Enverus VP Says Proprietary Data Fails to Create Enterprise AI Moat

Enverus副总裁表示专有数据无法构筑企业级AI护城河 若缺乏针对特定领域微调的模型、严格的治理以及允许人工监督的工作流控制，庞大的专有数据集将无法形成可持续的竞争优势。 Akash Sharma对业内普遍存在的假设提出了挑战，即认为拥有庞大的专有数据库会自动为企业级AI构筑起竞争护城河。他强调，多达95%的AI项目由于上下文膨胀、成本失控以及缺乏协同的试点而无法进入生产阶段，并据此阐述了Enverus的四大支柱架构：基础落地（Grounding）、治理（Governance）、组合编排（Composition）与控制（Control）。通过用于油气现场安全的Project Lifesaver等实际案例，Sharma展示了领域微调的嵌入模型、语义分层的智能体工作流以及明确的“人在回路”控制机制，如何赋能企业构建值得信赖且达到决策级标准的AI系统。 仅凭专有数据只是一个先决条件，若缺乏适当的垂直行业架构，它并不会自动为AI构建起可持续的竞争护城河。 (0:04) 95%的企业级AI项目未能进入生产阶段，同时预计到2027年，由于成本、目标不明确和扩展难题，将有40%的智能体AI项目被取消。 (2:08) 构想数量从少数几个激增至700多个潜在AI项目，在验证和生产规模化方面造成了巨大瓶颈。 (2:08) 在特定行业术语、地质区带及盆地概念方面，经领域微调的小模型和专用嵌入模型的表现显著优于通用前沿模型。 (3:43) 试图通过增大上下文窗口（检索大量分块）来弥补嵌入效果的不足，会导致上下文膨胀、延迟增加以及Token成本上升。 (3:43)

Proprietary data is an insufficient competitive moat for enterprise AI without a tailored vertical architecture, according to Akash Sharma, Vice President of Product Management at Enverus. The Enterprise AI Production Gap (2:08) Despite significant industry investment, enterprise AI projects frequently stall before reaching production. Sharma points to an MIT study indicating that approximately 95% of enterprise AI initiatives fail to deploy, while Gartner projects that 40% of agentic AI projects will be canceled by 2027 due to scaling challenges, unclear goals, and rising expenses. Sharma attributes part of this bottleneck to an unmanageable explosion in ideation: teams that previously vetted ten concepts to launch two now face over 700 potential ideas, complicating validation in capital-intensive, regulated sectors like energy. Domain Tuning Over General-Purpose Retrieval (3:43) General frontier models fail to handle complex vertical concepts out of the box through prompting alone. While off-the-shelf and domain-tuned models perform similarly on generic tasks such as identifying company names, specialized models deliver far higher precision on niche industry terminology like geological basins and plays. Enverus evaluated custom-tuned embeddings against Amazon's Titan Text Embeddings V2, finding that its domain-tuned model attained a 56% hit-rate accuracy on the very first retrieved chunk (K=1). In contrast, compensating for weaker retrieval by expanding context windows to 50 or 70 chunks causes context bloat, increased latency, and high token costs. Four Pillars of Decision-Grade Architecture (8:52) To transition data into an operational asset, Sharma outlines a four-part framework: Grounding, Governance, Composition, and Control. Grounding ensures system outputs tie directly to verified industry assets. Governance introduces evaluation frameworks that scale as multi-agent swarms increase in complexity. Composition requires building reusable, interconnected layers from the start to prevent fragmented enterprise pilots and unexpected token usage. Finally, Control incorporates human oversight into high-stakes scenarios, ensuring AI assists rather than unilaterally executes decisions affecting capital deployment or physical safety. Applied Field Safety and End-to-End Lineage (11:55) Enverus implemented this structured framework in Project Lifesaver, a field safety application currently in beta testing with a consortium of four major oil and gas operators. Because erroneous safety instructions pose immediate hazards to human life and machinery, the application relies on strict human-in-the-loop checkpoints, recognizing that field experience in variable conditions must supersede unmonitored model outputs. Sharma stresses that scalable agent architectures must maintain strict data provenance and traceability, enabling operators to trace the precise source chunk and agent behind every generated fact across multi-step workflows. Converting proprietary datasets into a defensible moat requires mapping where public model knowledge drops off, connecting point tasks to downstream business outcomes, and explicitly naming the human accountable for high-consequence decisions.

