Accend CEO Says AI Replaces Manual Credit Spreading by 2027

Accend CEO称AI将于2027年前取代人工信贷财务分拆 采用统一AI信贷操作系统的金融机构将通过消除人工数据录入错误并执行持续实时风险评估，超越依赖孤立单点解决方案的竞争对手。 Pranjal Daga阐释了商业信贷承销在2026年如何迅速从静态、碎片化的单点解决方案转变为持续、统一的AI操作系统。通过利用具备共享上下文和确定性验证闭环的智能体架构取代多达十个孤立工具及人工跨系统整合，机构消除了正负号颠倒及净收入误算等严重的交接错误。Daga提出了用于评估具备防御壁垒信贷技术栈的CVC框架（Context、Verification、Compounding），并预测人工财务分拆将不再作为一个岗位描述而存在。 到2026年，AI驱动的商业信贷决策已从试点全面步入生产环境，金融团队对生产级AI的采用率激增600%，信贷决策周期从数周大幅缩短至数小时或数分钟。 (0:04) 传统商业核保流程依赖多达10个割裂的点工具，人工充当集成层导致严重的交接风险，回测数据显示逾10%的借款人存在财务符号正负倒置错误，更有案例将100万美元净利润误算为7500万美元。 (3:43) 现代AI信贷操作系统通过共享上下文仓库整合全流程，模型本身仅占系统表现的20%，其余80%依赖确定性验证代理、评估回路及精确到单元格的源文件审计追踪。 (7:19) 评估信贷技术栈应遵循CVC框架（上下文Context、验证Verification、复利Compounding），基础大模型由于缺乏机构专有历史和深度工作流整合，无法直接颠覆专业信贷平台。 (14:04) 演讲预测到2027年，手动财报分拆录入（Spreading）将不再作为独立工作岗位存在，信贷分析师将全面转向管理AI系统、配置风险框架与执行高级金融裁决。 (17:19)

Pranjal Daga, Co-Founder and CEO of Accend, argues that commercial underwriting is rapidly moving away from fragmented point solutions glued together by manual human labor, shifting instead toward unified AI credit operating systems that perform continuous, context-aware risk evaluations. Fragility in Legacy Multi-Tool Workflows (3:43) Legacy commercial credit operations routinely rely on up to 10 isolated tools, spanning document intake, OCR readers, financial spreading software, ratio calculators, credit bureaus, loan origination systems, spreadsheets, and emails. Because these systems lack shared context, human credit analysts serve as the sole integration layer. Manual data transfers between disconnected software create severe information drops that introduce material credit risk. To demonstrate this fragility, Daga cites Accend's backtests on real bank loan portfolios. In one migrated portfolio, 48 out of 443 borrower files (over 10%) contained fundamental sign inversions (+/- reversals) that invalidated the underlying credit analysis. In another file, manual spreading errors caused an analyst to miscalculate a borrower's net income as $75 million when the actual verified figure was $1 million. Architecture of an Agentic Credit System (7:19) Modern underwriting platforms replace siloed tools with a unified sequential pipeline where document intake, spreading, credit memo drafting, live credit scoring, and portfolio monitoring share a single context store. The technical stack uses an Orchestrator Agent to break down complex files—such as packages containing more than 40 documents across three fiscal years—into sub-tasks executed by Specialized Agents that parse line items and draft narrative memos. To prevent hallucinations, deterministic Verification Agents reconcile balance sheets, verify cash flows, and recalculate debt service formulas. Daga emphasizes that base model accuracy accounts for only 20% of total system performance, while the remaining 80% depends on deterministic verification checks, evaluation loops, and context management. Source traceability links every calculated output directly to its source document page and spreadsheet cell, with human analysts positioned downstream for exception handling and credit approvals. Real-Time Intelligence Versus Periodic Reviews (12:05) The operating standard for risk management has shifted from periodic reviews to continuous monitoring. While running manual risk checks five times a day represented JPMorgan Chase's gold standard during the 2008 financial crisis, Daga states that by 2026 the baseline has become real-time portfolio surveillance, supported by a 600% increase in finance teams deploying AI into production. This architectural transition compresses credit decision turnaround times from weeks and days down to hours or minutes. Rather than relying on static, historical snapshots from legacy credit bureaus and annual portfolio reviews, modern credit intelligence synthesizes dynamic transaction streams, real-time banking data, and trade payment behavior into continuously updating scores. Institutional Moats and the CVC Framework (14:04) Daga disputes the idea that general foundation models from vendors like OpenAI or Anthropic will make specialized credit software obsolete. While base models commoditize general intelligence, they lack proprietary institutional credit judgment and bank-specific credit committee risk criteria. Consequently, an institution's competitive advantage lies in proprietary feedback loops and auditability. Daga evaluates credit platforms through the CVC Framework: Context (maintaining a unified data store across all underwriting stages to prevent handoff losses), Verification (deterministic audit trails tracking overrides and formula reconciliations to the source), and Compounding (the system's ability to learn from institutional decisions, ensuring higher accuracy on deal 10,000 than on deal 10). Daga predicts that by 2027, financial spreading will cease to exist as a standalone job description, as autonomous agents absorb manual data entry while human professionals transition into managing AI systems, configuring risk parameters, and exercising final credit judgment.

