及确保业务结果确定性。 - [16:22](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=982s) 通过将遗留 RPA 脚本无缝转化为纯英文流程资产，企业不仅能避免供应商技术锁定，还能将外部服务商与内部开发资源转化为标准化、可随时替代的产能。 - [18:08](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=1088s) 检验财务 AI 供应商的五大关键质询：能否完整单步回放审计、能否保证确定性输出、流程资产是否以英文归属客户、系统能否自主进化以及供应商是否敢于按实际业务成效挂钩收费。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=4s) 财务的核心是异常处理：缺乏上下文的智能是危险的 演讲在一场轻松的小插曲中拉开帷幕。Binny Gill 笑着向台下寻找分会场的观众打招呼，并幽默地指出现场系统刚刚上演了一出经典的“异常处理”（Exception Handling）——由于演示系统的下载进度缓慢，他们立刻启用了备用笔记本电脑来运行 Demo。他顺势切入主题：在财务领域，人类员工每天所做的工作本质上全都是围绕着“异常处理”展开的。无论是 SAP 还是 Oracle，现有的企业系统永远无法涵盖所有业务状况，这就是为什么必须依赖人类，因为“人类是极佳的异常处理机器”，我们的大脑中沉淀着海量的业务上下文。 Binny 强调，拥有人工智能带来的“智能”固然很好，但“缺乏上下文的智能是极度危险的”（Intelligence without context is dangerous）。AI 仅靠在公开互联网上预训练的模型远远不够，它必须理解特定企业的专属运作逻辑和上下文，才能真正在财务场景中创造价值。这场演讲聚焦于大型企业的落地实战而非空谈理论，Binny 还特别介绍了同台的搭档 Amit Gupta——一位拥有超过 23 年经验的资深注册会计师（Chartered Accountant），他将从实际业务落地与变革管理的角度解析如何让 AI 真正达到审计级标准。 ### [01:53](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=113s) Gartner 预警：40% 的 AI 项目为何会走向失败？ Binny 引述了 Gartner 的权威调研数据：到 2027 年底，当前启动的 AI 项目中有高达 40% 将以失败告终。他深入剖析了导致项目折戟的三大核心诱因： ```mermaid flowchart LR accTitle: AI项目失败的三大核心原因 subgraph FailReasons["AI 项目失败根源"] A["高昂的代币成本<br/>(Tokenomics 失控)"] B["模糊的业务价值<br/>(非痛点自嗨项目)"] C["未知的合规风险<br/>(无法通过审计)"] end FailReasons --> D["40% AI 项目在 2027 年前失败"] ``` 第一是成本失控。行业内常谈论“代币经济学”（Tokenomics），现实情况是模型的 Token 消耗极其昂贵且用量激增。Kognitos 通过更聪明的 Token 调用机制将运行成本直接削减了 90%。第二是业务价值模糊。许多自动化项目沦为开发者的“周末宠物项目”，做完后自我感觉良好，但两个月后业务部门根本不用；企业必须转变为“基于流程”的思考模式，精准解决财务团队真正的痛点环节。第三也是最关键的风险控制。许多人在测试环境（PoC）中看到惊艳的效果，却不敢将其推向生产环境，核心担忧在于：“如果审计人员进场质问‘这个数字从何而来’，而我能给出的最好回答是‘它来自一个大语言模型（LLM）’，那在财务部门我是会被开除的。”财务需要的是每一笔数据、每一个决策都具备可追溯、可证明的数据血统（Lineage）。 ### [03:33](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=213s) 财务智能体（Agent）的四大致命隐患 探讨当前企业部署 AI 智能体时面临的困境，Binny 总结了横亘在财务与 AI 之间的四大现实鸿沟： 1. **幻觉与审计信任危机**：审计制度诞生的初衷，正是因为人类经手财务时可能会犯错甚至舞弊；如果连人类都无法被盲目信任，我们就更不可能信任一个犯了错只需说声“对不起”、既不会被解雇也不会坐牢的 AI。尤其是在过去一年中企业可能切换了不同模型，导致 1 月与 7 月的决策逻辑截然不同，这种不可预测的学习行为在审计眼中是完全不可接受的。 2. **复杂任务引发的双重打击（Double Whammy）**：当 LLM 面临简单问题时消耗代币较少，但一旦遇到复杂业务决策，它往往会在逻辑圈子里打转，不仅更容易做出错误决策，还会伴随巨额的 Token 账单。 3. **不可预测的 AI 异常**：业界普遍认为 AI 处理 98% 的常规流程、人工兜底 2% 的异常即可，但在财务世界里，那 2% 的异常才是真正的核心所在。人类员工犯错时你可以问清缘由，而 AI 抛出的异常往往荒谬绝伦且无人知晓其背后的逻辑。 4. **拒绝黑盒输出**：审计人员绝不接受黑盒模型。当他们进场审查供应商入库或发票处理时，企业必须拿出通俗易懂的英文标准作业程序（SOP）和严格对应的真实案例，而不是一句“我们训练了一个随着时间变聪明的模型”。 ### [06:07](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=367s) 从 Python 代码泥潭到“以英语为代码”（English as Code） 为了解决智能体构建的难题，企业通常有两种老路：一种是构建无人知晓内部运作的黑盒模型；另一种则是用 Python 等传统代码编写长达 1400 行的工作流脚本。Binny 指出，这类脚本往往由多名工程师编写，随着人员离职和遗留注释中的“TODO”，代码维护成本会呈指数级上升。反观人类制定的 SOP，由于所有异常处理都在人脑及其上下文中完成，它并不会随着时间推移变得愈发昂贵。 Kognitos 提出了一种截然不同的范式——“以英语为代码”（English as Code），将常规流程（Happy Path）与异常上下文（Unhappy Path）完整捕获。系统内置了专利的英语解释器，它不是一个会产生幻觉的大模型，而是像 Python 解释器执行 Python 代码一样，严格、确定性地一步步执行英语 SOP。一旦遇到歧义或未定义的异常，神经引擎便会介入并呼叫人类参与决策，同时将该次处理沉淀为企业的“隐性知识”（Tribal Knowledge），进而构建起一张包含企业全部边缘案例与人工审批经验的“上下文图谱”（Context Graph）。 ### [08:17](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=497s) 传统流程挖掘的僵局与 Kognitos 的自适应进化机制 在谈及自动化实施路径时，Binny 对比了传统自动化路线与 Kognitos 模式的根本差异。传统方案（如借助 Celonis 等流程挖掘工具）试图在第一天就梳理出企业所有人员操作的完整复杂流程图，并将其交给 IT 进行硬编码开发。这种做法不仅交付周期漫长，而且一旦将部分子集固化为代码，系统便彻底失去了进化能力；随着实际业务环境的变化，第一天获得的自动化效率在第 30 天就会逐步衰减。 ```mermaid flowchart TD accTitle: 传统自动化与 Kognitos 自适应模式对比 subgraph Legacy["传统自动化路线 (Brittle & Frozen)"] L1["流程挖掘展示混乱业务流"] --> L2["IT 团队硬编码自动化"] L2 --> L3["系统固化，效率随业务变化逐月衰减"] end subgraph Kognitos["Kognitos 自适应模式 (Adaptive Learning)"] K1["第 1 天自动化 80% 常规流程"] --> K2["20% 异常由人工处理"] K2 --> K3["系统观察人工行为并回流图谱"] K3 --> K4["以英文形式持续进化，越用越聪明"] end ``` Kognitos 采取了敏捷务实的策略：在项目上线第一天，先自动化 80% 的常规流程（Happy Path）；剩下的 20% 边缘异常直接交由业务人员手动处理。在此过程中，Kognitos 会持续观察并学习人类的应对方式，将这些经验动态回流到上下文图谱或标准流程中。整个系统的进化完全以人类可读的英文呈现，使自动化率随着时间推移自然攀升。 ### [09:25](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=565s) Kognitos 的五大支柱与神经符号架构 Binny 总结了 Kognitos 支撑“审计级 AI”的五大技术支柱，并阐述了 Gartner 所定义的“神经符号执行”（Neuro-symbolic Execution）机制： - **以英语为代码（English as Code）**：企业历来依靠英文 SOP 运作，现在 AI 同样可以通过纯英语进行精准调度与管控。 - **神经符号执行（Neuro-symbolic Execution）**：架构由两层构成——“神经层”（Neural Layer）负责与人类协同进行常识推理，“符号层”（Symbolic Layer）负责严格的确定性执行，绝不浪费多余代币，彻底杜绝幻觉。 - **上下文图谱与持续学习（Context Graph & Learning）**：完整捕捉企业的隐性知识，即使像 John 或 Vanessa 这样的老员工离职，他们沉淀在大脑中的业务上下文依然能留在企业内部。 - **证据保险库（Evidence Vault）**：专为满足严苛审计设计。无论审计员指向任何一笔交易或决策，系统都能秒级回溯并精准解释“为什么会这样操作”。 - **流程版本化控制（Process Versioning）**：支持像 Git 一样进行业务版本回溯，企业可以清晰查阅并出具文档说明去年与今年的税务申报流程差异（Delta Changes）。 Binny 回顾道，团队早在 2020 年获得投资时就致力于将英语作为自动化引擎而非简单提示词（Prompt）进行研发，目前这套平台已在多家财富 500 强企业中大规模平稳运行，深度处理文档密集和人力密集型业务。 ### [11:27](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=687s) 产品实操演示：上下文图谱、200+ 智能体与人机协作工作台 现场视频演示直观展示了 Kognitos 的运行界面与核心功能模块： 1. **上下文图谱自动构建**：只需在系统后台配置连接到企业的 SAP、Oracle、NetSuite 等 ERP 或邮箱、Excel 数据源，系统便会在数天内自动抓取业务实体并建立关联信号，勾勒出 AI 决策所需的“周边上下文网络”。 2. **预置与动态智能体库**：平台出厂自带 200 多个开箱即用的预置智能体模板；若企业有专属需求，只需直接向智能体构建器（Agent Builder）输入 Word 文档格式的 SOP，甚至是员工培训交接的 Zoom 录屏转录文本，系统就会通过交互问答快速生成专属智能体。 3. **人类协作工作台（Human-in-the-Loop Workbench）**：智能体执行过程中遇到的异常审批、潜在欺诈预警或特定邮件回复请求，都会统一汇聚至工作台，供业务人员以原有的操作习惯进行审核；所有拟过账的日记账分录（Journal Entries）都会先暂存（Staged）供人工复核。 4. **企业级 AI 治理与安全驾驶舱（Cockpit）**：内置基于角色的访问控制（RBAC）和职责分离（SoD）机制；管理层可在驾驶舱仪表盘中清晰追踪“订单到收款”（O2C）、“采购到付款”（P2P）全流程的自动化收益指标，所有业务资产均以纯英语沉淀，绝非不可解释的模型黑盒权重。 ### [14:45](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=885s) 23 年资深注册会计师的视角：CFO 不可妥协的四大底线 曾先后在 Wipro、Accenture 和 Genpact 执业 23 年、帮助众多跨国客户搭建全球商业服务（GBS）和卓越中心（COE）的注册会计师 Amit Gupta 登台分享。他指出，当前企业运营正处于向“AI 赋能与增强”转型的历史拐点。针对全球 CFO 和财务总监关切的核心痛点，Amit 提出了财务管理者不可妥协的四大原则（Four Non-negotiables）： - **绝不捏造，一次都不行（Don't Fabricate, Not Even Once）**：“在财务工作中，猜测根本没有容身之地。” - **绝不丢失业务上下文（Don't Lose Context）**：企业保持卓越运营与核心竞争力的根本在于深厚的业务细微差异（Nuance），自动化绝不意味着简单套用现成标准模版。 - **保留所有凭证（Keep the Receipts, All of Them）**：面对突击审计，系统必须做到“相同输入必出相同输出”（Same Input, Same Output），确保证据链无懈可击，“尽一切努力别让我去坐牢”。 - **确保业务结果（Guarantee Outcomes）**：技术必须提供确定性的结果保障，而非概率性的猜测。 ### [16:22](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=982s) 资产与能力沉淀：从遗留 RPA 迁移到可替代的智能体平面 Amit 深入探讨了企业如何摆脱服务商与传统工具的绑架。依托 Kognitos 的专利解释器与智能体构建器，企业不仅能够吸收标准的 SOP 和会议记录，还能将过去在 Blue Prism、Automation Anywhere、UiPath 等遗留 RPA 工具中沉淀的代码资产无缝解析并转化为“以英语为代码”的标准化资产。这意味着，即便客户未来不再使用 Kognitos 平台运行自动化，企业也依然保留了纯英文形式的全部流程定义资产，彻底告别技术锁定。 在“智能体平面”（Agentic Plane）上，企业掌握着自身运行机制的核心密钥。传统的外部服务商、内部 COE 基地或 RPA 开发资源将转化为标准化的“可互换产能”（Fungible Capacity），企业可以按需计量并随时替换，不再受制于外包团队的漫天要价。平台出厂自带的 215 个流程线框模版（Process Wireframes）与 127 种预置异常及反欺诈检测信号，能让大型企业在短短数周内完成上线部署，而无需经历漫长的等待周期。 ### [18:08](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=1088s) 检验财务 AI 供应商的五大关键质询 在演讲尾声，Amit 向在场所有的财务决策者抛出了五道至关重要的质询题。他建议无论企业未来是否选择与 Kognitos 合作，只要有供应商宣称要在财务部门落地 AI，就必须要求对方正面回答以下五个硬核问题： | 评估维度 | 核心质询问题（Five Critical Questions） | 财务业务要求 | |:--- |:--- |:--- | | **审计复现能力** | 审计人员能否对系统执行的每一步操作进行完整回放？ | 必须支持单步精确追溯与复盘 | | **确定性执行** | 相同的业务输入是否每一次都能保证输出绝对相同的结果？ | 彻底杜绝大模型概率性随机输出 | | **流程资产归属** | 如果双方终止合作，这些沉淀下来的业务流程定义究竟归谁所有？ | 必须以英文资产形式完整归属客户自身 | | **自主进化机制** | 系统的自动化程度是动态持续提升，还是依赖 RPA 工程师手动修改维护？ | 需具备从日常人工异常处理中学习的能力 | | **商业风险共担** | 供应商是否愿意将自身的服务费用与经核实的实际成本节约直接挂钩？ | 供应商敢于为最终业务成效承担商业风险 | Amit 总结道，只有当这五个问题的答案全部为“是”时，企业才能放心地把账本交给 AI。最后，他热情邀请参会者前往 1213 号展位，现场基于企业自身的真实业务上下文与流程，即时构建可运行的实时 Demo。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=4s) 人类是极佳的异常处理机器 — Binny Gill > [00:04](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=4s) 缺乏上下文的智能是极度危险的（Intelligence without context is dangerous） — Binny Gill > [01:53](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=113s) 如果审计人员进场质问‘这个数字从何而来’，而我能给出的最好回答是‘它来自一个大语言模型（LLM）’，那在财务部门我是会被开除的。 — Binny Gill > [14:45](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=885s) 绝不捏造，一次都不行（Don't Fabricate, Not Even Once） — Amit Gupta > [14:45](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=885s) 在财务工作中，猜测根本没有容身之地。 — Amit Gupta > [14:45](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=885s) 绝不丢失业务上下文（Don't Lose Context） — Amit Gupta > [14:45](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=885s) 保留所有凭证（Keep the Receipts, All of Them） — Amit Gupta > [14:45](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=885s) 相同输入必出相同输出（Same Input, Same Output） — Amit Gupta > [14:45](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=885s) 尽一切努力别让我去坐牢 — Amit Gupta > [14:45](https://www.youtube.com/watch?v=PxC2T0_CS_w&t=885s) 确保业务结果（Guarantee Outcomes） — Amit Gupta ## Mentioned Binny Gill, Amit Gupta, Kognitos, Ai4, Gartner, SAP, Oracle, NetSuite, Celonis, Blue Prism, Automation Anywhere, UiPath, Wipro, Accenture, Genpact, Zoom, Python, Git, Word, Excel ## Video description Binny Gill - CEO, Kognitos Amit Gupta - Vice President, Kognitos Autonomous finance is the promise of the decade, but here's what nobody pitching it will tell you: an AI agent that guesses, even brilliantly, will never survive your audit. Boards have stopped asking whether finance uses AI and started asking why the pilots haven't moved the needle. The answer is almost always trust. In this talk, Binny Gill, founder and CEO of Kognitos, shows what audit-grade autonomy actually requires: deterministic execution instead of probabilistic guessing, and step-by-step replay of every action an agent takes. He'll make the case that when it comes to trusting AI with your ledger, context is all you need: a context graph that grounds every agent decision in your vendors, your policies, and your history, the approach Gartner recognized Kognitos for in two 2026 Hype Cycles. He'll demonstrate it live on the workflows finance leaders care about most: touchless journal entries, zero-touch close, and AP fraud detection. You'll leave with the six-step playbook for taking your first agent from pilot to production, and the bar to hold every AI vendor to. [Source](https://www.youtube.com/watch?v=PxC2T0_CS_w)

[0:04] Binny Gill: 人类是极佳的异常处理机器 [0:04] Binny Gill: 缺乏上下文的智能是极度危险的（Intelligence without context is dangerous） [1:53] Binny Gill: 如果审计人员进场质问‘这个数字从何而来’，而我能给出的最好回答是‘它来自一个大语言模型（LLM）’，那在财务部门我是会被开除的。 [14:45] Amit Gupta: 绝不捏造，一次都不行（Don't Fabricate, Not Even Once） [14:45] Amit Gupta: 在财务工作中，猜测根本没有容身之地。 [14:45] Amit Gupta: 绝不丢失业务上下文（Don't Lose Context） [14:45] Amit Gupta: 保留所有凭证（Keep the Receipts, All of Them） [14:45] Amit Gupta: 相同输入必出相同输出（Same Input, Same Output） [14:45] Amit Gupta: 尽一切努力别让我去坐牢 [14:45] Amit Gupta: 确保业务结果（Guarantee Outcomes）

[0:04] Binny Gill: intelligence without context is dangerous when applied to enterprise finance. [1:53] Binny Gill: 40% of current AI projects are slated to fail by the end of 2027. [1:53] Binny Gill: it came from an LLM [6:07] Binny Gill: English as Code [14:45] Amit Gupta: Do Not Fabricate [14:45] Amit Gupta: Preserve Business Context [14:45] Amit Gupta: Keep All Receipts (Auditability) [14:45] Amit Gupta: Guarantee Outcomes [17:21] Amit Gupta: Can your auditor replay every step of every run? [17:21] Amit Gupta: Does the exact same input generate the exact same output every time? [17:21] Amit Gupta: Who owns the process definition if we stop working together? [17:21] Amit Gupta: Will the system's autonomy improve over time, or is it capped once the developer leaves the room? [17:21] Amit Gupta: Will you put your fees at risk based on validated savings? [17:21] Amit Gupta: If the answer to all four or five of them is yes, then yes, please go ahead. Otherwise, come over to Booth 1213 and we'll show you a demo live.