Enverus 产品管理副总裁 Akash Sharma 指出，专有数据本身并不能自动构成人工智能的竞争护城河；企业必须建立涵盖锚定、治理、系统组合与人类控制的垂直 AI 架构，才能解决绝大多数企业试点无法投入生产的困境。 企业AI投产瓶颈与数据误区 (0:04) Sharma 强调，尽管 Enverus 拥有 25 年历史并管理着 2.7 PB 的能源数据以及超过 3.5 亿条土地管理记录，但拥有海量专有数据仅仅是入局的先决条件，而非完整的解决方案。现实中企业普遍陷入落地困境：Gartner 预测，到 2027 年将有 40% 的 Agentic AI 项目因成本失控、目标模糊或难以扩展而宣告取消；而麻省理工学院（MIT）的一项研究也显示，约 95% 的企业 AI 试点项目从未成功进入生产阶段。造成这种瓶颈的部分原因在于创新漏斗前端的爆炸式增长——团队需要评估的想法从以往的 10 个激增到 700 多个，导致验证和扩展严重受阻。此外，部分企业管理者误以为只需将大型数据库直接接入前沿大语言模型即可解决业务报告问题，这种忽视架构设计和工作流集成的尝试往往以失败告终。 领域微调与嵌入检索优化 (3:43) 在检索增强生成（RAG）系统中，通用前沿模型无法仅靠提示词或技能工程来稳定处理垂直领域的复杂业务。Sharma 认为，针对特定任务对较小模型进行领域微调比盲目追求通用大模型更加有效。测试表明，尽管前沿大模型与垂直微调模型在识别公司代码等通用实体上表现相当，但在识别地质构造、盆地和油气区块等深层行业术语时，领域微调模型展现出显著优势。 在向量检索层面，基于 Enverus 专有语料微调的嵌入模型在单块检索命中率（K=1）上达到了 56%，在租约名称和地质构造等专业概念的检索精度上大幅超越了 Amazon Titan Text Embeddings V2。Sharma 警告称，如果试图通过将检索上下文扩大到 50 至 70 个切片来解决“大海捞针”式的召回问题，必然会引发上下文膨胀，导致系统响应延迟增加和 Token 成本急剧攀升，最终破坏生产级系统的可行性。 决策级垂直架构四大支柱 (8:52) 为将企业数据转化为真正的护城河，Sharma 提出了构建决策级垂直 AI 系统的四大支柱。第一是“锚定（Grounding）”，将前沿模型的输出与垂直行业经验证的操作现实和数据结构直接对齐；第二是“治理与评估（Governance & Evaluation）”，建立严密的量化评估体系，且评估严格程度必须随着多智能体集群或工具构建智能体的自主权提升而同比例提高；第三是“系统组合（Composition）”，必须自始至终进行全局协同设计，避免孤立试点导致数十个分散系统和失控的 Token 账单；第四是“控制（Control）”，在能源等涉及重资产分配和人身安全的高风险行业，核心决策必须保留人类专家的最终干预权。 基于这一框架，Enverus 设计了向后兼容的语义分层架构：底层技能支撑工作流，工作流驱动智能体，智能体再赋能具体应用，所有流程均统一接入中央治理层，确保工程组件能够跨既有软件套件持续复用。 高风险场景实践与工作流整合 (11:55) 为了在现实环境中验证分层架构，Enverus 与四家大型油气客户联盟联合推出了处于测试阶段的“Project Lifesaver”现场安全应用。在油气作业现场，模型生成虚假或错误的指令会直接威胁重型设备与工人生命安全。由于工人的现场经验往往优于任何算法，该系统被严格设计为向作业人员提供精准参考信息，而非直接做出自主操作决策。 基于实践经验，Sharma 建议企业在部署垂直 AI 时遵循明确原则：首先绘制自身垂直领域的“知识悬崖”，在公共大模型能力急剧下降但专家直觉决策密集的断层区集中投入资源；其次在多智能体系统中建立端到端的数据血统与追踪机制，确保能够精确定位是哪一个特定智能体检索到了支撑决策的微小事实；最后，不能仅停留在单点任务加速上，必须将应用与下游最终业务结果紧密挂钩，并明确指定该环节承担风险与机会成本的具体责任人。 Sharma 总结指出，单纯的数据堆积无法形成防御壁垒，只有将专有数据与严谨的垂直分层架构相融合，并建立清晰的追踪溯源与人类把关机制，企业才能跨越原型试验与生产落地之间的鸿沟。

## Background 本视频由 Ai4 于 2026 年 8 月 23 日发布在 YouTube 平台上，内容为 Enverus 产品管理副总裁 Akash Sharma 的演讲。视频简介指出，演讲探讨了在能源等高风险领域中，如何通过具备数据溯源、领域评估、工作流整合与人工干预机制的垂直 AI（Vertical AI）架构，将企业私有数据转化为具备“决策级”能力的核心护城河。 ## TL;DR Akash Sharma 首先打破了“私有数据即护城河”的传统迷思，指出绝大多数企业 AI 项目由于缺乏系统架构而止步于试点。他结合 Enverus 的实践经验，介绍了如何通过小型的领域微调模型与专用嵌入模型解决专业术语识别难题，并提出了包含接地（Grounding）、治理（Governance）、组合（Compose）与控制（Control）四大支柱的垂直 AI 架构。此外，他还分享了用于现场高风险环境的“Project Lifesaver”案例，并为企业在多智能体协同、数据溯源及端到端业务流程设计方面提供了可落地的避坑指南。 ## Atmosphere & style 演讲风格专业、务实且充满洞察力，演讲者语气沉稳自如，善于抛出反主流观点来引发观众思考。整体氛围兼具行业前沿的科技感与落地高风险领域的谨慎严谨，打破了市场对 AI 狂热的泡沫，给人一种理智、深刻且注重实战应用的切实感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=4s) 即使拥有一流的数据基础，私有数据本身也不直接构成护城河，必须依靠正确的垂直 AI 架构将其转化为决策级的生产系统。 - [02:08](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=128s) 企业 AI 项目落地困难重重，Gartner 预测 40% 的 Agentic AI 项目将在 2027 年前被取消，而 MIT 研究也指出 95% 的项目未能成功进入生产环节。 - [03:28](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=208s) RAG 中的检索（Retrieval）地位类似于传统机器学习中的回归分析，虽然基础但极具挑战，盲目扩大检索量会导致严重的“上下文膨胀”和高成本。 - [05:08](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=308s) 通用大模型缺乏行业特定术语和关联逻辑，对特定领域的微调与专用嵌入模型（Embedding）能够显著提升检索准确率度并降低延迟。 - [08:52](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=532s) 构建企业级垂直 AI 架构必须具备四大核心要素：真实化 grounding、严苛的治理与评估、跨系统工作流的组合性，以及人类掌控（Human-in-the-loop）。 - [11:55](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=715s) 单独将数据库直连大模型或依赖单一 Tool Use 无法应对复杂的企业场景，项目架构需要分层解耦并实现语义上的向后兼容。 - [13:55](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=835s) 在高风险场景（如油田现场）中，经验丰富的现场专家必须被作为安全检查点融入系统，确保安全的同时加快迭代。 - [16:46](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=1006s) 推进企业 AI 落地需寻找行业专属的“决策断崖”，在设计初期即保障数据溯源与安全，并确保架构能够无缝复用到后续的多个工作流中。 - [18:30](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=1110s) 梳理需求时应采用端到端的全局流程视角，避免仅解决单一节点的检索问题，并必须明确指定人机环路中的具体责任人。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=4s) 重新审视数据护城河与 Enverus 的数据底座 演讲者 Akash Sharma 是 Enverus 的产品管理副总裁。他在午餐前的这一场演讲中，首先抛出了一个反主流的观点：大家都认为拥有私有数据（Proprietary Data）就是企业在 AI 时代的护城河，但事实并非如此简单。Enverus 作为全球最大的专注于能源领域的 SaaS 软件公司，成立至今已有 25 年，由黑石集团（Blackstone）持有。公司提供覆盖整个能源价值链的数据分析、咨询以及 AI 解决方案。 尽管 Akash 提出数据本身不直接构成护城河，但 Enverus 本身确实坐拥极其庞大的数据资源。公司拥有大约 2.7 PB 的数据，涵盖私有数据、公开数据以及经过深度加工的增强数据集，内容横跨地下地质（Subsurface）、地面设施、可再生能源以及数据中心选址等领域。在非结构化数据方面，仅土地管理（Land Management）这一项，公司就拥有超过 3.5 亿条可追溯至 19 世纪（1800年代）的历史记录。Akash emphasize 强调这些数字并非为了自夸，而是要说明：无论处于哪个垂直行业，企业都会深受其行业专属数据特性的影响，如何将这些数据转化为真正的竞争壁垒，既是挑战也是巨大的机遇。 - **公司成立年限**：25 年 - **母公司/控制方**：黑石集团（Blackstone） - **数据总量**：约 2.7 PB（含私有、公开及增强数据） - **非结构化数据规模**：仅土地管理板块即拥有 3.5 亿+ 条记录，最早可追溯至 1800 年代 --- ### [02:08](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=128s) AI 落地困境与检索增强（RAG）的本质 为了把问题说清楚，Akash 引用了两份业内非常著名的报告数据。首先是 Gartner 的一份报告预测：到 2027 年，将有 40% 的 Agentic AI（智能体 AI）项目被取消。虽然目前市面上充满了狂热与兴奋，但绝大多数项目未能真正进入生产环境（Production）并实现规模化，原因包括成本超支、目标模糊等。在 B2B、高监管、重资本投入的能源行业，这种落差显得尤为严峻。 另一个广为流传的数据来自麻省理工学院（MIT）的著名研究：95% 的企业 AI 项目从未能够成功落地生产。Akash 提到，前段时间他几乎每次和客户开会，对方都会提到这个 95% 的统计数字。不过他也提醒大家需要客观对待这一现象：AI 革命急剧放大漏斗顶部（Top of the Funnel），过去企业可能只评估 10 个想法并挑选 2 个推进，现在则有 700 个想法等待筛选，这自然拉低了转化率。 谈到技术实现，在当前 RAG 架构的主流讨论中，Akash 提出了一个深刻的比喻："RAG 中的检索（Retrieval）对于生成式 AI 落地，就像回归分析（Regression）对于传统机器学习一样。" 虽然检索逻辑看似简单直接，但并不意味着它不起作用，真正的难点在于如何处理好其中的各种细节与行业现实。 - **Gartner 报告预测**：到 2027 年，40% 的智能体 AI 项目将被取消。 - **MIT 统计数据**：95% 的企业 AI 项目未能进入生产阶段。 - **项目筛选变化**：从过去的“10 个想法选 2 个”，演变成当下“700 个想法待评估”。 --- ### [05:08](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=308s) 领域微调与专用嵌入模型（Embeddings）的突破 在构建智能体系统时，通常需要关注模型、数据集与上下文层（Context Layer）三个前沿。Akash 明确表示，Enverus 的目标并不是要做一个“能源界的 Claude”，因为在通用逻辑推理和公开知识上，前沿大模型已经非常强大。但在特定领域内，行业术语（Jargon）极其繁复。如果每次交互都将庞大的术语表作为上下文传入，成本与效果都不理想。因此，Enverus 对小型的领域专属模型进行了微调（Domain Tuning）。在公开指标（如公司代码 Ticker、公司名称等）上，微调模型与前沿通用模型不相上下；但一旦涉及地质地层（Geological Formation）、油气区带（Play）、盆地（Basin）等专业概念时，通用模型的准确率大幅下滑，差距被迅速拉大。 更为关键却常被忽视的是嵌入模型（Embedding Models）的微调。嵌入决定了数据在检索生态系统中的表现。Enverus 将自身在特定语料库上微调的多个嵌入模型，与亚马逊的 Titan V2 进行了对比。Titan V2 作为通用嵌入模型固然极其优秀，但在处理诸如“租约名称（Lease Name）、所有者名称、所在地层及其地质关联”这类看似杂乱的行业术语时，专用微调模型优势极其明显。 在评估指标上，Enverus 最好的领域微调嵌入模型在 Top-1 检索（K=1，即检索出的第一个文本块是否准确）上的准确率达到了 56%。虽然有人会问“为什么不直接检索 50 或 70 个文本块来覆盖答案？大模型不是越来越擅长针大海捞针了吗？”，但 Akash 指出，这会导致严重的“上下文膨胀（Context Bloat）”和高延迟，每次检索都要消耗大量 Token。一个高效的专用嵌入模型，能让检索系统同时实现更高准确率、更低延迟和更低成本。 - **K=1 检索准确率**：Enverus 最优领域微调嵌入模型的 Top-1 准确率达到 56%。 - **对比对象**：Amazon Titan V2 通用嵌入模型。 - **盲目增加 K 值的弊端**：产生上下文膨胀（Context Bloat）、增加系统延迟与 Token 运行成本。 --- ### [08:52](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=532s) 垂直 AI 架构的四大核心支柱 要打造真正将数据转化为护城河的 enterprise 架构，Akash 提出了必须思考与回答的四个核心问题： 1. **Grounding（接地/真实化）**：如何将前沿大模型与行业特定的现实逻辑、专业术语进行精准锚定？ 2. **Governance（治理）**：如何对系统建立严格的评估框架（Evaluations）？系统越复杂（如引入多智能体编排 Swarms、自主构建工具的 Agent），就越需要极其严苛的评估机制。 3. **Compose（组合性）**：许多企业 AI 项目停留在单一孤立的 Pilot 阶段。如果缺乏系统性架构，最终会导致企业内部散布着 25 个互不相通的试点项目，Token 账单飞涨。因此必须从一开始就从整体系统进行组合化设计。 4. **Control（控制）**：在能源这种涉及重资本投入乃至人身安全的行业，最终决策必须由人类掌控（Human-in-the-loop），AI 系统应当提供信息辅助而非直接越权决策。 Akash 强调："我不认为仅仅拥有数据本身就能构成护城河。" 他透露，大约一年半前，曾有一位能源公司的高管信心满满地表示，他们只需把庞大的石油和天然气生产数据库直接连接到一个大语言模型（LLM）上，就能解决所有的自动报表需求。但事实证明，这种想法最终完全落空。单纯依赖某个特定模型的工具调用（Tool Use）或智能体编排能力，根本无法支撑复杂的企业级应用。 --- ### [11:55](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=715s) 分层解耦架构与“救生员项目”实战案例 Akash 展示了 Enverus 内部采用的分层架构图。该架构在语义上实现了向后兼容（Backwards Compatible）：应用层可以使用智能体（Agents），智能体可以使用工作流（Flows），工作流可以使用技能（Skills）。这种互联互通性解决了组合性难题，任何团队成员构建的工作流都不会孤立存在，而是可以作为上层的资源被调用，或通过可组合的 UI 生态系统触发。整个体系通过统一的治理层控制，无论是连接数据资产、智能体还是工作流，都遵循相同的安全与治理规则。 为了验证这套架构的威力，Akash 分享了一个具体的客户案例。几个月前，四大大型油气公司找到 Enverus 表示，AI 在办公室分析工作中表现优异，但石油工业的许多核心生产活动发生在油田现场（Oil Field）。现场属于高风险环境，对安全标准的要求极高。一旦 AI 系统给拉取销售报表的分析师提供错误信息，代价可能只是修改数据；但如果给操作重型机械的现场工人提供错误的安全建议，后果不堪设想。 针对这一需求，Enverus 与这几家油气公司组成的联盟合作推出了名为 **"Project Lifesaver"（救生员项目）** 的 Beta 版本。该系统将上述架构的六个环形图层落地到油田现场场景中。系统并不直接代替现场专家做决策，因为“再多的模型训练和微调，也比不上数十年的现场经验”。现场专家能根据当天的气温和环境，精准判断各项工序的操作节奏。Project Lifesaver 将这些有机知识（Organic Knowledge）作为安全检查点（Safety Checkpoint）融入系统，既保障了现场安全，又大幅加快了新功能的迭代速度。 - **案例项目名称**：Project Lifesaver（救生员项目，目前已推出 Beta 版） - **合作方**：由 4 家大型油气巨头组成的客户联盟 - **核心原则**：将经验丰富的现场工人作为系统的最终决策者与安全检查点 --- ### [16:46](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=1006s) 避坑指南：寻找行业悬崖与数据溯源 在演讲的结尾，Akash 为企业推进 AI 落地总结了四条核心建议： 1. **寻找你所在行业的“断崖”（Find your own cliff）**：不要试图为金融或能源行业重新造一个通用模型。公共领域的通用知识大模型已经处理得很好，关键是要找到通用知识落空的那个“断崖点”。行业专家在回答一个问题前，大脑中可能已经隐性地做出了 25 个子决策，必须找出这些让行业工作流程区别于其他领域的具体细节。 2. **设计即保障数据溯源与安全（Data provenance and security by design）**：虽然这些词听起来像营销套话，但一旦开发铺开，项目将以极快速度推进。如果不事先搭建好数据溯源与安全结构，事后补救将异常困难。 3. **多智能体环境下的深度溯源与评估**：随着系统从单智能体走向多智能体、甚至能自主构建工具的 Agent Team，能够精准查明“第 17 个智能体是在哪里提取到了那条关键事实并影响了最终输出”，对于验证和评估系统至关重要。 4. **资助第二个工作流（Fund the second workflow）**：避免只针对单一孤立场景开发，要确保架构能够无缝拓展并复用到后续的多个工作流中。 - **行业决策深度**：现场专家在回答一个问题前，通常会在潜意识中做出大约 25 个串行决策。 - **追溯复杂性**：在多智能体系统下，必须能够精准定位到例如“第 17 个智能体”的数据获取源头。 ### [18:30](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=1110s) 从端到端流程视角洞察客户需求与人机协作的风险把控 在企业内部推进 AI 系统落地时，绝不能仅仅满足于局部效率的微小提升。与客户或内部业务部门直接沟通时，不要简单地追问“你最需要解决的核心痛点是什么”，而应当引导他们梳理出完整的业务全流程。以 HR 系统为例，如果仅仅构建一个“帮助 HR 更快检索信息”的工具，并没有触及问题的本质；必须追问这个动作最终指向什么业务结果，以及当这个问题被解决后，紧接着会触发什么新问题、这两个问题之间是否存在前后关联。深入开展这种端到端的流程调查与剖析，对于准确把握需求至关重要。 与此同时，必须明确指定“人机环路（Human-in-the-loop）”中具体的责任人。随着 AI 模型与技术框架在不同风险偏好的业务场景中深入应用，忽视这一环节所带来的机会成本与合规风险是极其巨大的。在评估已有的 AI 项目或计划引入的新项目时，都需要从流程关联与人机协同等四个核心维度进行深刻审视。 - **全局流程洞察**：避免孤立地解决单一节点问题（如单纯加速信息检索），需打通前后逻辑链条与最终业务目标。 - **明确人机责任**：将“Human-in-the-loop”落到具体的人头与岗位上，有效规避高风险场景下的误决策损失。 ### [19:27](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=1167s) 构建垂直 AI 架构，将数据转化为核心护城河 企业现有的私有数据本身并不能直接构成不可替代的壁垒，唯有通过匹配正确的 Vertical AI（垂直 AI）架构，将其转化为具备“决策级”产出能力的系统，数据才能真正成为企业的核心护城河。在演讲的最后，演讲者邀请现场观众在午餐休息期间前往展厅 333 号展位（Booth 333），与 Enverus 团队展开进一步的交流探讨，无论是赞同演讲观点还是持有不同见解，都欢迎进行思想碰撞。 - "your data is not your moat but it can become with the right architecture."（你的数据现在还不是护城河，但只要搭配正确的架构，它就能成为护城河。） - **展位交流信息**：Enverus 团队位于展厅 333 号展位（Booth 333），欢迎参会者针对垂直 AI 架构与决策级落地应用展开深入讨论。 ## Notable quotes > [02:08](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=128s) RAG 中的检索（Retrieval）对于生成式 AI 落地，就像回归分析（Regression）对于传统机器学习一样。 — Akash Sharma > [08:52](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=532s) 我不认为仅仅拥有数据本身就能构成护城河。 — Akash Sharma > [11:55](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=715s) 再多的模型训练和微调，也比不上数十年的现场经验。 — Akash Sharma > [19:27](https://www.youtube.com/watch?v=Xiq_E20yqmc&t=1167s) your data is not your moat but it can become with the right architecture. — Akash Sharma ## Mentioned Akash Sharma, Enverus, Ai4, 黑石集团 (Blackstone), Gartner, 麻省理工学院 (MIT), Claude, Amazon Titan V2, Project Lifesaver ## Video description Akash Sharma - VP Product Management, Enverus Every enterprise has proprietary data. Far fewer have an AI system they’d trust to make a decision that carries real consequence. RAG can retrieve the right context, but production takes more than retrieval: governed data with lineage, domain-specific evaluation, workflow integration, and execution a human can override. Akash Sharma draws on Enverus’ work taking AI to production in energy, one of the hardest domains there is, where a wrong answer is a misallocated capital budget, a failed compliance filing, or a safety incident. If a system earns trust there, the architecture behind it holds anywhere. He lays out that architecture, the Vertical AI blueprint for turning proprietary data into decision-grade systems, then names why promising pilots stall, what production-ready systems do differently, and how to find the weakest layer in your own program. You’ll leave with a diagnostic you can run on your own architecture on Monday, and a clear read on what to build next. [Source](https://www.youtube.com/watch?v=Xiq_E20yqmc)

[2:08] Akash Sharma: RAG 中的检索（Retrieval）对于生成式 AI 落地，就像回归分析（Regression）对于传统机器学习一样。 [8:52] Akash Sharma: 我不认为仅仅拥有数据本身就能构成护城河。 [11:55] Akash Sharma: 再多的模型训练和微调，也比不上数十年的现场经验。 [19:27] Akash Sharma: your data is not your moat but it can become with the right architecture.

[16:46] Akash Sharma: Energy professionals are not a collection of energy information—there are sequences of tasks we all do. When I work with somebody working in the field and ask them a question, there's about 25 decisions they make subconsciously before even answering that question. [19:27] Akash Sharma: Your data is not your moat, but it can become [one] with the right architecture.