Accend联合创始人兼首席执行官Pranjal Daga指出，商业信贷审批正在从依赖多款分散工具及人工搬运数据的旧模式，演进为由统一上下文和自主智能体驱动的“AI信贷操作系统”。该方案旨在通过全流程集成与确定性验证消除流转误差，并将信贷决策与资产风控由数周周期的静态评估转变为分钟级的全天候连续监控。 传统多工具割裂与人工整合隐患 (3:43) 传统商业信贷决策流程通常依赖多达10种相互孤立的软件工具，涵盖文档摄入、OCR识别、财务报表重述（Spreading）、比率计算、征信数据源、信贷备忘录模板、贷款发放系统（LOS）、电子表格、电子邮件以及外部资产组合管理看板。这些离散工具之间缺乏共享的上下文数据，迫使信贷分析师充当唯一的系统集成桥梁。分析师在紧迫的时限下跨多个屏幕手工转移数据，不仅耗费大量精力，且每一次系统间的数据交接都会造成信息丢失，直接转化为底层信贷风险。 根据Accend对实际银行贷款档案的回测数据，此类人工搬运与重述错误极为普遍。在某家银行迁移的贷款资产组合中，443名借款人中有48名（占比超过10%）的财务数字存在正负符号反转的基础错误，导致原有的信贷分析彻底失效；在另一份信贷档案中，因人工电子表格录入错误，借款人的净利润被严重误算为7500万美元，而核实后的实际真实净利润仅为100万美元。 智能体流水线与确定性验证架构 (7:19) 为消除人工交接风险，新型AI信贷操作系统将文档摄入、财报重述、比率计算、备忘录起草、动态评分与资产组合监控整合至单一流水线，前序输出直接作为后序环节具备上下文感知的输入。该技术栈包含明确的分工层级：编排智能体（Orchestrator Agent）负责将跨越数个财年的海量复杂借贷材料拆解为离散任务；专业智能体（Specialized Agents）协同提取非结构化文本、归类财务科目并草拟备忘录；确定性验证智能体（Verification Agents）则执行硬性校验，核查资产负债表平衡、偿债公式与现金流勾稽关系以拦截模型幻觉；人工审核员（Human-in-the-Loop）则保留在末端负责异常处理与最终裁决。 Daga强调，基础大模型的单点精度仅占整个系统性能的20%，其余80%的表现取决于确定性验证核查、评估闭环以及上下文窗口工程。此外，系统具备精细的源头可追溯性，计算生成的每一个数据指标均能直接反查关联到源文件中的具体页码、行项目及电子表格单元格。 授信与风控流程的全面实时化 (12:05) 信贷底层架构的革新正在重塑各项运营指标。Daga回顾指出，在2008年金融危机期间，摩根大通（JPMorgan Chase）CEO Jamie Dimon将每日执行5次人工风险检查视作保持资产负债表稳固的黄金标准，践行“重防备而非重预测”的理念；而到2026年，在财务团队生产级AI应用激增600%的驱动下，行业基准已从周期性抽检全面转向全天候实时资产监控，授信决策时间亦由原本的数天至数周大幅压缩至数小时乃至数分钟。 在业务流程层面，传统数小时至数天的文档录入与备忘录撰写，转变为由AI数分钟内生成初稿并由分析师审核微调；资产组合审查由静态的年度复核升级为持续动态风险追踪。评估依据也从依赖征信机构单一时点的静态历史快照，转变为融合实时银行账户数据、动态交易流水、商业支付行为以及行业上下文的实时动态信用评估。 机构专有壁垒与CVC评估框架 (14:04) 针对通用大模型提供商（如OpenAI或Anthropic）可能淘汰垂直信贷软件的观点，Daga分析认为，通用模型虽然使基础认知能力商品化，但缺乏金融机构特有的专有上下文，既不了解特定银行信贷委员会的偏好，也无法理解垂直行业的专有会计科目逻辑。因此，金融机构构建AI防御壁垒的核心在于专有的反馈数据积累与源头级可审计性。 Daga提出了评估信贷技术栈的“CVC框架”：第一是上下文（Context），考量平台是否具备端到端共享的统一上下文存储，避免因工具割裂遗失借款人信息；第二是验证（Verification），考量系统是否具备确定性校验闭环与追溯至源文件的自动审计追踪，以记录模型运算与人工干预；第三是复利效应（Compounding），考量平台能否通过持续沉淀信贷人员的专业判断与反馈，使系统在处理第10,000笔交易时比第10笔交易更具智慧。 Daga预测，到2027年，作为独立职能的“财务报表重述”岗位将彻底消失，基础数据录入工作将完全由自主智能体替代，信贷专业人士的核心职责将转为配置风险框架与行使高阶商业判断。不过，该观点来自垂直领域软件供应商，其关于全自动效率提升与近乎零失误的论断建立在理想的单一平台架构之上；对于现有IT架构异构、受强合规约束且依赖遗留工具的大型传统机构而言，跨部门替换迁移的实际阻力与模型黑箱治理等挑战仍有待时间验证。

## Background 该内容由行业峰会频道 Ai4 于 2026 年 8 月 20 日在 YouTube 发布，主讲人为 Accend 联合创始人兼 CEO Pranjal Daga。视频基于真实生产环境数据，探讨了金融机构如何通过统一的 Agent 操作系统取代 10 种相互割裂的信贷单点工具，从而实现端到端商业信贷核保的重塑与复利进化。 ## TL;DR Pranjal 指出，传统信贷流程依赖分析师充当“人肉胶水”拼接多个独立软件，不仅极其低效，更隐藏着灾难性的数据错漏与风控隐患。2026 年的现代金融机构正在转向一体化的“AI 信贷操作系统”，通过协调器、专业化 Agent 与确定性验证循环的协同，将核保周期从数周压缩至数分钟，并实现了 100% 可追溯的审计链条。信贷模型准确率的 80% 取决于工程编排与闭环验证，而非单纯的基础大语言模型能力；评估信贷技术栈的核心在于共享上下文、审计验证和复利进化的“CVC 模型”。最终，人工手动分拆报表的岗位将被全面淘汰，未来十年的行业赢家将属于那些率先构建起统一闭环系统的机构。 ## Atmosphere & style 演讲节奏沉稳干练且极具行业穿透力，主讲人以 2008 年金融危机的历史教训切入，迅速营造出一种深刻的危机感与紧迫感。通过披露真实贷款迁移中触目惊心的财务失真案例，整场分享在冷静的技术架构剖析与激进的行业终局预言之间张弛有度。听众不仅能感受到 Agentic AI 对传统金融运营体系的巨大冲击，更能体会到专业信贷判断力在技术重塑下向更高维度跃迁的必然趋势。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=4s) AI在商业信贷决策中已实现爆发式普及，2026年金融机构在生产环境中部署AI的比例激增600%，信贷审批周期从过去的数周压缩至数小时乃至数分钟。 - [00:04](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=4s) 风险控制标准发生根本性范式转移，从2008年金融危机时期杰米·戴蒙推崇的“每日5次人工核查”升级为全天候的“AI实时持续监控”。 - [03:43](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=223s) 传统信贷流程依赖多达10个割裂工具，并将昂贵的信贷分析师当作脆弱的“人工胶水”，在交接中造成严重信息断层并引发灾难性数据失真（如10%报表数据符号颠倒、净利润虚增75倍）。 - [07:19](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=439s) 一体化“AI信贷操作系统”取代分散工具链，将进件提取、财务分拆、备忘录生成及投后监控无缝整合进统一平台，并提供点击即可反查原始PDF和Excel单元格的穿透式审计能力。 - [07:19](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=439s) 信贷核保AI的可靠性遵循80/20工程架构法则：底层基础大模型仅贡献约20%的准确度，其余80%完全由确定性核查、闭环验证机制与上下文工程决定。 - [07:19](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=439s) 现代信贷架构采用多智能体编排系统（Agentic Stack），通过任务协调器、专业执行Agent与验证对账Agent实现闭环，并遵循“安全边界内AI自主执行，重大商业决策交由人类裁决”的人机协同原则。 - [12:05](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=725s) 融合实时银行流水与动态支付行为的“动态企业信用评分（Live Score）”，正在彻底超越传统征信局数月更新一次的静态快照评分。 - [12:05](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=725s) 通用基础大模型的能力正被快速大宗商品化，垂直信贷平台的核心护城河并非基础算法，而是机构内部专有的反馈数据沉淀与端到端审计追踪能力。 - [15:32](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=932s) Pranjal提出评估信贷操作系统的CVC核心模型：即是否具备打通工具孤岛的共享上下文（Context）、全流程源头级可追溯审计（Verification）、以及随着处理交易量增加而越变越聪明的复利进化能力（Compounding）。 - [16:51](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=1011s) 2026年的“堡垒式信贷系统”不再是单点工具的机械拼凑，而是演化为一个能纵览借款人全貌、穿透检验每个财务数字并伴随每笔新交易持续进化的统一Agent平台。 - [17:37](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=1057s) 传统的手工财务报表分拆/录入（Spreading）岗位将在未来一年内彻底消亡，信贷专业人士的角色将转变为“AI系统的管理者”与高维度商业风险判断者。 - [17:37](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=1057s) 基础模型的普惠让算法独占优势不再存在，未来十年商业信贷领域的最终赢家，必然属于那些率先构建起最佳统一操作系统（Best System）的机构。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=4s) 商业信贷的预测与历史镜鉴：从2008危机到2026全面落地 一年前在同样的讲台上，Pranjal 曾大胆预测：到 2027 年，商业信贷领域几乎所有的决策都将在 AI 的辅助下完成。然而时间来到 2026 年，事实证明当年的预测甚至保守了——商业信贷决策几乎已经全面被某种形式的 AI 所接管，AI 在生产环境中的应用迎来爆发式普及。在过去一年中，金融团队在生产环境中部署 AI 的比例实现了 600% 的惊人增长，信贷审批速度也直接从过去的“数天或数周”大幅缩短至“数小时，甚至在部分场景下仅需数分钟”。 为了理解当前这场范式转移，Pranjal 将视线拉回 2008 年全球金融危机。彼时贝尔斯登（Bear Stearns）化为乌有，雷曼兄弟（Lehman Brothers）轰然倒塌，全美最大的储蓄银行华盛顿互惠银行（Washington Mutual）也难逃崩溃厄运。然而，摩根大通（JPMorgan Chase）却在危机之后变得比以往任何时候都更加强大。摩根大通 CEO 杰米·戴蒙（Jamie Dimon）曾透露，在危机期间乃至危机之前，他们每天都会进行 5 次全方位的风险排查。他们之所以这样做，并非对自己模型的预测能力缺乏信心，而是因为他们坚守一个核心信条——“准备重于预测”（preparation over prediction）。 在 2008 年，每天 5 次人工风险检查被奉为构建“堡垒式资产负债表”（fortress balance sheet）的黄金标准；但在 2026 年的今天，支撑这套旧堡垒的世界已经彻底瓦解，新的行业基准已经演化为“全天候持续监控（Continuous Monitoring）”。 - 2026 年金融机构在生产环境中应用 AI 的比例激增 600%。 - 信贷决策周期从昔日的“数周/数天”跨越式压缩至“数小时/数分钟”。 - 风险控制的核心理念从 2008 年的“每天 5 次人工核查”升级为 2026 年的“AI 实时持续监控”。 --- ### [03:43](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=223s) 传统信贷的痛点：10个割裂工具与“以分析师为胶水”的巨大风险 审视目前大部分机构的商业信贷决策流程，会发现一个由多达 10 种割裂工具拼接而成的混乱工作流：首先是一个收集各类客户文件的进件门户（Intake portal），接着是用于识别扫描件的 OCR 工具，再到财务报表分拆软件（Spreading tool）以提取并计算财务比率；随后信贷员还要从各大征信局拉取外部数据，套用模板起草信贷备忘录（Credit memo），再把数据手动录入贷款发起系统（LOS），其间还要在 Excel 电子表格、邮件往来记录以及投资组合监控看板之间来回搬运数据。 ```mermaid flowchart LR accTitle: 传统割裂信贷工具链与脆弱的人工集成层 subgraph S1["割裂的工具孤岛"] A["进件门户 / OCR"] --> B["报表分拆 / 征信数据"] B --> C["LOS / Excel / 邮件看板"] end subgraph S2["脆弱的集成层"] D["人类分析师<br/>(双屏操作 / 赶工易错)"] end S1 <-->|"多重交接造成信息丢失 (风险)"| S2 ``` 这 10 个工具之间彼此孤立，毫无共享上下文（Shared Context）可言。而串联起这整个庞杂系统的唯一“集成层”，竟然是人类分析师。分析师本应是团队中最昂贵、判断力最核心的资产，但在这种架构下，他们却不得不坐在双显示器前，在死线逼近的巨大压力下充当人工“数据搬运工”。每一次在不同工具之间的交接（Handoff），都会造成严重的信息衰减与遗漏；而在信贷与 AI 的语境中，这种信息的脱节就直接等同于“风险”。 Pranjal 创办的 Accend 致力于协助银行与金融科技公司实现更快、更精准的商业客户核保。在为合作银行做历史贷款档案回测与系统迁移时，团队发现了令人触目惊心的真实数据： - **符号颠倒错误**：在一个迁移的贷款组合中，超过 10% 的财务数据正负符号出现了颠倒。在 443 位借款人中，有 48 位的关键数据正负号完全搞反，直接导致后续的所有财务分析全盘皆错，造成灾难性的信贷误判。 - **75 倍净利润失真**：在一笔具体的贷款档案中，系统计算出的借款企业净利润竟然高达 7500 万美元，而实际数字仅为 100 万美元。这种离谱的偏差导致所有偿债比率和信贷备忘录彻底失真，而其根源正是系统工具之间根本没有上下文传递。 --- ### [07:19](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=439s) 2026年的堡垒式资产负债表：AI信贷操作系统与智能体架构 针对工具割裂与人工搬运带来的巨大风险，2026 年现代金融机构的“堡垒式资产负债表”演进为一体化的“AI 信贷操作系统（AI Credit Operating System）”。这一系统将进件摄取（银行流水、财务报表等）、报表分拆与财务分析、信贷备忘录生成、投后组合管理以及最终的企业信用评分无缝整合进同一个工作流中。系统中的每一个引擎的输出直接作为下一个引擎的输入，串行协同运行，并通过全局的统一上下文存储（Context Store）实现全流程可解释性——用户只要点击系统中的任何一个数据指标，就能一键反查其在 Excel 中的具体单元格位置、PDF 文档的精确页码以及原始文件出处。 ```mermaid flowchart TD accTitle: AI信贷操作系统多智能体协同架构 A["贷款档案摄取<br/>(40+文档 / 3财年历史)"] --> B["协调器 Orchestrator<br/>(任务分解与调度)"] subgraph S1["专业化执行智能体 (Agent 集群)"] C1["文档提取 Agent"] C2["行项目拆解与分类 Agent"] C3["信贷备忘录生成 Agent"] end subgraph S2["确定性验证与对账"] D["验证智能体 Verification Agents<br/>(资产负债表平衡 / 勾稽对账)"] end subgraph S3["人机协同与决策层"] E["人类信贷分析师<br/>(例外处理 / 最终商业裁决)"] end B --> S1 S1 --> S2 S2 -->|"自动闭环检查"| S1 S2 -->|"通过校验 / 触发例外"| S3 ``` 在底层技术实现上，这套体系依赖于严密编排的多智能体架构（Agentic Stack）： 1. **协调器（Orchestrator）**：接收包含 40 余份文档、跨越 3 个财年的复杂借款人档案，将其自动化拆解为细粒度的具体任务。 2. **专业化智能体（Specialized Agents）**：分别负责文档信息提取、逐行拆解与规范化分类、信贷备忘录起草等环节，犹如一场排练精密的“交响编舞”。 3. **验证智能体（Verification Agents）**：引入确定性检查与逻辑对账循环，自主核实资产负债表是否平衡、各项数据勾稽关系是否成立，一旦发现异常立即在流程内拦截报错。 4. **人机协同（Human-in-the-loop）**：明确界定“在安全边界内给予 AI 充分自主权（Autonomy where it's safe），在涉及重大商业决断与例外裁量时交由人类处理（Humans where the decision matters）”。 许多银行高管常问：“你们使用的 AI 模型的准确率是多少？”Pranjal 强调这是个“虽然不错但从根本上问错了方向的问题”。信贷核保的真正准确率并不取决于基础模型本身，而是取决于整套系统的工程架构。在整个体系中，基础大语言模型的能力只占最终准确率的 20%，其余 80% 的可靠性均由确定性检查机制、闭环验证循环、评测基准（Evals）和上下文工程共同决定。 --- ### [12:05](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=725s) 新旧模式对比：动态信用评分与专有上下文护城河 从 10 种割裂工具切换到 AI 信贷操作系统后，信贷作业流程迎来了彻底的重构： | 维度 | 传统信贷模式 | 2026 AI 信贷操作系统 | |:--- |:--- |:--- | | **文档分析** | 耗费数天手工翻阅与录入 | 数分钟内全自动提取与分拆 | | **备忘录起草** | 分析师耗费数天从零手动撰写 | AI 自动化草拟，分析师专注于审核与微调 | | **核保决策时效** | 需数周审批流 | 数小时甚至数分钟即可完成 | | **数据可信度** | 依赖手工录入，极易出现正负号错漏 | 每一个数字均具备完整的端到端审计溯源链条 | | **投后与信用评分** | 年度被动审查；依赖征信局静态快照评分 | 结合交易流水与支付行为的实时动态评分（Live Score） | 在这套全新体系中，最大的价值沉淀在于最终生成的**动态企业信用评分**。以往的评分往往依赖征信局几个月才更新一次的静态历史数据；而现在，系统融合了企业实时的银行交易明细、动态支付行为、历史履约记录以及深度的行业/客群上下文，能够根据企业业务的即时变化做出动态调整。 