Trusted Data Is Where Enterprise AI Begins
可信数据，才是企业 AI 真正的起点
Why governance-qualified data and a governed semantic layer are the foundation for enterprise AI ontology, reasoning, and trusted action.
为什么治理合格数据与受治理的语义层，是企业构建 AI Ontology、可信推理与可控行动的起点和基石。
A semantic layer is more than a BI modeling layer. It is a shared business language across teams, tools, dashboards, notebooks, APIs—and increasingly, AI agents.
As enterprises enter the AI era, attention naturally goes to LLMs, agents, knowledge graphs, and automated decisions. But the capability that determines whether AI can scale is more fundamental: does AI understand the business through data that is trusted, consistent, and governed?
If the underlying data is unstable, “revenue” has three definitions, and business relationships are ambiguous, AI does not remove the problem. It amplifies it faster and with more confidence.
Start with business questions, not the tool
When building a semantic layer, I would not begin with a product selection. I would begin with the decisions the organization needs to support:
- Which decisions matter most?
- Which KPIs are critical to those decisions?
- Where is metric drift already damaging trust?
- Which concepts must be shared by people, systems, and agents?
From there, the work becomes structured: define core entities, build clean foundations, standardize metrics, add governance, optimize performance, connect consumers, and make the layer AI-ready.
Clean data is not governance-qualified data
Most teams can land data in a lake or warehouse. The harder work is turning “raw” into something consistently trustworthy and reusable.
Governance-qualified data needs ownership, upstream data contracts, validation and monitoring, lineage, freshness, permissions, and documented fitness for purpose. A revenue field with no null values may still be unusable if no one knows whether it means booked, billed, or recognized revenue.
The semantic layer is now part of the intelligence layer
Historically, the value was straightforward: prevent “revenue” from meaning one thing in Looker, another in Tableau, and a third in a notebook. A governed semantic layer creates one place for metrics, business logic, dimensions, ownership, and reusable KPI governance.
With AI agents, that value becomes larger. An agent cannot reliably explain why revenue changed or what the business should do next if it does not know what revenue means, which time window applies, and which evidence is authoritative.
AI ontology connects facts, meaning, and action
Governance-qualified data provides trusted facts. The semantic layer gives those facts shared business meaning. Ontology then connects entities, metrics, events, rules, and possible actions so machines can reason in business terms.
For an LLM or analytics agent to reason correctly, it needs more than raw tables. It needs metric meaning, business context, lineage, evidence, permissions, and confidence. Agent intelligence should not stop at “What happened?” It should help explain why, show the evidence, express uncertainty, and recommend the next appropriate action.
Why did “Chloe” and “China” return the same Top 5?
为什么 “Chloe” 和 “China” 返回了相同的 Top 5？
I once tested retrieval in a graph database built around a movie domain, similar to a Netflix-style catalog. I entered two very different queries—“Chloe” and “China”—and received exactly the same Top 5 results.
The graph database was not necessarily broken. It may have executed the query perfectly. The deeper problem was that the data model could not express the semantic difference between the two requests.
The system matched strings, not business meaning
“Chloe” might be an actor, character, director, title, or user. “China” might mean country of production, filming location, market, language, or subject. Without entity typing and context, both become text tokens rather than distinct business concepts.
The graph had connections, but weak relationship semantics
acted_in, filmed_in, produced_in, and available_in carry different meaning. Flatten them into a generic related_to, and highly connected popular movies dominate every query.
Embedding similarity is not business relevance
Embeddings can retrieve statistically similar content without respecting entity type, relationship direction, or user intent. A vector score can look precise while the business answer is wrong.
A mature system would first resolve “Chloe” as an Actor and traverse acted_in → Movie. It would resolve “China” as a Country and choose among produced_in, filmed_in, or distributed_in based on intent. If the results still overlap, the system should show why, through which evidence path, and with what confidence.