[0:04] Hey everyone, uh welcome. And if you're looking  for the palazzo room, stay here. This is going to [0:09] be fun for you. So, um as you might have noticed,  we did some exception handling here. I mean, this [0:16] uh system has been downloading our uh demo. It's  been taking a while. What we did classic exception [0:24] handling. We have the backup plan running here.  This is a laptop and we'll run the demo through [0:29] that. In finance all the work that people do is  around exceptions you know what what do you do [0:36] when the system you're running SAP Oracle that's  not doing everything that's why you have humans [0:41] we are great exception handling uh machines right  that's all we do we have a lot of context in our [0:46] head today we going to talk about that context  that is all you need of course peace of mind [0:52] is all you need but the way you get to peace of  mind is first you get intelligence now you have [0:57] intelligence with AI Okay, intelligence without  context is dangerous. So you need context so that [1:04] intelligence knows okay this is the context this  is how this company works not all that I've been [1:09] trained on on the open internet this company  is different so I need the context after that [1:15] you can go do something meaningful in finance  today what we're going to talk about is uh not [1:20] theory but actually what we have implemented in  large companies um bringing AI in finance that [1:26] is audit grade and I'll talk about that with me I  have Amit who has 23 plus years experience being [1:33] a chartered accountant and um he's going to tell  us how AI can be deployed you know what are the [1:41] practices that you need to go through to actually  deploy it it's not just tech it's also change [1:46] management you have to bring it in your business  um so he'll be with me on this talk first thing [1:53] is to understand what is the industry seeing  this is Gartner slides they're talking about [2:01] 40% of the projects are going to fail um by the  end of 2027 that were AI projects right now. Why [2:08] is that? There are three reasons. One is the cost.  Sometimes, you know, we we talk about tokconomics, [2:16] but really the real challenge is that tokens are  too expensive and we are burning more and more [2:20] of them. How do you deal with that? We're going  to talk about that. We have brought down costs [2:25] by 90% by being smart about how to use tokens.  Number two is unclear value. Somebody did some [2:32] automation. It was a pet project. They felt happy  over a weekend and 2 months later people are not [2:38] using it. So that part is important. Let's talk  about that. Why a a processbased thinking has [2:45] to be brought in. It's not tech based. The goal  is to not deploy tech. The goal is to solve real [2:50] problems that the hotspots in in your finance,  you know, people are working on that. And finally, [2:56] it's about the risk controls. like I built  something but I'm worried. So it's it's in P. [3:03] It's working beautifully but I'm worried if I put  it in production what will the auditor say? Will I [3:07] lose my job? If somebody shows up and says okay  this number where did it come from? If the best [3:12] answer I have is it came from an LLM I'm going to  lose my job if I'm in finance. That is something [3:18] we are talking about how every data every every  decision has a lineage that is provable. Even [3:25] months later you can say this is where it came  from. Now in terms of agents there are four [3:33] problems that are happening there. One is LLMs  can hallucinate. We all know that but we don't [3:41] want a number to be just coming out of thin air  right that's the big problem. That is why it won't [3:47] survive an audit. You know why audit was created?  When humans run finance still other humans come [3:53] and say I want to audit you because you could  have done a mistake or actually done fraud, [3:59] right? So I want to make sure. So we don't trust  even humans. Now imagine AI which doesn't have [4:04] any problems about making mistakes. It just has to  say I'm sorry and carry on. Doesn't lose its job. [4:09] Doesn't get fired. Doesn't go to jail for fraud.  