Euno Unveils Automated Context Graph to Curb Enterprise AI Failures

CIG: Complete, Instant, Governed Context for Enterprise AI With Euno — Ai4

Ai4

Connecting AI agents directly to enterprise databases fails because manual semantic documentation degrades rapidly, requiring automated metadata lineage and governance to reach production accuracy.

AI-assisted summary of CIG: Complete, Instant, Governed Context for Enterprise AI With Euno — original source. Check the original for full context.

Source published: 2026-08-17

Detailed notes

## Background Published on YouTube by Ai4 on August 17, 2026, this presentation features Evan Fleisch, Head of Sales at Euno. The session addresses why connecting AI agents directly to enterprise data platforms like Snowflake or Databricks fails without deep business context, introducing Euno's Enterprise AI Context Platform to unify lineage, governance, and semantic definitions across data stacks.

## TL;DR Enterprise AI agents routinely fail when plugged directly into private databases because they lack the implicit institutional knowledge, operational quirks, and edge-case awareness needed to interpret structured data. Attempts to manually maintain semantic layers or Markdown documentation create a Sisyphean burden of stale artifacts and degraded performance. To succeed, enterprise context must be automated, strictly governed, and intelligently scoped to eliminate latency and security risks. Euno implements this through a multi-tiered Context Pyramid—grounded in an automated metadata knowledge graph with layered governance, glossaries, persona scoping, and guardrails—to push agent accuracy from an 80% baseline to over 95% within two weeks.

## Atmosphere & style The session carries the crisp, pragmatic energy of a fast-paced enterprise tech brief, cutting briskly through general AI hype to pinpoint messy operational realities. Fleisch speaks with direct, experienced authority, framing manual data-engineering workflows as a cautionary Sisyphean trap while humorously calling out familiar pains like the dreaded eight-minute spinning wheel. The delivery feels purposeful, architectural, and grounded, moving steadily from enterprise frustration to a tidy, systematic blueprint for governed reliability.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=4s) AI agents perform well on open web data, but hooking them directly to enterprise databases via basic MCP servers often fails due to a lack of business context. - [00:04](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=4s) The bottleneck in enterprise AI adoption is not raw model intelligence, but access to organizational nuances, history, and operational edge cases. - [01:24](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=84s) While unstructured search tools handle documents effectively, structured data queries require deep context hidden in transformation code, lineage, and usage patterns. - [01:24](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=84s) Manually maintaining semantic layers and static Markdown documentation for AI agents creates a Sisyphean loop where documentation rapidly becomes stale. - [02:54](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=174s) Enterprise AI context must be automated, governed, and scoped to deliver scalable value without devolving into a glorified chat interface for legacy BI. - [02:54](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=174s) Without proper scoping, agents attempting downstream impact analysis overload context windows, leading to multi-minute latency and execution timeouts. - [03:53](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=233s) Euno's automated foundation crawls lineage, metadata, and existing semantics into a knowledge graph, establishing baseline accuracy around 80% on day one. - [03:53](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=233s) Achieving over 95% enterprise accuracy requires stacking automated governance, business glossaries, persona scoping, and strict guardrails on top of the metadata graph. - [03:53](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=233s) Persona and departmental scoping restrict agent visibility to relevant domains, tailoring access differently for roles like product engineering versus sales. - [03:53](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=233s) Enterprise guardrails enforce data freshness cutoffs and redact PII across downstream BI layers to ensure safe agent execution.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=4s) Context vs. Raw Intelligence in the Enterprise Walled Garden

Evan Fleisch introduces the fundamental disconnect between public AI agent capabilities and their performance inside corporate environments. While AI agents perform remarkably well on the open internet—a challenge Evan previously tackled at Tavily by pulling down live web results—bringing those same agents inside the enterprise "walled garden" tells a very different story. When organizations attempt to plug agents into their enterprise cloud environments using standard Model Context Protocol (MCP) servers, the immediate outcome is often an outright disaster.

The breakdown occurs not from a lack of raw model intelligence, but from an absence of context. Intelligence alone does not make an enterprise employee effective; the most valuable team member is the one who understands business nuances, historical context, operational quirks, and edge cases. Transforming an AI agent into an enterprise-ready asset requires moving beyond baseline intelligence to make the agent deeply knowledgeable and context-aware.

```mermaid flowchart TD accTitle: The Context Gap in Enterprise AI accDescr: Moving from raw LLM intelligence to context-grounded enterprise AI agents.

subgraph LLM["Raw Agent Intelligence"] A["LLM Base Model"] --> B["Standard MCP Connections"] end

subgraph Problem["Enterprise Failure"] B --> C["Hallucinations & Broken Context"] end

subgraph Solution["Euno Context Foundation"] D["Lineage, Governance & Semantics"] E["Domain & Persona Scoping"] D --> F["Context-Grounded Enterprise Agent"] E --> F end ```

- **The Enterprise Disconnect**: Agents excel on public web data but fail inside private enterprise systems when hooked directly up to databases without context. - **The Core Differentiator**: Enterprise value is driven by understanding organizational edge cases and institutional knowledge rather than raw model IQ.

---

[01:24](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=84s) The Pitfall of Manual Semantic Layers and Stale Markdown

Reflecting on tools like Glean, Evan notes how natural language processing initially unlocked unstructured documents and files across organizations. However, enterprise structured data introduces an entirely different challenge. The context required to navigate structured data already exists implicitly across the stack—hidden "between the lines" in query usage patterns, transformation code, asset ownership, and data freshness metrics.

Despite this existing metadata, organizations repeatedly attempt to bridge the gap through manual, Sisyphean efforts. Data teams spend endless hours attempting to hand-map every question to an answer in semantic layers or drafting static Markdown documentation for AI agents to consume. Because humans are forced to manually "feed the robots," these documentation files are not treated as first-class citizens and become stale almost immediately, causing agent accuracy to degrade rapidly.

- **Unstructured vs. Structured Data**: Enterprise search successfully handled documents, but structured database queries require deep contextual knowledge of code and schema relationships. - **The Sisyphean Semantic Trap**: Manually crafting ontologies and 1:1 question-to-semantic-layer mappings fails because manual artifacts go stale in moments. - **The Implicit Pseudo-Semantic Layer**: Critical operational context already lives within system metadata, lineage, code history, and asset ownership.

---

[02:54](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=174s) Core Requirements for Enterprise AI Context: Automated, Governed, Scoped

For AI agents to function sustainably on enterprise data, the contextual framework must satisfy three non-negotiable criteria: it must be automated, governed, and strictly scoped. Relying on human intermediaries turns an agent into nothing more than a "pretend AI"—essentially a traditional business intelligence tool retrofitted with a chat interface, providing zero scalable enterprise value.

Context must also be strictly governed to prevent unauthorized access, incorrect downstream consumption, and data leaks. Finally, context must be intelligently scoped. Without scoping, connecting agents to broad MCP endpoints results in the dreaded "infinite spinning wheel" where an LLM spends eight minutes or more attempting to parse downstream impact analysis across massive schemas, overloading context windows and failing to return timely answers.