针对外界普遍担忧的“通用基础大模型（如 ChatGPT、Claude）未来是否会直接吞噬垂直应用”这一疑问，Pranjal 指出，基础模型的发展确实带来了强大的技术红利，但前沿模型本质上只是将“通用能力”商品化（让解析更多文档变得更便宜、更快速、更精准）。通用模型永远无法拥有机构内部沉淀的特定上下文——比如资深分析师过去对同类特定企业的裁决逻辑、机构特有的风险偏好与行项目分类规则。因此，信贷科技真正的护城河并非基础算法本身，而是机构自身沉淀的**专有反馈数据（Proprietary Feedback Data）**与**全流程审计追踪能力（Audit Trails）**。 ### [15:32](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=932s) 信贷评估的核心评估框架：CVC 模型 在商业信贷与核保决策中，输出的每一个数据都必须具备源头级别的可追溯性（source-level traceability）。针对如何评估现有的核保决策流程，Pranjal Daga 提出了一个关键的 **CVC 模型**（Context 上下文、Verification 验证、Compounding 复利）： ```mermaid flowchart TD accTitle: CVC商业信贷决策评估模型 subgraph CVC["CVC 决策评估框架"] C["Context（共享上下文）<br/>打通分散工具，消除数据孤岛"] V["Verification（审计验证）<br/>完整审计追踪与人工干预记录"] M["Compounding（复利飞轮）<br/>随着处理规模增加而越变越聪明"] end C --> V --> M ``` - **上下文（Context）**：评估机构在进行决策时，到底是在统一共享的上下文中做得更好，还是在依赖 10 个互不相通、各自为政的工具。不同工具之间实现上下文共享至关重要，孤立的数据流会导致信息断层。 - **验证（Verification）**：当模型运行并得出结论后，团队能否随时倒推回溯其运行过程？系统是否具备完整的审计追踪（audit trails）？中途是否有且仅在必要时经过了人工干预与修改？ - **复利（Compounding）**：这是决定平台长期价值的试金石。当系统处理到第 10,000 笔交易时，所拥有的上下文丰富度与决策精度，是否显著高于处理第 10 笔交易时？如果答案是否定的，那么机构在未来势必会被更底层的通用基础模型（foundation models）所颠覆与淘汰；只有随着数据积累持续自我迭代的系统，才能构建起真正的竞争壁垒。 --- ### [16:51](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=1011s) 构筑“堡垒系统”与信贷运营的未来形态 真正的“堡垒（Fortress）”模式绝非简单拼凑 5 到 10 个互不相干的单点工具，而是让所有环节在统一平台上无缝共享上下文。回顾历史，在 2008 年金融危机时期，机构可能需要依赖每天五道严格的人工核查来打造“堡垒式资产负债表”（Fortress Balance Sheet）。而展望 2026 年，这一堡垒的形态将被彻底重塑为：**一个能够纵览每一位借款人全貌、穿透检验每一个财务数字、并且伴随每一笔新交易而不断进化变聪明的统一 Agent 平台**。 ```mermaid flowchart LR accTitle: 2008年与2026年信贷风控堡垒对比 subgraph Y2008["2008 模式"] A["多道分散人工核查"] --> B["繁杂的单点工具"] end subgraph Y2026["2026 模式：AI 统一平台"] C["全视角覆盖借款人"] --> D["穿透级数据溯源"] --> E["单笔交易驱动复利进化"] end ``` --- ### [17:37](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=1057s) 行业终局预言：财务报表转录员消亡与专业判断力重塑 Pranjal 对未来商业信贷领域的岗位演进与竞争格局做出了清晰的预判： - **传统报表录入（Spreading）职能将彻底消失**：作为一门独立岗位描述的“报表转录/分拆（Spreading）”将不复存在，目前仍在由人工手动处理的基础杂务在未来一年内将被全面取代。 - **信贷专业人员的定位转变**：从业者的核心工作将转向**管理 AI 系统**，并在商业信贷领域充分发挥自身的专业底蕴，去制定更高维度的商业与风险判断（judgment）。 - **未来十年的赢家画像**：无论是一线银行、Fintech 公司、信用社（Credit Unions）还是专业借贷机构（Specialty Finance），获取顶尖的基础模型本身将变得极为普及，不再构成独占优势。未来十年真正能胜出的机构，必然是那些率先为自己搭建起最佳统一系统（Best System）的团队。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=4s) 准备重于预测 — Jamie Dimon > [07:19](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=439s) 在安全边界内给予 AI 充分自主权，在涉及重大商业决断与例外裁量时交由人类处理 — Pranjal Daga > [07:19](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=439s) 信贷核保的真正准确率并不取决于基础模型本身，而是取决于整套系统的工程架构。在整个体系中，基础大语言模型的能力只占最终准确率的 20%，其余 80% 的可靠性均由确定性检查机制、闭环验证循环、评测基准和上下文工程共同决定 — Pranjal Daga > [12:05](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=725s) 前沿模型本质上只是将“通用能力”商品化 — Pranjal Daga > [12:05](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=725s) 信贷科技真正的护城河并非基础算法本身，而是机构自身沉淀的专有反馈数据与全流程审计追踪能力 — Pranjal Daga > [15:32](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=932s) 当系统处理到第 10,000 笔交易时，所拥有的上下文丰富度与决策精度，是否显著高于处理第 10 笔交易时？