Nothing like that. Right? So you need audit. [4:14] Auditors come in and look at AI stuff and say  like what? Over the last year, you had different [4:20] models running. It's almost as if you had churn in  your business. Every model thinks differently. And [4:25] you're telling me an AI system is learning. That  means in January the behavior is different than [4:29] in July. All of that is unoduditable. So that's  the challenge. We got to figure it out. One thing [4:36] we have noticed is that if you're using LLMs to  do decision making, when it's a simple problem, [4:43] it burns less tokens. When it's a complex problem,  that's when it goes into C circles and kind of [4:49] makes a wrong decision a lot of the time. So, not  only it makes a wrong decision, it also gives you [4:54] a large bill of tokens when it is actually giving  you the wrong answer. It's the worst. It's a [5:00] double whammy. That's what we see is a problem.  Third thing is a lot of the AI thinking is okay, [5:07] it does 98%, so 2% exceptions you can handle. The  challenge is no that those 2% is the real problem. [5:14] Today in finance we are doing all all sorts of  things that are exception handling. The exceptions [5:19] from AI are the worst kind of exceptions. Nobody  knows why it said that. At least if it's a human [5:24] I can ask like why did you do it? AI is just  going to give some bad answers. That's the one [5:29] uh we need to solve. And finally auditors do not  want a black box. They want to come to a finance [5:37] department says okay how do you do your vendor  onboarding? How do you do your invoice processing? [5:42] You should show a standard process in a language  that they understand English. And then the audit [5:48] says okay show me three examples of this. You  should be able to show and that should map to [5:52] this. It should not be that I have trained a model  and over time it has become smart. That's not [5:57] audited. Okay. Now how do you build these agents?  There are two ways. One is you can have somebody [6:07] create a blackbox model. the worst worst way of  doing it. Second way is that okay I create an an [6:14] automation in some workflow language. I mean here  we picked Python 1400 lines three engineers wrote [6:21] it two have left the company there's a todo you  know in a comment there that part is you know over [6:28] time gets more and more complicated no wonder like  you know software projects become more expensive [6:33] over time whereas if you have created a human  instruction that doesn't become more expensive [6:39] over time the same SOP runs the business year  over year the exception handling is done in the [6:45] human's brain the context that they have. So we  have built a system where the happy path as well [6:49] as the context the unhappy path is captured in  the system. The happy path we are showing here [6:55] in English as code. So Cogntos is a system where  you can take English SOPs and run it as if it was [7:03] code step by step. We have an interpreter. It's  not an LLM and therefore it does not listenate. [7:08] Just like Python would read Python instructions,  we have Cogntos English interpreter that reads [7:13] English instructions, runs it steps by step by  step. Anytime there is a doubt or a confusion, [7:19] it zooms out a neural engine comes in and that  pulls in a human and to understand what's the [7:25] next best action and that gets recorded as  tribal knowledge and the next time it can be [7:30] better. Now for the next time getting better,  that's where the context is created. We create [7:37] a context graph. This context graph is a a  collection of what all the humans have told [7:45] the system are the things that you need to do  outside of the standard of the process. So no [7:53] no company runs purely on the standard process,  right? There's all sorts of variations and edge [7:57] cases. The way we do it is you collect connect  to your systems, grab whether it's SAP or your [8:03] Outlook mailbox or bunch of Excels lying around.  