```mermaid flowchart LR accTitle: Pillars of Enterprise Context accDescr: Three core pillars required for enterprise AI agents: Automated, Governed, and Scoped.

subgraph Pillars["Context Requirements"] P1["Automated Context"] --- P2["Governed Data Access"] --- P3["Scoped Query Context"] end

Pillars --> OUT["Reliable, High-Speed Agent Execution"] ```

- **Automated**: Removes manual human-in-the-loop bottlenecks to prevent agents from devolving into glorified BI chatbots. - **Governed**: Guarantees safety, prevents access to unverified assets, and enforces enterprise security boundaries. - **Scoped**: Prevents context-window bloat and multi-minute latency by providing only the exact downstream metadata necessary for the specific task.

---

[03:53](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=233s) The Euno Context Pyramid: Achieving >95% Enterprise Accuracy

Euno productizes the contextual groundwork—such as the manual methodologies outlined in Anthropic's research—into an automated, multi-tiered Context Platform. At the foundation sits a metadata knowledge graph that crawls enterprise lineage, metadata, and existing semantic definitions in a couple of hours. This automated baseline provides roughly 80% accuracy across the enterprise on day one.

```mermaid flowchart TD accTitle: The Euno Context Pyramid accDescr: The tiered layers of Euno's platform that lift agent accuracy from 80% to over 95%.

subgraph Accuracy95["Target: >95% Accuracy in < 2 Weeks"] L5["Guardrails: Freshness limits (>60d cutoff), PII redaction"] L4["Personas & Skills: Departmental scoping (Product vs. Sales)"] L3["Glossary & Business Logic: Confluence, catalogs, KPI definitions"] L2["Automated Governance: Asset safety, data health, certification"] L1["Foundation: Unified lineage, metadata & semantics (Knowledge Graph)"] end

L1 --> L2 L2 --> L3 L3 --> L4 L4 --> L5 ```

Because an 80% success rate is still a failing grade for enterprise decision-making, Euno stacks four additional governance and contextual layers on top of the knowledge graph to reach over 95% accuracy in less than two weeks:

1. **Unified Lineage, Metadata, and Semantics**: The baseline crawler that connects the entire data estate into a unified knowledge graph within hours. 2. **Automated Governance Rules**: Automated verification identifying which data assets are certified, fresh, and safe for AI consumption. 3. **Glossary and Business Logic**: Integration of existing business catalogs, Confluence documentation, and metric definitions. 4. **Personas and Skills**: Departmental and role-based scoping (e.g., granting broader system access to product engineers while scoping sales representatives to commercial domains). 5. **Enterprise AI Guardrails**: Strict execution guardrails, such as ignoring datasets that have not been refreshed within 60 days or blocking PII propagated downstream to BI layers.

By implementing this automated contextual pyramid, organizations can avoid the "Sisyphean joke" of building endless manual ontologies and successfully answer the long tail of enterprise AI queries.

## Notable quotes > Simply connecting agents to platforms like Snowflake or Databricks isn't enough. The result is often hallucinations, inconsistent answers, broken trust, and a growing governance burden that can't keep pace with the speed of AI. — Evan Fleisch

> [00:04](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=4s) The breakdown occurs not from a lack of raw model intelligence, but from an absence of context. — Evan Fleisch

> [00:04](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=4s) Intelligence alone does not make an enterprise employee effective; the most valuable team member is the one who understands business nuances, historical context, operational quirks, and edge cases. — Evan Fleisch

> [00:04](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=4s) Transforming an AI agent into an enterprise-ready asset requires moving beyond baseline intelligence to make the agent deeply knowledgeable and context-aware. — Evan Fleisch

> [01:24](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=84s) Data teams spend endless hours attempting to hand-map every question to an answer in semantic layers or drafting static Markdown documentation for AI agents to consume. — Evan Fleisch

> [01:24](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=84s) Because humans are forced to manually 'feed the robots,' these documentation files are not treated as first-class citizens and become stale almost immediately, causing agent accuracy to degrade rapidly. — Evan Fleisch