[0:04] Uh thank you everyone for uh coming out to this  session right before lunch. So this should be [0:09] should be fun. Uh my name is Akash Sharma. I'm the  vice president of product management at Envirus. [0:14] Uh Invers is uh the world's largest uh energy  focused software as a services firm. uh we've [0:19] been around for 25 years and uh the topic that I  want to cover today uh is is a bit of a contrarian [0:26] to how we hear a lot about conversations around  AI with the importance of your proprietary data [0:32] and in line with the track theme so to speak. But  I want to highlight some of the things that are [0:37] important to actually make uh this information  and this asset you know differentiated and emote [0:43] for your organization. So if we yeah a little bit  about us um as [clears throat] I mentioned uh we [0:52] are we're the largest software as a services firm  uh focused on energy uh we've been around for 25 [0:58] years we're owned by Blackstone and uh across  the energy spectrum we provide data analytics [1:03] consulting and now AI solutions and services uh  contrary to what I said at the beginning a couple [1:10] of minutes ago we do sit on a large trove of data  uh with about 2.7 seven pabytes of proprietary [1:16] public as well as enhanced data sets that span  subsurface surface uh renewables data center [1:22] sighting what have you across the spectrum.  We also work a lot with structured as well as [1:27] unstructured data. In our unstructured data drove  uh specifically focus on land management. We have [1:31] about 350 million plus records dating all the  way back to 1800s uh which are critical for any [1:39] type of operation that you're working on. And my  objective of highlighting some of these metrics [1:43] for users um whichever uh vertical, whichever  organization, whichever domain your company sits [1:49] in, it is going to be impacted by the specific  nature of the data and information it pertains [1:55] to. Right? And that that's that is a challenge and  an opportunity that our organization has faced for [2:02] several years. I want to talk through some of the  methods that we adopted to solve for that. All [2:08] right, let's get right into it. So couple of stats  to ground the problem before I get into uh the [2:15] actual solution. These are not our stats. These  come from two really infamous reports uh that came [2:20] out a few months ago. One was by Gartner that says  40% of Agentic AI projects are cancelled by the [2:26] year 2027. Right? There's a lot of momentum.  There's a lot of energy. There's a lot of [2:31] excitement, but it's not reaching production. It's  not reaching scale. Right? It could be exceeding [2:37] costs. It could be unclear objectives. It could  be a variety of reasons but that is something [2:43] that's happening across enterprise AC you know  across the world and we sit in a B2B focused you [2:50] know working with enterprises directly sort of a  space in a highly highly regulated highly capital [2:55] intensive industry. So this becomes even more  more important and more serious in our space. [3:01] The other part of it which I'm sure a lot of you  heard about I'm guessing if you guys are here you [3:05] work in the AI space the infamous AI study the MIT  study I remember for a while I could not go into [3:12] a meeting with a customer without hearing about  the MIT study which said 95% of AI initiatives [3:18] never reach production right uh which is true and  there's a lot of reasons behind it but one of the [3:23] things that's very important to also understand  is the acceleration at the top of the funnel that [3:27] the AI revolution has caused right we went from  looking at 10 ideas and validating and picking [3:33] two to actually take to production to now we have  700 ideas that need to be figured out. Right? So [3:38] that also impacts these numbers. But there are  also other things that are important for us to [3:43] consider when we take things to production.  So first things first given that it's a rag [3:48] track retrieval is necessary for any enterprise  implementation. Uh I one lens that helps me think [3:56] about retrieval is retrieval to generative AI  implementation is what regression was to classical [4:02] ML implementation. Just because it's simple and  straightforward doesn't mean it doesn't work, [4:07] right? But it's important to understand how  we implement it because the nuances is what [4:12] make things complicated. Uh the other thing to  also remember is that as you're putting your [4:18] retrieval systems together with frontier models  and you know world-class horizontal AI systems [4:24] uh grounding it into the reality of your industry  and your space becomes critical. Uh prompting and [4:32] you know even using skills to some extent has  limits to how much you can affect and influence [4:37] the behavior of that model and that system to  pertain to what your industry needs. And so the [4:42] way we start started approaching this is looking  at the three frontiers on a agent or an AI system. [4:49] Right? I've got the model, I've got the data set,  and I've got the context layer. So I'm going to [4:55] talk about the model first since we just I just  covered some of the data pieces that our company [4:59] has been doing for a while. So the first thing we  started doing is how do we define the impact of [5:08] domain tuning your model, right? uh we we work in  the energy space. Our intention is not to create [5:14] claude for energy. Uh that given the amount  of information that sits in the public domain [5:19] and how powerful and reasoning structures these  models have, that's not where we win or lose this [5:25] uh conversation. But like with any industry,  we have a lot of jargon for one example, [5:32] right? we have a lot of very specific terminology  that I could either pass as a context layer file [5:38] each and every interaction that my system does  or identify what task it's executing at any [5:43] given point in the overall AI system. So we uh  tuned a small domain model and you can see in [5:50] uh in metrics and KPIs that are much more publicly  available uh like company ticker, company name, [5:57] it is very competitive with what the frontier  models are offering. But as I get into domains [6:02] that are much more specific to our industry, for  example, a geological formation, a play, a basin, [6:09] the gap starts to widen. And more often than not,  the specific metrics are going to be coming in [6:16] whether you're using it for internal systems or  your customers are using it. They're defining and [6:21] communicating things in very very different  ways. uh the other part of uh the other part [6:27] of u fine-tuning and domain during these models  which make retrieval specifically very important [6:33] uh which I think is often under um overlooked  is the strength of embedding models. Embeddings [6:39] define how your data manifests in your retrieval  ecosystem. Uh this is the comparison of several [6:46] different embedding models that we tuned on our  own corpus. uh