capture all of that, create a context graph that [8:09] becomes the input for the agents to figure out how  to handle this edge case with a human approval. [8:17] And the difference between the legacy ways of  doing automation and our way is that the legacy [8:25] way of automation people come and say you know  what we need to do um put in something like a [8:30] salonus or you know some other process mining tool  first let's figure out all the edge cases they'll [8:35] show you a massive spaghetti of process this is  what your humans do now let's give that to it and [8:42] we're going to automate all of that once you've  automated a subset of it you're frozen there. It's [8:48] not going to get smarter over time, you know, and  and that's whatever efficiency you got on day one, [8:54] in day 30 it's going to be slightly lower and so  on because the fidelity of that spaghetti that [8:59] initially you saw is actually also changing.  In our way, we say look even if you handle [9:05] u the happy path on day one, 80% is automated.  The rest of the 20% goes to a human who does it [9:12] manually and then over time our system observes  how they do manually and the system feeds in back [9:18] into the context graph or the standard process  and it's getting better everything in English. So [9:25] these are the five things that cogntos provides  and I'll also show you a video of the features [9:30] there. English is code we have run businesses on  English SOPs are English. Now the time has come [9:37] to run even AI with English. English is a way you  control intelligence at scale whether it's human [9:42] or AI. That's the next standard we are saying.  Neurosy symbolic execution. Um Gartner uses this [9:48] term to explain what we do. There's a neural layer  that works with a human to reason. Symbolic layer [9:54] executes it deterministically. It won't use tokens  or hallucinate. The context graph and learning [10:01] is what is describing how your business runs.  So when John or Vanessa leaves your business, [10:07] whatever they leave with is what we are trying  to capture. So that that stays with the business [10:12] once our system has been running for some time.  Evidence vault, whatever the auditor wants to [10:17] do. They point at anything, you should be able  to say exactly why it happened. And finally, [10:23] the process itself is versioned. So your business  has evolved over time. You can go back in time and [10:28] say, how did we do um this tax reporting last  year? And you should be able to answer that [10:36] um with a document and these are the delta  changes. One thing is we have been working [10:41] on it. We started with uh doing research in this  space um 2020. Um we were funded to create English [10:50] as an automation engine uh rather than English as  prompt. LLMs have come after that. They help ease [10:56] out the burden of handling exceptions, building  the plan, being creative, but anything you run [11:02] deterministically runs through our engine. Um,  it's running at scale right now. So again, this [11:08] is not something new that we are pitching. Um,  Fortune 500 companies are running at scale. I can [11:14] go through these list, but B, you know, anything  around document heavy, people heavy processes, [11:20] um, we have there's some examples here. So you can  take a look. Um let me play this video. So this is [11:27] um showing um what we do in terms of uh creating a  context graph. You can connect to any one of your [11:33] uh systems whether it's SAP, Oracle, Netswuite.  