> [02:54](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=174s) For AI agents to function sustainably on enterprise data, the contextual framework must satisfy three non-negotiable criteria: it must be automated, governed, and strictly scoped. — Evan Fleisch

> [02:54](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=174s) Relying on human intermediaries turns an agent into nothing more than a 'pretend AI'—essentially a traditional business intelligence tool retrofitted with a chat interface, providing zero scalable enterprise value. — Evan Fleisch

> [02:54](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=174s) Without scoping, connecting agents to broad MCP endpoints results in the dreaded 'infinite spinning wheel' where an LLM spends eight minutes or more attempting to parse downstream impact analysis across massive schemas. — Evan Fleisch

> [03:53](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=233s) At the foundation sits a metadata knowledge graph that crawls enterprise lineage, metadata, and existing semantic definitions in a couple of hours. This automated baseline provides roughly 80% accuracy across the enterprise on day one. — Evan Fleisch

> [03:53](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=233s) Because an 80% success rate is still a failing grade for enterprise decision-making, Euno stacks four additional governance and contextual layers on top of the knowledge graph to reach over 95% accuracy in less than two weeks. — Evan Fleisch

## Mentioned Evan Fleisch, Euno, Claude, Cursor, Snowflake, Databricks, Tavily, Model Context Protocol (MCP), Glean, Anthropic, Confluence

## Video description Evan Fleisch - Head of Sales, Euno

In 2026, organizations expect AI to just work on their enterprise data. But simply connecting agents to platforms like Snowflake or Databricks isn't enough. The result is often hallucinations, inconsistent answers, broken trust, and a growing governance burden that can't keep pace with the speed of AI.

In this session, we'll show how Euno customers take a different approach. By leveraging Euno's Enterprise AI Context Platform, they enable agents like Claude and Cursor to work accurately and safely on top of enterprise data, consistently querying the right data, applying the right business logic, and returning reliable, explainable results.

AI agents need context to understand enterprise data, and that context is fragmented across the modern data and AI stack. Through a live walkthrough, we'll show how Euno automatically unifies lineage, ownership, governance, semantic definitions, and institutional knowledge across warehouses, ETL/ELT pipelines, observability platforms, security tools, BI platforms, and more. That trusted context is then surfaced directly to the AI agents your teams already use.

This session is designed for data and AI leaders who are deploying AI on top of their core data platforms to accelerate decision-making across both technical and business teams. You'll leave with a practical blueprint for grounding AI agents in complete, governed, real-time context, and an emerging enterprise standard for making AI work reliably on enterprise data.

[Source](https://www.youtube.com/watch?v=q-yIVW_b-0k)

中文简报

Euno发布自动化上下文图谱，遏制企业AI失误将AI Agent直接连接至企业数据库往往以失败告终，因为人工编写的语义文档会迅速退化，需要依靠自动化的元数据血缘与治理才能达到生产级的准确率。企业AI Agent在直接接入私有数据库时常常失败，因为它们缺乏解释结构化数据所需的隐性机构知识、运营特性和边缘场景认知。试图通过手动维护语义层或Markdown文档来解决问题，会造成文档过时和性能下降的西西弗斯式重担。为了走向成功，企业上下文必须实现自动化、严格治理并进行智能化范围界定，从而消除延迟和安全风险。Euno通过多层级的“上下文金字塔”（Context Pyramid）来实现这一点——该架构基于自动化元数据知识图谱，并辅以分层治理、业务词汇表、角色范围界定和安全护栏，可在两周内将Agent的准确率从80%的基线提升至95%以上。 AI Agent 在公网表现优异，但若仅通过标准 MCP 服务直连企业数据往往会引发严重错误；企业落地的核心瓶颈在于缺乏业务隐性知识与运营背景，而非模型的基础智力。 (0:04) 企业结构化数据的上下文隐性存在于转换代码、数据血缘和使用模式中；人工手动维护语义层或静态 Markdown 文档注定会迅速过时，导致 Agent 准确率迅速衰退。 (1:24) 企业级 AI 上下文必须满足自动化、强治理与严格限定范围（Scoped）三大核心要求，否则 Agent 会因分析大规模模式导致上下文窗口超载，出现长达数分钟的延迟与超时。 (2:54) Euno 平台在数小时内自动抓取血缘与元数据构建知识图谱，首日即可实现约 80% 的基线准确率，随后通过叠加自动化治理、角色限定与严格护栏，在两周内将准确率提升至 95% 以上。 (3:53)

Reading summary