compared to creating an embedding [6:52] model with a Titan v2 uh embedding system by  Amazon. This is by no means, by the way, a shot [6:59] against Titan V2's performance. As a generalist  system, that is probably one of the better, [7:04] if not best, embedding models out there. I've not  seen all the benchmarks with the new Gemini model, [7:08] but suffice to say, when I'm trying to build  a specific solution that applies to multiple [7:14] domains, then the green lines will not really  meet the criteria. But we see it in a very [7:20] specific domain where the understanding and  interreationship between what a lease name is, [7:26] a proprietor's names is, the formation it sits on  and the geology associated with it are words that [7:32] could sound like messy jargon to a standard  embedding structure. And the way those things [7:37] are embedding, those embeddings manifest not  only impact accuracy but also overall system [7:42] performance. uh if you look at the curves that  are out there the K1 value which represents if the [7:49] first chunk that was retrieved is accurate enough  sits at 56% for our best domain tuned model and [7:55] this is still an R&D initiative we're continuing  to push that further and further so and then as I [8:01] increase the number of K values that are retrieved  you can see the accuracy starts to flatten out or [8:06] you could say that well why don't I just retrieve  50 60 70 chunks and I'll most likely get the right [8:12] answer and these models are getting better at the  so-called needle in the haststack problem. Well, [8:17] while you could do that, the challenge other  challenge becomes is um you're creating a lot [8:23] of context bloat. You're also creating a lot of  latency. The model has to process 70 50 whatever [8:29] number of chunks each and every time. So, by  having a much more efficient embedding model, [8:34] you have a more accurate retrieval system. You  also have a much more faster retrieval system and [8:39] a more affordable retrieval system which goes back  to some of the reasons why you know the enterprise [8:46] adoption of uh these generative AI systems is  not as successful. So as you think about your [8:52] architecture that should define the implementation  and make the data that you have your mode, [8:56] there are four questions you should think about.  How are you grounding your information? Right? [9:02] How are you governing this framework? Because  that is just as important, right? Evaluations, [9:08] evaluations, evaluations. If you cannot  capture and quantify how a system performs, [9:14] which is increasingly hard as we go into more  and more complex styles of systems, we are adding [9:19] agent swarms, multi- aent orchestration, open  claw style agents that are building tools on the [9:24] fly. The more autonomy the agent have, the more uh  rigorous our evaluation framework needs to be. Uh [9:31] the other part is compose. This is the other part  where we believed a lot of the enterprise systems [9:37] fail because the amount of pilots that will start  off as a one-off, you know, an evaluation. Let's [9:43] go solve this problem. Oh, that's great. It worked  out. We can't really scale it to production. Now, [9:47] let's start a different problem. Now, let's start  to solve a different problem. And suddenly you [9:51] have 25 pilots which don't have any interconnected  system associated with it. And uh we've all seen [9:57] the news reports about token bills and things of  that nature, right? So, think about the overall [10:02] system first. even if it takes a second longer  to figure out how these pieces connect to each [10:06] other. And then finally control in a industry  like energy the final decision that has critical [10:14] capital and in a lot of cases human impact  needs to be a human making that decision and [10:19] I'll talk through a case study where we cover  that. So um as I mentioned uh this part before [10:26] uh what these different components are right so  grounding governance uh composition and control [10:32] and when you think about each of these pieces and  you build these systems those are components I [10:36] would highly recommend implementing again going  back to how you make your data your organization [10:42] and your system a truly differentiated asset as  you go through your AI transformation journey. So, [10:49] what I mean to say by the original title is that I  don't believe that your data is a moat by itself. [10:55] I had a conversation about a year and a half  ago with a senior executive at an energy company [11:01] who told me that they're just looking to connect  their massive oil and gas production database to [11:06] an LLM and that would solve all their problems of  doing all types of reporting. As you can imagine, [11:12] that did not come to fruition for a variety of  reasons. I think a lot of them were covered in [11:16] the last talk that you were here. It's because  architecting the system that services your [11:21] customers and your users the way they're supposed  to is far more important than whether you're using [11:26] Opus 56 or uh Fable 5 or whichever new model  there is. Right? Reliance purely on a particular [11:35] model's tool use and agent orchestration  abilities uh is not sufficient especially [11:40] when you're thinking about implementation in the  enterprise environment. So with that let's see [11:55] uh I don't this thing is not working. Oh thank  you. So that is a very uh brief diagram of [12:02] our architecture uh that we have implemented in  our system. I would not call it an architecture [12:07] diagram. I work in product and saying things  like that is why my tech teams are mad at me. [12:12] So this is a this is a very basic visual  representation of how we've stacked these [12:17] layers and they're designed semantically to  be able to be backwards compatible. So every [12:22] application in our system can use agents which  can use flows which can use skills and so on and [12:27] so forth and that interconnectedness helps us with  the composition problem. Any flow that anybody on [12:34] my team builds does not sit in isolation. It can  be used as a resource by the layer above it. It [12:40] can also be triggered outside through a composable  UI ecosystem which as an organization that's been [12:47] around for 25 years, we have a lot of those uh as  well as our customers are building a lot of their [12:51] own user interfaces. So that interconnection  becomes critical and it all flows through our [12:57] governance layer. So that whether you're trying  to connect to a data asset, an agent, a workflow, [13:01] an application, it all happens through the same  framework. And again the models whether it's are [13:08] uh you know proprietary models, frontier models,  they all fit into the same framework. So thinking [13:12] about a framework like this, I'm sure some of you  have a system similar to this slightly adjusted, [13:17] but the idea is that I think more often than not  in a lot of customers that I speak with, this step [13:23] gets skipped way too