Quickly we connect and we gather the various [11:41] um entities and also create the signals around  that. So this is the context graph. You can see [11:48] it shows essentially what is available for the  AI engines to observe um because that creates [11:57] efficiency and also accuracy. If your AI agent  is just given a open-ended question then it [12:03] cannot figure out because you don't even give  the data the context is needed. Even when you [12:08] hire an employee you give the employee access  to all the data they need to in order to make [12:12] good decisions. This is what it is doing. This  contest graph gets created in the first few days [12:18] of us connecting to your backend systems the ERP  systems and the agent is able to reason about the [12:24] neighborhood around the context you know that you  are working on. So if it's a um you know edge case [12:30] that you're looking at it's going to do that. We  come up with a system where there are 200 plus [12:37] agents already templates provided in our um in  our product that you can refine into what your [12:43] business needs. Those agents run and they create  human in the loop requests on the workbench. So [12:51] this workbench is where people will come and  handle exceptions whether it is an approval [12:56] or looking at some signal of fraud or looking at  uh sending an email to somebody in response to a [13:04] request. Those requests can be handled here by the  same people who would be working on these things [13:10] before automation came in before AI came in. Now  whenever AI makes decisions like you know posting [13:16] certain things in your journal entries those  are staged so that you can review them. There [13:21] is a platform substrate that we are building  for arback access control AI governance making [13:27] making sure things are running properly. It can  manage that and on top of that all these agents [13:33] and SOPs are driven by the business side. So they  know what the business needs. Um these uh roles [13:40] are can be defined custom by your uh company. Um  we also come with some standard roles of you know [13:47] separation of duties and so on. Those things are  available to you. If any agent is not available [13:52] in our template library we have 200 plus. You  can go build agents dynamically here as well. [13:58] You throw in your SOPs, word documents or even  a transcript of a Zoom recording that you did [14:03] with somebody in the company and our agent builder  will look at that, ask you certain questions and [14:08] help you build that. Over time in the cockpit  you can see the benefits that you're getting [14:15] uh in terms of what's happening in order to  cash, procure to pay and these um signals [14:21] will become better over time as we are gathering  more and more about your uh tribal knowledge. So [14:27] everything that we create is finally documented  in English. It's not weights in an AI model that [14:33] you don't understand. So if we leave, we leave  the English behind. And that is precisely how [14:38] your business used to run before we left. Right?  So that's um what we are providing. I'll invite [14:45] uh Amit to talk about how we are taking it to  customers and helping them deploy it. Thank you [14:50] very much. Uh I'm audible, right? All right. So u  I've been a chartered accountant a practitioner in [14:57] the business processes space for about 23 years.  I've worked with companies like Vipro, Accenture, [15:03] Genpack for long period of time help clients like  you build uh your global business services models [15:09] or centers of excellence. Uh now the inflection  point has come where everybody is converging more [15:15] towards how can all of that become AI additive  or augmented by AI and that's the question that [15:21] I'm going to share a little bit of my thoughts  with. what are CFOs asking? Uh what does the [15:26] technology that Ben uh showcased, how does that  support it? And then what's the next step? So [15:31] when you go back from this conference on Monday,  what could be the next step here? Right? So four [15:35] non-negotiables for every CFO controller in the  room. You know, don't fabricate not even once. [15:42] Uh guessing is not part of what we do in finance.  