Evan Fleisch asserts that enterprise AI agent initiatives fail not from deficient model intelligence, but from a total lack of automated, operational context within corporate data environments. Context Gaps and the Pitfalls of Static Documentation (0:04) According to Fleisch, while AI agents handle public web data effectively, connecting them directly to enterprise cloud environments via standard Model Context Protocol (MCP) servers frequently causes breakdowns. The issue is not model intelligence, but an inability to interpret business-specific nuances, operational quirks, and historical institutional knowledge. Enterprise search tools previously addressed unstructured documents, but structured data context remains hidden within query usage patterns, code transformations, and asset lineage. Fleisch notes that enterprises often attempt to bridge this divide through manual efforts, such as hand-mapping queries into semantic layers or writing static Markdown files. Because these resources require continuous human updates, they are rarely maintained as first-class operational assets, quickly turn stale, and ultimately degrade agent accuracy. Architectural Mandates: Automated, Governed, and Scoped Context (2:54) To prevent agents from devolving into brittle conversational wrappers over business intelligence tools, Fleisch outlines three mandatory criteria: context must be automated, governed, and scoped. Automated ingestion eliminates reliance on human intermediaries, while enterprise governance ensures that agents access only certified, secure datasets and respect corporate security perimeters. Scoping is equally critical to prevent operational latency. Without strictly targeted context, connecting agents to broad schemas can trigger multi-minute processing delays—such as eight-minute impact analyses—that overload context windows and stall task execution. The Tiered Context Pyramid for High Enterprise Accuracy (3:53) Euno productizes contextual discovery by stacking five progressive layers over an automated metadata knowledge graph. According to the company, crawling existing lineage, metadata, and semantics provides an initial baseline of approximately 80% accuracy within hours. However, enterprise workflows demand higher reliability to be viable for decision-making. To exceed 95% accuracy in under two weeks, the platform adds automated governance validation, business glossaries from sources like Confluence, role-based persona scoping (such as distinguishing sales access from product engineering), and strict operational guardrails. These guardrails enforce security and data quality by blocking personally identifiable information (PII) from downstream business intelligence layers and automatically excluding data assets that have not been refreshed within 60 days. Fleisch concludes that automating context generation replaces unsustainable manual ontology mapping, enabling AI agents to handle the long tail of enterprise queries while remaining bound to rigorous freshness and access boundaries.

精炼阅读摘要

Evan Fleisch 指出，AI 智能体在企业私有云中直接接入数据库往往导致运行混乱，其根本症结在于模型缺乏业务上下文而非基础智力不足；企业必须摆脱低效的手动语义维护，采用自动化、受治理且限定范围的上下文平台。企业围墙花园内的上下文断层 (0:04) Fleisch 认为，在公网环境表现优越的 AI 智能体一旦通过标准 Model Context Protocol（MCP）服务器直接接入企业私有云环境，往往会遭遇严重失败。这种故障并非由于基础大模型的智力水平低下，而是源于上下文的缺失。在企业环境中，高价值员工的核心竞争力在于掌握业务的细微差异、历史背景、操作惯例和边缘特例。直接将智能体连入未加修饰的数据表，无法提供这些关键的机构知识，只会引发幻觉与上下文断裂。人工语义层与静态文档的维护陷阱 (1:24) 企业结构化数据与自然语言文档的处理逻辑存在本质差异。结构化数据所需的关键业务逻辑与上下文，其实已经隐式存在于查询使用模式、转换代码、资产归属和数据新鲜度等系统元数据之中。然而，许多企业依然耗费大量人力，试图在语义层中为每个问题手工映射答案，或编写静态 Markdown 文档供智能体调用。Fleisch 批评这种“人工喂养机器人”的做法是西西弗斯式的无用功，因为手工维护的文档极易过时，进而直接导致智能体查询准确率严重下滑。自动化、受治理与范围限定三大准则 (2:54) 为了让企业 AI 智能体保持稳定运行，上下文框架必须满足自动化、受治理以及严格限定范围三个核心要素。若依赖人工介入，智能体将退化为套着聊天界面的传统商业智能（BI）工具。此外，上下文体系必须具备严格的治理能力，以阻断未认证资产并防止数据泄露。同时，系统必须精确限定上下文的投递范围；若直接暴露宽泛的 MCP 端点，大模型将被迫解析庞大的底层下游依赖，不仅会耗尽上下文窗口，还会导致长达 8 分钟以上的响应延迟。 Euno 五层上下文金字塔的实施路径 (3:53) Euno 将相关上下文构建策略工程化为一个分层的自动化平台。底层元数据知识图谱通过爬取企业数据血缘、元数据和既有语义，可在数小时内建立基础连接，初始提供约 80% 的准确率。由于该准确率仍无法满足企业决策要求，平台进一步叠加了四层机制：用于检验资产认证与健康度的自动化治理规则、整合 Confluence 与 KPI 定义的业务词汇表、基于部门（如产品工程与销售）的权限与技能范围划分，以及执行安全护栏（如排除超过 60 天未刷新的陈旧数据、拦截流向下游的 PII 信息）。Euno 称该方案可在两周内将智能体准确率提升至 95% 以上。通过自动构建元数据知识图谱并叠加治理防护栏，企业有望摆脱手工维护本体的沉重负担并应对长尾查询。不过，对于缺乏完整操作日志、代码记录与元数据沉淀的遗留系统，该自动化上下文框架的实际适应能力仍有待进一步验证。

Detailed notes — zh