如果答案是否定的，那么机构在未来势必会被更底层的通用基础模型所颠覆与淘汰 — Pranjal Daga > [16:51](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=1011s) 一个能够纵览每一位借款人全貌、穿透检验每一个财务数字、并且伴随每一笔新交易而不断进化变聪明的统一 Agent 平台 — Pranjal Daga > [17:37](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=1057s) 作为一门独立岗位描述的“报表转录/分拆”将不复存在，目前仍在由人工手动处理的基础杂务在未来一年内将被全面取代 — Pranjal Daga > [17:37](https://www.youtube.com/watch?v=Zq67jU6I_iI&t=1057s) 未来十年真正能胜出的机构，必然是那些率先为自己搭建起最佳统一系统的团队 — Pranjal Daga ## Mentioned Pranjal Daga, Accend, Jamie Dimon, JPMorgan Chase, Bear Stearns, Lehman Brothers, Washington Mutual, ChatGPT, Claude, Excel ## Video description Pranjal Daga - Co-Founder & CEO, Accend What happens when a credit team replaces ten disconnected tools with one agentic platform? We will explore that question through real production data from institutions processing hundreds of thousands of financial statement line items through AI from document intake and spreading to credit memos, business credit scoring and portfolio monitoring. In the past year, agentic AI has moved commercial underwriting from pilots to production, but a dangerous gap has opened between institutions running unified credit platforms and those stitching together point solutions that foundation model companies can disrupt overnight. We will discuss the before-and-after of AI-native underwriting: what actually changed in speed, accuracy, headcount leverage and risk visibility; and why a business credit score built on live financial data now outperforms anything a legacy bureau can offer. Attendees will leave with a practical framework for judging whether their credit stack will survive the next wave of AI disruption, and a roadmap any institution from regional bank to Fortune 500 — can follow to build an underwriting operation that compounds in accuracy with every deal. If your team is making credit decisions the same way it did eighteen months ago, this talk will cover exactly what your fastest competitors are doing differently. [Source](https://www.youtube.com/watch?v=Zq67jU6I_iI)

[0:04] Jamie Dimon: 准备重于预测 [7:19] Pranjal Daga: 在安全边界内给予 AI 充分自主权，在涉及重大商业决断与例外裁量时交由人类处理 [7:19] Pranjal Daga: 信贷核保的真正准确率并不取决于基础模型本身，而是取决于整套系统的工程架构。在整个体系中，基础大语言模型的能力只占最终准确率的 20%，其余 80% 的可靠性均由确定性检查机制、闭环验证循环、评测基准和上下文工程共同决定 [12:05] Pranjal Daga: 前沿模型本质上只是将“通用能力”商品化 [12:05] Pranjal Daga: 信贷科技真正的护城河并非基础算法本身，而是机构自身沉淀的专有反馈数据与全流程审计追踪能力 [15:32] Pranjal Daga: 当系统处理到第 10,000 笔交易时，所拥有的上下文丰富度与决策精度，是否显著高于处理第 10 笔交易时？如果答案是否定的，那么机构在未来势必会被更底层的通用基础模型所颠覆与淘汰 [16:51] Pranjal Daga: 一个能够纵览每一位借款人全貌、穿透检验每一个财务数字、并且伴随每一笔新交易而不断进化变聪明的统一 Agent 平台 [17:37] Pranjal Daga: 作为一门独立岗位描述的“报表转录/分拆”将不复存在，目前仍在由人工手动处理的基础杂务在未来一年内将被全面取代 [17:37] Pranjal Daga: 未来十年真正能胜出的机构，必然是那些率先为自己搭建起最佳统一系统的团队

[0:04] Pranjal Daga: preparation over prediction [17:19] Pranjal Daga: Spreading as a job description will not exist.