Uh don't lose the context of what we do. all the [15:48] nuance of why we are a phenomenally run business,  why we are competitive is because of the context. [15:53] So automation does not mean bringing stock,  it does not mean just bringing uh things out [16:00] of the box and and using it. Uh keep the receipts  all of them. So if my auditors show up tomorrow, [16:06] would you be able to show everything same input,  same output every time? So do everything to not [16:11] put me in jail and then um can you guarantee the  outcomes? Right? So and the answer for all four [16:16] of them is yes with with cogntos and how I'll  I'll show you. So Bini spoke about English as [16:22] code significantly. You know this is what keeps  you hostage with if you have service providers or [16:28] uh folks who do this business in silos. Uh this  is what we extract out of people because of our [16:34] patented interpreter and the patented agent  builder. We can take any SOP, we can take [16:40] transcripts, we can take your Python code. If you  have RPAs that have been deployed, blue prism, [16:44] automation anywhere, UiPath, all of that can  be ingested without needing technical support. [16:50] That's the beauty of the solution that it can  be ingested. It can be then shown as English as [16:54] code and that's what we leave behind. So even if  you would want to bring us to document everything [16:59] just as pure English and not run the automation  as a platform, that's doable as well. Three four [17:06] things that come out of English as code. just that  now more people in the organization or subject [17:11] matter experts understand it but also on uh uh can  you use it to then negotiate right uh and other [17:18] things so I in the interest of time I'll quickly  move um so the agentic plane right this is what is [17:24] most important is that you use English as code the  context graph you keep the keys with you of what [17:30] you're running and then the capacity becomes  fungeible so if you have a service provider [17:33] today or your COE your captive center or an RPA  provider that then becomes your funible capacity [17:39] that can be metered and can be replaced at will.  You do not uh you know they're not negotiating [17:44] with you on what needs to happen next. Uh like  Benny said, we've seeded the platform with 215 [17:51] process wireframes about 127 signals that detect  suspicious transactions, anomalies, frauds, [17:57] whatnot. So you can be ready within the first few  weeks of deploying this platform, not uh you know [18:02] wait for months. Now I'll just leave you with this  last slide in the interest of time is even if you [18:08] don't want to work with cogntos these are the  five questions that you have to ask of anybody [18:11] who comes in and says I want to deploy AI for  finance in your organization is you know can you [18:17] can your auditor replay every step of every run  can the same input generate same output who owns [18:23] the process definition if you leave I decide not  to work with you uh who owns those process areas [18:28] and then will the autonomy improve over time or  am I locking myself into a process uh a developer [18:34] or a RPA developer once he leaves out of the room  that's where I I leave the autonomy at and will [18:41] you put your fees at risk at the validated  saving right so if the answer to all four [18:45] or five of them is yes then yes please go ahead  otherwise we are at booth 1213 come over we'll [18:51] show you a demo live your context your process  we'll build it live with you thank you very