## Background 该视频由 Ai4 于 2026 年 8 月 17 日发布在 YouTube 上。根据页面描述，Euno 销售主管 Evan Fleisch 在本场演讲中介绍了如何借助 Euno 的企业 AI 上下文平台，统一跨数据仓库、ETL 管道与 BI 平台的血缘与元数据，为 Claude 和 Cursor 等 Agent 提供完整、受治理的上下文，从而彻底解决企业数据上的幻觉与信任问题。 ## TL;DR Evan Fleisch 指出，模型智力并不能直接转化为企业生产力，单纯人工维护语义层和元数据已经沦为徒劳的“西西弗斯陷阱”。要让 Enterprise Agent 可靠落地，必须满足高度自动化、严格受治理与精准限定范围（Scoped）三大核心要求。Euno 通过构建五层上下文金字塔——从底层自动抓取血缘与隐式语义，到叠加治理规则、业务术语表、角色技能划分及安全护栏，帮助企业在两周内将 Agent 在核心数据上的回答准确率从 80% 提升至 95% 以上。 ## Atmosphere & style 整场演讲风格轻松幽默且直击行业痛点。面对台下人数不多的小型研讨会场，讲者以邀听众上台同坐的打趣开场，用生动的“聪明员工 vs 优秀员工”类比和犀利的吐槽打破了企业数据架构的枯燥沉闷。叙事节奏明快，既有对业界“人工喂养机器”现状的辛辣解构，又有切实落地的技术蓝图，最后在邀大家去喝一杯、去威尼斯人酒店碰运气的调侃中干脆利落地收尾。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=4s) 开放网络与企业内部环境存在本质差异，AI Agent 在企业“围墙花园”内直接直连云数据库往往会导致失效与灾难。 - [00:04](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=4s) 企业级 Agent 的瓶颈不是基础大模型的通用智力，而是缺乏对企业专有业务逻辑、隐式规则和边缘情况的深度上下文理解。 - [01:24](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=84s) 结构化数据所需的上下文隐式分散在代码、血缘关系与日常调用中，形成了企业内部碎片化的“伪语义层”。 - [01:24](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=84s) 依靠人工编写 Markdown 和硬编码映射来维护语义层是“西西弗斯式的苦役”，规则迅速过时会导致 Agent 频繁产生幻觉。 - [02:54](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=174s) 企业级 Agent 落地必须具备三大基础支柱：高度自动化（Automated）、严格受治理（Governed）以及精准限定范围（Scoped）。 - [02:54](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=174s) 如果 Agent 背后仍需大量人工介入清洗和提示词干预，本质上只是给传统 BI 工具套了一个聊天壳，无法实现规模化扩展。 - [02:54](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=174s) 缺乏精准作用域限定（Scoping）会导致 Agent 陷入元数据与工具调用的计算过载，引发严重的执行超时与效率低下。 - [03:53](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=233s) Euno 上下文金字塔将原本耗费大量人力的血缘与语义统一流程彻底产品化，旨在将企业回答准确率从 80% 提升至 95% 以上。 - [03:53](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=233s) 知识图谱自动抓取血缘与元数据可在数小时内提供 80% 的基础解析准确率，但面对严肃生产环境，80% 的准确率依然等同于不及格。 - [03:53](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=233s) 上下文金字塔向上逐层融合自动化治理规则、业务术语表（Glossary）以及按部门严格划分的角色与技能范围（Personas & Skills）。 - [03:53](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=233s) 顶层的企业级 AI 护栏（Guardrails）负责强制实施安全规范，可自动拦截时效性超期数据及渗透到 BI 展现层的 PII 敏感信息。 - [03:53](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=233s) 采用自动化上下文生成平台替代耗时数月的手工本体论构建，是让企业 AI Agent 高效、可靠解答长尾业务问题的关键路径。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=4s) 企业 AI Agent 的核心困境：聪明有余而缺乏企业业务上下文 Euno 的销售主管 Evan Fleisch 在开场时面对台下人数不多的听众打趣道：“到了这个阶段，你们其实完全可以跟我一起坐到台上来了……我们这十来个人正好能来一场热烈而兴奋的探讨。”他指出，在开放的互联网环境中，AI Agent 的表现已经非常出色，例如他几年前在 Tavily 所解决的问题——直接抓取并处理公开网络数据。然而，一旦把这一套搬进企业内部的“围墙花园”（walled garden），情况就急转直下。如果你在一家大型企业工作，直接接上几个 MCP（Model Context Protocol）工具，连上云端数据库并期望它直接返回可靠结果，其结局往往是“一场灾难”。这种失败并不是因为 AI 模型本身的智商不够。Evan Fleisch 强调，业内一直在反复讨论模型的能力上限，但企业真正需要的往往不是单纯的智力。他向听众抛出一个类比：“在你们公司里，最聪明的人是谁？大概率就是坐在台下的各位；但谁才是**最优秀的员工**？”最优秀的员工往往是那些掌握全部上下文、洞悉所有边缘情况（edge cases）、并且对业务运行逻辑有深刻理解的人。当前构建 Enterprise Agent 的核心痛点，正在于如何将 Agent 从一个单纯“高智商的推理源”，转变为一个“既聪明、又懂业务、且通晓全局上下文的可靠实体”。 - **互联网与企业内部环境的割裂**：开放 Web 上的信息检索已基本解决，但企业内部异构系统让原生 Agent 屡屡碰壁。 - **“聪明员工”与“优秀员工”的区别**：模型智力已足够强大，缺失的关键拼图是对企业专有业务逻辑与边缘场景的深度上下文理解。 --- ### [01:24](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=84s) 结构化数据的断层与人工维护语义层的“西西弗斯陷阱” 谈到企业数据的整合，Evan Fleisch 引用了一个哈利·波特的经典梗图来形容 Glean 刚问世时的惊艳感：“那就像魔法一样，把所有文档文件连起来，非结构化数据简直是 LLM 的天生绝配。”但当面对结构化数据时，真正的麻烦才刚刚浮出水面。事实上，理解结构化数据所需的上下文并非不存在，它们化作一种“伪语义层”（pseudo-semantic layer）隐式藏在企业代码、日常数据调用、资产归属权以及新鲜度状态的字里行间。企业并不缺少这些元数据，缺少的是将它们系统化抽取出来的元数据底座。然而，企业目前的解决手段却陷入了西西弗斯式的苦役（Sisyphean task）——试图人工将业务问题与数据语义层进行一对一的硬编码映射。数据团队构建了执行层和语义层，满怀信心地认为这套方案万无一失，但现实是这些规则很快就会过时。Evan 尖锐地指出：“我们现在本质上是在靠人工喂养机器（feeding robots by humans）。”工程师们写下一份份未被视作一等公民的 Markdown 说明文档，结果短短几分钟后就彻底与生产环境脱节。指望在 Agent 调用的每一个环节都靠人工介入去维护上下文，不仅根本走不通，也彻底宣告了 Agent 自动化的失败。 ```mermaid flowchart TD accTitle: 传统人工维护与自动化上下文平台的模式对比 subgraph Bad["传统方式：西西弗斯式的苦役"] direction TB B1["人工编写 Markdown 文档 / 映射规则"] --> B2["业务逻辑与表结构变更"] B2 --> B3["语义层迅速陈旧过时 (Stale)"] B3 --> B4["Agent 获取错误上下文并产生幻觉"] end subgraph Good["Euno 方式：自动化元数据与上下文体系"] direction TB G1["底层知识图谱自动抓取血缘与元数据"] --> G2["自动化治理与业务术语表对齐"] G2 --> G3["基于角色权限 (Personas) 动态限定范围"] G3 --> G4["安全护栏过滤 (时效性/PII) 并输出可靠结果"] end ``` --- ### [02:54](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=174s) 落地企业级 Agent 的三大立足基石：自动化、受治理与精确范围控制 Evan Fleisch 随后阐明了 Euno 打造企业级 Agent 上下文平台的底层逻辑。要让 Agent 在企业复杂业务中稳定运行，必须具备三项不可妥协的基础能力（table stakes）： 1. **高度自动化（Automated）**：上下文必须自动化捕获与更新。如果 Agent 背后依然需要大量人工来清洗、标注和提示，那么“这无非是给传统 BI 工具套了一个聊天界面（chat interface）而已”。这种伪 AI 既无法横向扩展，也根本无法为企业释放实际价值。 2. **严格受治理（Governed）**：缺乏治理的 Agent 在企业内极度危险。它可能会无差别访问未经授权的数据资产，甚至引用错误的报表来做决策，因此自动化上下文必须内置安全策略与合法性校验。 3. **精准限定范围（Scoped）**：用过 Claude 并接入 MCP 的人都体会过那种令人绝望的“无限旋转加载指示圈”——明明只是问了一个简单的问题，系统却在背后调用大量无序工具和元数据，直接转了整整 8 分钟；对于需要进行下游影响分析（impact analysis）的工程师而言，这种计算过载更是灾难。Agent 的上下文绝不能超载，必须根据具体任务严格框定作用域。 --- ### [03:53](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=233s) Euno 上下文金字塔：从 80% 的及格线迈向 95% 以上的企业级准确率在演讲的最后，Evan Fleisch 展示了 Euno 上下文平台的核心架构金字塔，并向台下一位正在拍照的观众致意：“感谢那位拍照的朋友，我很感激。”他提到，Anthropic 此前曾发表过一篇优秀的博客，详细阐述他们如何靠纯人工把血缘、元数据和语义统一起来；而 Euno 的核心突破，就是将这套极度耗费人力的流程完全**产品化**。 ```mermaid flowchart TD accTitle: Euno 上下文金字塔架构 L5["企业 AI 护栏 Guardrails (例如：数据新鲜度 > 60 天则禁用、BI 层 PII 拦截)"] L4["业务角色与技能 Personas & Skills (按产研、销售等业务域严格划分访问范围)"] L3["业务术语表 Business Glossary (融合 Confluence 文档、传统数据目录、业务指标定义)"] L2["自动化治理规则 Automated Governance Rules (评估数据资产安全性、新鲜度与可用性)"] L1["统一血缘、元数据与隐式语义 Unified Lineage & Semantics (知识图谱自动爬取，数小时内完成建模，提供 80% 基础准确率)"] L5 --> L4 L4 --> L3 L3 --> L2 L2 --> L1 ``` - **底层基础：统一血缘、元数据与语义（Unified Lineage, Metadata & Semantics）**：利用知识图谱自动爬取全套技术栈中的元数据，企业通常在一天乃至几小时内就能完成全系统的血缘构建。这一层能够提供约 80% 的企业数据解析准确度。但在严肃的企业生产环境中，“80% 的准确率就等同于不及格（fail）”。 - **第二层：自动化治理规则（Governance Rules）**：自动评估并标定哪些数据是安全的、新鲜的、且可被 AI 直接消费的。 - **第三层：业务术语表（Glossary & Business Logic）**：整合传统数据目录与 Confluence 文档中的业务指标、计算公式及核心定义，让机器理解指标背后的真实业务含义。 - **第四层：角色与技能范围（Personas & Skills）**：根据不同部门的业务域进行精准的作用域隔离（Scoping）。例如，产研团队与销售团队应分配截然不同的访问上下文与工具权限。 - **顶层：企业级 AI 护栏（Guardrails）**：强制落地企业级安全规范。例如，若某张报表的数据时效性已超过 60 天，直接阻止 Agent 读取；一旦检测到敏感个人信息（PII）穿透传导到了 BI 展现层，坚决拦截访问。 Evan 总结道，通过这套分层金字塔，企业可以在两周之内将 Agent 在核心数据上的回答准确率推高到 95% 以上。行业中很多人依然在花费数月手工拼凑语义层、死磕本体论（ontologies），在他看来那纯粹是“一场西西弗斯式的笑话”。随着更多人转向这种自动化的上下文生成方案，AI 长尾业务问题的解答才真正具备可行性，否则企业团队将永远在徒劳地“推石上山”。随后他感谢全场听众，幽默地宣布演讲结束，大家可以去喝一杯、吃点东西，或者去威尼斯人酒店碰碰运气了。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=4s) 到了这个阶段，你们其实完全可以跟我一起坐到台上来了……我们这十来个人正好能来一场热烈而兴奋的探讨。 — Evan Fleisch > [00:04](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=4s) 在你们公司里，最聪明的人是谁？大概率就是坐在台下的各位；但谁才是最优秀的员工？ — Evan Fleisch > [01:24](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=84s) 那就像魔法一样，把所有文档文件连起来，非结构化数据简直是 LLM 的天生绝配。 — Evan Fleisch > [01:24](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=84s) 我们现在本质上是在靠人工喂养机器（feeding robots by humans）。 — Evan Fleisch > [02:54](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=174s) 这无非是给传统 BI 工具套了一个聊天界面（chat interface）而已 — Evan Fleisch > [03:53](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=233s) 感谢那位拍照的朋友，我很感激。 — Evan Fleisch > [03:53](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=233s) 80% 的准确率就等同于不及格（fail） — Evan Fleisch > [03:53](https://www.youtube.com/watch?v=q-yIVW_b-0k&t=233s) 一场西西弗斯式的笑话 — Evan Fleisch ## Mentioned Evan Fleisch, Euno, Tavily, Glean, Anthropic, Claude, Cursor, Snowflake, Databricks, Confluence, Model Context Protocol (MCP) ## Video description Evan Fleisch - Head of Sales, Euno In 2026, organizations expect AI to just work on their enterprise data. But simply connecting agents to platforms like Snowflake or Databricks isn't enough. The result is often hallucinations, inconsistent answers, broken trust, and a growing governance burden that can't keep pace with the speed of AI. In this session, we'll show how Euno customers take a different approach. By leveraging Euno's Enterprise AI Context Platform, they enable agents like Claude and Cursor to work accurately and safely on top of enterprise data, consistently querying the right data, applying the right business logic, and returning reliable, explainable results. AI agents need context to understand enterprise data, and that context is fragmented across the modern data and AI stack. Through a live walkthrough, we'll show how Euno automatically unifies lineage, ownership, governance, semantic definitions, and institutional knowledge across warehouses, ETL/ELT pipelines, observability platforms, security tools, BI platforms, and more. That trusted context is then surfaced directly to the AI agents your teams already use. This session is designed for data and AI leaders who are deploying AI on top of their core data platforms to accelerate decision-making across both technical and business teams. You'll leave with a practical blueprint for grounding AI agents in complete, governed, real-time context, and an emerging enterprise standard for making AI work reliably on enterprise data. [Source](https://www.youtube.com/watch?v=q-yIVW_b-0k)

Source quotes

[0:04] Evan Fleisch: 到了这个阶段，你们其实完全可以跟我一起坐到台上来了……我们这十来个人正好能来一场热烈而兴奋的探讨。 [0:04] Evan Fleisch: 在你们公司里，最聪明的人是谁？大概率就是坐在台下的各位；但谁才是最优秀的员工？ [1:24] Evan Fleisch: 那就像魔法一样，把所有文档文件连起来，非结构化数据简直是 LLM 的天生绝配。 [1:24] Evan Fleisch: 我们现在本质上是在靠人工喂养机器（feeding robots by humans）。 [2:54] Evan Fleisch: 这无非是给传统 BI 工具套了一个聊天界面（chat interface）而已 [3:53] Evan Fleisch: 感谢那位拍照的朋友，我很感激。 [3:53] Evan Fleisch: 80% 的准确率就等同于不及格（fail） [3:53] Evan Fleisch: 一场西西弗斯式的笑话

Source quotes