[0:04] My name is Pranjal and I am going to talk  about the AI credit operating system. [0:14] Now I want to go back one  year ago when I was right here [0:21] and I was talking about what the prediction  is going to be for commercial credit in 2027 [0:30] and I had said almost all decisions in commercial  credit will be done with the help of AI. [0:40] I was wrong. We are in 2026 and we  already see almost all of the decisions [0:48] in commercial credit being taken over in  some way, shape or form with the help of AI. [0:56] But let's go back in history a little bit. Let's  go back to 2008. I'm sure you uh remember the [1:04] crisis that happened at that time. Bear Sterns was  gone. Lehman Brothers just collapsed completely [1:14] and then one of the biggest banks at that  time, Washington Mutual had crashed as well. [1:20] But there was one financial institution  that came out of the 2008 crisis much better [1:29] than when it stepped into it and that was JP  Morgan Chase. Jamie Diamond, the CEO of JP Morgan, [1:40] had mentioned that they used to do risk checks  five times a day during that time and before that. [1:51] Why did they do that? It was not because they did  not have confidence in their predictions. they did [2:01] but they always focused on preparation  over prediction. So now the question is [2:10] what does this fortress balance sheet that  they'd mentioned before in 2008 by JP Morgan [2:19] doing these five checks a day. What does  this fortress looks like in 2026? So [2:28] you all should have like some um answer in  your mind and then we'll talk about like what [2:34] we think could be the f fortress. So the world for  the fortress that was built on is actually gone. [2:44] What we've seen in the past one year is a  lot of pilots going into production mode. [2:53] We've also seen many finance teams using  AI in production for their um operations [3:03] and for um credit decisioning as well. So you  can see here 600% jump in that and the speed [3:13] decisions are actually going from days and weeks  to hours and now in some cases even minutes. [3:22] So the delta we see is pretty high. Five manual  checks a day was gold standard before in 2006. [3:31] In 2020 uh in 2008, excuse me. In 2026, what we  see is the baseline is continuous monitoring. [3:43] So how this actually happens, you see here there's  like 10 different tools for uh commercial credit [3:53] decisioning. Firstly, there's an intake  portal. You take all the documents. [3:59] Then there's an OCR tool which is like reading the  documents, spreading software, understanding it, [4:06] then you're creating ratios. Then you  go to credit bureaus and pull the data. [4:12] You create the memos. You have  some template in mind to do that. [4:16] And then you manually put all of those things  together in your loan origination system. [4:21] Put that back into spreadsheets. Then  you have email threads. Then you have [4:26] portfolio dashboards to be doing like all of this  stuff. There's a bunch of tools that are just [4:33] siloed from one another. You see here 10  different tools. There's no shared context at all. [4:41] But you see there's one integration layer.  This is the analyst. This is the most expensive [4:50] and also unfortunately the most errorprone  uh entity in in this relationship. [4:57] Um and you're focusing on somebody who is taking a  look at two screens, has a deadline, has a lot of [5:05] judgment and capabilities, but they're doing  all of the manual work to do that as well. [5:12] You also see like all these different  tools have different handoffs. [5:16] And when you have handoffs multiple times there  is drop in information and in credit an AI world [5:26] drop in information has a term associated  with that and it's all it's called risk [5:33] when you do this you have a lot of  risk what does this uh quantify as [5:41] so I've given you two numbers here let's focus  on the first on um I run a company called Excend [5:48] and we help fintex and banks do their business  customers underwriting faster and more accurately. [5:55] When we partner with a bank or a financial  institution, we take in their loan files as well [6:01] and we do a back test of all of the loans that  they have provided. We find out that in a migrated [6:08] portfolio the sign the positive or negative  sign of the numbers was off by more than 10%. 48 [6:19] out of 443 borrowers had a sign change which means  that everything in in the analysis was incorrect [6:30] catastrophic. Now the second number actually  makes me um not sleep at night. And why is that? [6:38] Um you see the net income on one of um the  loans that was provided was calculated out [6:47] to be $75 million when the actual number was 1  million. You can see the discrepancy here. It's [6:57] bizarre how different those things would be.  It's bizarre how different the ratio calculation [7:04] and the memos will be as well. So those loan  decisions are not great. And why this happened? [7:10] Because no um no information was  passed from one system to the other. [7:19] What does a fortress balance  sheet actually looks like in 2026? [7:24] I actually call it the AI credit operating  system. You have the intake workflow [7:31] where you have ingestion of bank statements,  financial statements and other uh documents. [7:38] You do the spreading which is financial analysis,  bank analysis and this um the whole process around [7:44] that. Create a memo, do portfolio management  and ultimately create a business credit score. [7:52] all of those in one particular workflow and  they're shared context. The output of one of [8:00] this uh one of these engines is the input  to the other. So it's going sequentially. [8:08] An important thing to note here is all of these  are agentic workflows and every single agent [8:17] sees the whole borrower's profile  everything from the beginning till the end. [8:23] So there's one context store which means if there  is um if there's like some correction that needs [8:30] to be done on one particular number you'll  be able to you should be able to click on it [8:36] find it where it is in the exact same Excel cell  the page number and also uh the entire document [8:46] and obviously the credit score that  you create should be coming from like [8:50] multiple data sources which we'll talk about in  a minute. But for the practitioners in the room, [8:57] I want to go a little bit deeper. There's this  agentic stack. Now, the word agentic um has [9:05] been thrown around a lot in this conference. So,  actually, let's go a little bit deeper into that. [9:12] The first layer is an orchestrator. You give  it a loan file maybe 40 different documents [9:20] uh three fiscal years um and  multiple loan files as well. [9:25] It will decompose everything into specific  tasks and these task will be taken up by agents. [9:33] Now these specialized agents will have  specific tasks as I mentioned before [9:38] document extraction converting everything into  specific line items and then categorizing those [9:44] line items and creating a credit memo. All of  those are separate tasks. So separate agents [9:50] come together. It's like a very coordinated uh  well choreographed dance that they do. And then [9:59] the important thing to note here is you  can create those workflows very easily [10:05] but you need to create checks in place.  