quickly. You just look at  the data and we look at the model and we look at [13:28] delivery, right? And that's where a lot of these  systems start to fail. So it's very important to [13:33] actually look at the whole system, look at those  four pillars that we spoke about and how they [13:37] connect to each other. Oh, sorry. All right. So,  uh, one example and case study of how we actually [13:45] implemented these cases. So, about a few months  ago, four of our customers, uh, big oil and gas [13:51] focused companies came to us and said that, you  know, AI is having a huge impact on how our teams [13:57] are doing analysis and how they're working in the  office, but a lot of the work in the oil field [14:02] happens in the oil field, right? How do we take  this technology to benefit that space and that [14:09] group of our persona? As you can imagine, working  in the oil field with a high risk environment, [14:15] the the safety criteria and threshold become  much more higher. And then it really brought [14:22] us back to that same governing principle, right?  How are we grounding this information? Me and my [14:28] agent accidentally giving wrong information to an  analyst pulling sales reports has a very different [14:34] cost associated with giving wrong safety advice  to a field worker working with heavy machinery [14:39] and equipment in the oil field. Right? So the  risk profile is very different. Governance is [14:44] critically important. We are not just talking  about organizational stats now. you're talking [14:49] about PII and you know personal safety incident  information and then how do we connect these [14:55] pieces together again the if you're going to make  behavioral changes and impact on the oil field [15:02] we're not going to have standalone implementations  this all all of these things need to fit together [15:07] and then finally what I mentioned initially the  control aspect we want the system to inform the [15:13] different teams that are working in the oil field  we don't want the system to make final decisions [15:19] Because no amount of training, embedding, tuning  beats field experience, right? These are these are [15:27] experts who've been doing this thing for decades  and working in the field for decades. They can [15:31] tell which operation needs to work in what cadence  given the temperature for the day. Right? So [15:37] that field value that organic knowledge is very  important to implement and it's very important to [15:42] add as a safety checkpoint. And so what you'll see  here is uh this is uh we call project lifesaver. [15:49] We just released the beta out uh to our consortium  of customers that we're working with. Those are [15:54] the uh companies on the left. Um and what you  see are the six barrels or circular layers of [16:00] the same stack of architecture that you saw a  couple of slides ago now implemented in this case [16:05] study. Right? So not only does the architectural  structure helps us differentiate the products and [16:12] technologies we bring to market, it allows us  to do it so much faster now because the design [16:17] and the framework is already in place. Right? So  if if there is a specific fe feature request or [16:23] a specific data asset that needs to be added, the  team doesn't have to think about how to bring it [16:28] in. It's already built into the framework. And so  this this as I mentioned is currently deployed and [16:33] we're getting really good feedback from the teams  and it's allowed us to do some really interesting [16:38] things with both the context layer, the model  layer as well as the data layer. So to close this [16:46] out, I would leave you guys with four things. As I  mentioned, you know, these examples that I shared [16:51] are from my industry, from the energy industry.  While some of you might be from it, I would assume [16:56] not all of you are from it. So think about what is  your own cliff right there is no point in building [17:03] you know uh a open AI or claude for financial  services model all lot of that information is [17:09] public where does that drop off happen and it  is it is in the specifics because gen in in [17:15] general terms there isn't a differentiation  needed for energy but when you think about [17:19] it energy professionals are not a collection of  energy information there are sequences of tasks [17:26] we all do when I work with somebody working  in the field and ask them a question, there's [17:30] about 25 decisions they make subconsciously  before even answering that question. Right? [17:34] Think about those things that differentiate how  your industry works versus the others. uh data [17:41] provenence and security by design uh is critical.  As as buzzwordy as they these things have become, [17:49] they have huge ramifications because once you set  the process in place, development and deployment [17:54] happens as a at a breakneck capacity. So if you  don't have those structures in place, adding them [17:59] later on is is much harder than actually planning  and thinking them through it. And as these systems [18:05] get more and more complex, we've gone from single  agents to multi- agents to open claw style tool [18:11] building agents to agent team frameworks and so  on and so forth. Identifying where the 17th agent [18:18] found that little fact that informed your model  is critical in validating and evaluating your [18:23] systems. Uh and then uh fund the second workflow.  This is this is really important. We work a lot [18:30] with customers directly and we always not ask  for like hey what's the most important thing [18:33] you need to do like actually talk about what the  whole process is. If you're building something [18:37] internally let's not just build a system that  helps let's say HR find some information faster. [18:42] What is the end result it's tying to? And once  we solve this problem what is the next problem [18:47] it's going to require to be solved. Do these  two problems connect to each other? Right? Like [18:51] actually doing that investigation is critically  important. And then finally name your human in [18:58] the loop human because that is a very important  step to identify the opportunity cost and the risk [19:04] associated especially depending on the risk  profile of where these models and frameworks [19:09] are implemented is way too high to avoid this  step. So identifying these four things as you [19:14] go back and identify you know implementations that  you're already doing maybe new things you want to [19:19] implement. Uh I I would question these four things  as you're thinking about those. And with that uh [19:27] I would uh I would leave you to it uh the uh and  leave you to to the lunch break. As I mentioned uh [19:32] your data is not your mode but it can become with  the right architecture. And yeah would love to [19:39] chat more. Uh me and some of my colleagues are in  the exhibition hall on booth 333. And if you agree [19:44] with me great if you disagree with me would also  love to hear those ideas. So, thank you so much.