Evan Fleisch: Simply connecting agents to platforms like Snowflake or Databricks isn't enough. The result is often hallucinations, inconsistent answers, broken trust, and a growing governance burden that can't keep pace with the speed of AI. [0:04] Evan Fleisch: The breakdown occurs not from a lack of raw model intelligence, but from an absence of context. [0:04] Evan Fleisch: Intelligence alone does not make an enterprise employee effective; the most valuable team member is the one who understands business nuances, historical context, operational quirks, and edge cases. [0:04] Evan Fleisch: Transforming an AI agent into an enterprise-ready asset requires moving beyond baseline intelligence to make the agent deeply knowledgeable and context-aware. [1:24] Evan Fleisch: Data teams spend endless hours attempting to hand-map every question to an answer in semantic layers or drafting static Markdown documentation for AI agents to consume. [1:24] Evan Fleisch: Because humans are forced to manually 'feed the robots,' these documentation files are not treated as first-class citizens and become stale almost immediately, causing agent accuracy to degrade rapidly. [2:54] Evan Fleisch: For AI agents to function sustainably on enterprise data, the contextual framework must satisfy three non-negotiable criteria: it must be automated, governed, and strictly scoped. [2:54] Evan Fleisch: Relying on human intermediaries turns an agent into nothing more than a 'pretend AI'—essentially a traditional business intelligence tool retrofitted with a chat interface, providing zero scalable enterprise value. [2:54] Evan Fleisch: Without scoping, connecting agents to broad MCP endpoints results in the dreaded 'infinite spinning wheel' where an LLM spends eight minutes or more attempting to parse downstream impact analysis across massive schemas. [3:53] Evan Fleisch: At the foundation sits a metadata knowledge graph that crawls enterprise lineage, metadata, and existing semantic definitions in a couple of hours. This automated baseline provides roughly 80% accuracy across the enterprise on day one. [3:53] Evan Fleisch: Because an 80% success rate is still a failing grade for enterprise decision-making, Euno stacks four additional governance and contextual layers on top of the knowledge graph to reach over 95% accuracy in less than two weeks.

Original transcript

[0:04] My name is Evan Flesh. I mean, you all are just like welcome to come on stage with me [0:07] at this point. Um, so I'm going to talk to you about ASG. Uh, it's actually a really [0:13] really interesting problem whether it's relevant for you or not, which is we're all using AI agents [0:18] and it works beautifully on the internet. That's a problem that I solved with a company called [0:23] Tilli just a few years ago saying, "Okay, we're going to pull down results. Fantastic. What about [0:29] inside the walled garden of the enterprise, right? How do we make an AI agent there work? [0:34] If you're working at a big company, what's going to actually happen when you put in a few [0:39] MCPS, plug it into your cloud, and try and get a result. The result is a disaster, right? And it's [0:45] not because the AI is not intelligent enough. It is. People have talked about that over and over. [0:51] But think about in your company, who's the most intelligent person? It's it's probably you, [0:57] right? But what about the mo the best employee? The best employee is the one who has the context, [1:03] understands the edge cases, understands the business itself in a deep deep way, [1:09] right? And that's what we're trying to solve for with an agent is how do you turn an agent [1:13] from an intelligent source to an intelligent knowledgeable and understanding of context source. [1:19] So that's what we're here to talk about. All all 10 of us going to have an exciting chat about it. [1:24] So, if you're like me, you're a Harry Potter fan, which you recognize this meme. And this to me is [1:29] Glean when it came out. The magic of Glean. You could connect all of your files. Hooray. [1:33] Agents. Wonderful. Unstructured data. Perfect for LLMs. But what about the structured data? [1:40] That's where the problem comes in. Yet, the context for them, how they work, [1:44] what they should know, it already exists. It's a pseudossemantic layer that's already there. [1:50] It's between the lines. It's the usage. It's the code. It's the ownership. Is it [1:54] fresh? Is it not fresh? And all of these will make an agent work on the enterprise. [1:59] And it exists. And what you need is a metadata layer to get there. And that metadata layer [2:05] doesn't