So you need to create verification agents. [10:12] Is the balance sheet balanced? Is everything  reconciling? Agents should be able to do those [10:18] checks in place for you and then provide you  the output whether something's off or not. [10:26] But you see at the very end there are humans.  If you want to make judgment or if you want [10:33] to do exception calls, you should be able to  have humans in the loop to be able to do that. [10:41] Important thing to note here, we see in  the first three things there's autonomy. [10:46] So autonomy where it's safe and humans where  the decision comes in where it actually matters. [10:59] Now the other side of the equation is a  lot of executives when we talk to them [11:04] when I'm talking to them we're going to these  conferences meeting them they always ask me [11:09] um what's the accuracy of your AI or what's the  accuracy of the uh of the AI that you're using. [11:15] It's a good question but I think it's the wrong  question. It's slightly wrong but importantly [11:23] wrong question. What should be the question  is what's the accuracy of the system [11:29] and the accuracy is actually not based on  the model it's based on the system in place. [11:36] So um you see here it depends on like multiple  things. Deterministic checks need to be in place. [11:43] You need to have verification loops and then  the model is actually 20% of the entire thing [11:51] of the entire system. You have like other  things verification, eval context windows [11:58] and so on so forth that takes up the  majority of the rest of the system in place. [12:05] So let's see what's before and after. Obviously we  can go through all of these things. So, a lot of [12:10] time gone into analyzing documents. Um, it takes  days to craft a credit memo, weeks for a decision. [12:19] Um, you do annual portfolio reviews and then  your score, your business credit score is a [12:26] snapshot. It's a one-time static looking score.  However, if you go into the agentic side in 2026, [12:36] it looks like this. Everything happens in  minutes and you have verification trail, [12:42] audit trail on every number, every decision.  You want to craft memos now. You craft those [12:48] memos with the help of AI and then your analysts  on the team. They're actually not writing them, [12:55] they're reviewing and refining those memos.  Um, you make a decision very quickly. Um, [13:01] and then you have a score that's live. It lives  it changes based on the changes in the business [13:08] and how it changes the payoff of all of  this that we have talked about is the score. [13:15] So you don't only have snapshots, you don't only  have bank statements, financial statements. Those [13:19] are very important things, but you also have all  the transaction data, you have payment behaviors, [13:26] history of things. All of those things  combined can give you a score. Previously [13:32] you have the scores based  on static bureau data only. [13:38] Now things are changing. You are taking a look  at scores which are deeply um contextdriven [13:46] based on the industry, based on the business,  based on the customer customer type and so forth. [13:55] And then what you see here is there's a big  tailwind. Models actually matter. So foundational [14:04] models that come up matter a lot. But data matters  so much more. And when you have a system that [14:11] sees the entire borrower portfolio, you can  have a better system, better business score. [14:21] Now the other question that comes up is okay [14:26] when Chad GBT and Claude are going to create these  things this will be fine we're going to be good [14:35] but foundation models they only are commoditizing  capabilities they're not giving you the context [14:41] of things what I mean by that is once you have a  better model of course It's going to be the case [14:51] that more documents will be analyzed in fewer  amount of time or less amount of time. Um, [14:58] and they're going to be much better in terms  of accuracy, but they would not have context on [15:04] what your analyst has said about the same type of  company before, how they're making decisions, how [15:10] you classify things, so forth. So your moat when  you're taking a look at solutions or providers [15:20] is proprietary feedback data and also the  audit trails that you bring to the table. [15:29] You have like every agent that's like coming in [15:32] but then every number that goes out  needs to have source level traceability. [15:40] Now what are the questions to ask uh when you take  a look at underwriting decisions the CVC model. [15:46] So context when you have the context  in place are you doing better [15:54] or you have 10 different tools that are not  talking to each other. So context is not shared [16:00] extremely important to have context shared  among different tools. Second verification. [16:09] When the model is run already, can you go back and  see how it was run? What the audit trails? Has any [16:16] human touched it or not? And then importantly, the  compounding. When you take a look at deal 10,000, [16:28] do you get more context than when you're taking  a look at the deal number 10? If you are saying [16:35] that that's not going to be the case, yes, I  agree with you. You will be disrupted with uh [16:43] the foundation models. But if you are continuously  improving, then you can go ahead with that. [16:51] So now let's take a look at all of these  things. Again, the fortress is this model. [16:59] It's like five different tools or 10 different  tools coming together and having shared context. [17:06] Going back in history in 2008, there were five  checks a day to create a fortress balance sheet. [17:19] What does that look like in 2026? one platform  that sees every borrower, takes a look at [17:27] each and every number and also with every  single deal it gets smarter and smarter. [17:37] I made a prediction last year and AI has gone  uh and come around very quickly from there. [17:47] But I want to make a prediction this year  as well and I hope things get better and [17:53] I'm incorrect next time as well. Spreading  um as a job description would not exist. [18:03] The things that you're doing manually next year  would not exist. What will exist and what the job [18:09] will be will be to manage systems and creating  better judgment uh based on your expertise uh in [18:18] the commercial credit world and the institutions  who are going to be winning in the next decade [18:24] the banks the fintexs the credit unions or  specialty finance uh institutions they will [18:31] have access to the best models because it's going  to be so widespread and everybody will have access [18:36] to it But they will have access by themselves to  have the best system in place as well. Thank you [18:45] so much. My name is Pranjul. I'll be around uh for  your questions. Um and yeah, appreciate your time.