need to wait. Since it's already there, you can start today. [2:10] But how are people doing it now, right? And how do you fill the gap between what's going [2:13] now? It's a sysophician task. It's you're pushing a rock up a hill of [2:19] mapping one to one questions to answers on a semantic layer. You have an execution layer, [2:25] your semantic layer and you are confident it is going to work and it is stale. We are feeding [2:30] robots by humans. We are making markdown files that we are not treating as first class citizens [2:35] that becomes that become stale in moments. That's the current reality across the enterprise. [2:43] Right? So assuming that there's a human in the loop is a huge problem. [2:49] It's not working and it's not working for agents. [2:54] This is why I'm here to talk very briefly now about the actual approach that UNO does, [2:58] right? Is that an agent needs a few different things in order to be successful? It needs to be [3:04] automated, right? The context must be automated because if you're doing humans behind an agent, [3:10] what's the point? You're just making a new BI tool with a chat interface. That's not fun. It's not [3:15] scalable. It doesn't extend. You're getting zero of the value. It's pretend AI. It must be governed [3:22] because otherwise it could be unsafe. You're going to access all the information I need not list that [3:27] could be wrong. So in order to actually activate this, the governing needs to be there and then it [3:31] needs to be scoped because we've all used claude. If you've connected it to MCPs, you know that [3:36] infinite spinning wheel of claude, you've asked a simple question and it's gone on eight minutes. [3:41] Let alone for engineers what that takes for an impact analysis. We're going downstream, [3:46] right? So never overloading and scoped. These are the fundamental table stakes to how to get there. [3:53] So that's where we came in and we built the context platform for how agents work. [3:57] And this is my last slide before we all get some whatever we do after this drinks or food or [4:03] gamble at the Venetian. And what you're seeing here is and thank you for the guy [4:08] picking pictures. I appreciate you. Um, so we have you what what represents the bottom and by [4:15] the way there's an amazing blog post by anthropic about how they did this by hand and this is the [4:19] part that we productized this clear part on the bottom unified lineage metadata and semantics [4:24] right which gets you to an 80% answer across an enterprise and you could set this up in one day [4:29] at this point with with what we've leveraged it's a knowledge graph that goes on crawls [4:34] all of your metadata literally like in a couple hours you have full lineage across your system 80% [4:38] accuracy hooray. Hey, that's a fail, right? From there, it's what you put on top of it. That's the [4:43] basis of what makes you successful, right? You need governance rules that need to be automatic, [4:48] saying what is what is safe, what is fresh, what is useful for AI. Then on top of that, a glossery, [4:55] this is more of the traditional data catalog stuff that still needs to be there. If you have a [4:58] confluence pages, it's saying what are the actual definitions? What are the business logic? What are [5:03] those questions? Integrate that. Then on top of that, it's personas and skills, right? who should [5:09] be accessing what and all the details of that the details of the department the details of of [5:16] the details of if it's a domain if the domain is correct so is somebody from product sure give [5:21] them wider access are they from sales sales access that's part of scoping and then finally the guard [5:27] rules whatever those enterprise AI rules might be saying hey we have freshness rules if this is over [5:32] 60 days we don't want to access it if it's um any kind of PII that's at uh that's propagated down [5:39] to the BI layer. Don't access it. This pyramid leads you to what's over 95% accuracy across the [5:47] enterprise and it honestly it could be done in less than two weeks. We're in a world out there [5:51] where people are still building semantic layers where people are building ontologies and frankly [5:56] it's a sysophician joke and more people will be aligning to this to get the li to get the long [6:01] tale of AI questions answered and until then we'll be pushing rocks up hills. Thank you everybody.