我曾经在一个类似 Netflix 电影目录场景的 Graph Database 中做过检索实验。我分别输入了两个完全不同的关键词——“Chloe” 和 “China”——系统却返回了完全相同的 Top 5 结果。
图数据库不一定出了故障。它可能非常准确地执行了查询，只是底层数据模型无法表达两次请求之间的语义差异。
系统识别了字符串，却没有识别业务含义
“Chloe” 可能是演员、角色、导演、电影名称或用户；“China” 可能代表制片国家、拍摄地、发行市场、语言或题材。缺少实体类型和上下文时，它们只是两个文本 Token，而不是两个含义不同的业务概念。
图中有连接，但关系语义太弱
acted_in、filmed_in、produced_in 和 available_in 的含义完全不同。如果全部简化为 related_to，高连接度的热门电影就可能主导每一次查询。
向量相似不等于业务相关
Embedding 可以找到统计上相近的内容，却未必尊重实体类型、关系方向和用户意图。相似度分数看起来很精确，业务答案却可能是错误的。
成熟的系统应该先把 “Chloe” 解析为 Actor，再沿着 acted_in → Movie 检索；把 “China” 解析为 Country，再根据意图选择 produced_in、filmed_in 或 distributed_in。即使最终结果有重合，系统也应该说明为什么重合、证据路径是什么，以及置信度有多高。
The ROI is larger than BI consistency
The immediate return is fewer duplicated models, less metric debate, and faster analytics delivery. The strategic return is a reusable enterprise meaning system: one metric can serve multiple BI tools; one entity can be understood by APIs, models, and agents; one governance policy can constrain access, reasoning, and automated action.
Without this foundation, every new AI application must relearn the business. With it, business knowledge becomes shared infrastructure.
From pipelines to a system of trust
The real shift is not simply from warehouses to lakehouses, or from dashboards to agents. It is from moving data to serving governed data products—and from producing answers to supporting decisions.
- Trusted data: the facts are reliable.
- Trusted metrics: definitions remain consistent.
- Trusted context: meaning, lineage, and permissions are explicit.
- Trusted reasoning: evidence paths and uncertainty are visible.
- Trusted action: recommendations are bounded and accountable.
Enterprise AI transformation does not begin with buying more AI tools. It begins when data can be trusted, business language can be shared, and machines can understand the entities, relationships, rules, and evidence behind a decision.
语义层早已不只是 BI 建模层。它是团队、工具、Dashboard、Notebook、API，以及越来越多 AI Agent 之间共享的业务语言。
进入 AI 数据时代，企业很自然地把注意力放在大模型、智能体、知识图谱与自动化决策上。但真正决定 AI 能否规模化落地的，是一个更基础的问题：AI 所理解的业务，是否建立在可信、统一、可治理的数据之上？
如果底层数据不稳定，“收入”存在三种定义，业务关系含义模糊，AI 不会消除这些问题，只会以更快的速度、更确定的语气放大它们。
不要从工具开始，要从业务问题开始
建设语义层时，我不会先选择工具，而会先回答：
- 哪些决策最需要被支持？
- 哪些 KPI 对这些决策最关键？
- 指标漂移正在什么地方伤害信任？
- 哪些业务概念必须被团队、系统和 Agent 共同理解？
从这里出发，工作才会变得结构化：定义核心业务实体、建设干净的数据基础、统一指标、加入治理、优化性能、连接消费端，并最终让语义层具备 AI-ready 能力。
清洗过的数据，不等于治理合格的数据
大多数团队都能把数据送进数据湖或数仓。真正困难的，是把 Raw Data 转化为可以被持续信任和复用的数据产品。
治理合格数据需要明确的数据所有权、上游数据契约、质量验证与监控、完整血缘、数据时效、权限边界与适用范围。一个没有空值的收入字段，如果没人知道它代表签约、开票还是确认收入，仍然不能支持可靠决策。
语义层正在成为 Intelligence Layer 的一部分
过去，语义层最直接的价值，是避免“收入”在 Looker、Tableau 和 Notebook 中分别代表不同含义。受治理的语义层为指标、业务逻辑、维度、所有权与 KPI 治理建立统一来源。
在 AI Agent 时代，它的价值更大。如果 Agent 不知道收入到底是什么、适用哪个时间范围、什么证据最权威，它就无法可靠地解释收入为什么变化，更无法建议企业下一步应该做什么。
AI Ontology 连接事实、含义与行动
Governance-Qualified Data 提供可信事实，Semantic Layer 为事实提供共享的业务含义，Ontology 再把实体、指标、事件、规则与行动连接起来，让机器能够用业务语言推理。
LLM 或分析型 Agent 要正确推理，需要的不只是原始表，还需要指标含义、业务上下文、血缘、证据、权限与置信度。Agent Intelligence 不应该停留在“发生了什么”，还应该解释为什么、展示证据、表达不确定性，并提出合适的下一步行动。
语义层的 ROI 远大于 BI 一致性
它最直接的回报，是减少重复建模、降低指标争议、加快分析交付。更重要的战略回报，是建立可复用的企业语义系统：一个指标服务多个 BI 工具，一个实体被 API、模型和 Agent 共同理解，一套治理规则同时约束访问、推理和自动化行动。
没有这套基础，每一个新的 AI 应用都要重新学习企业业务；有了它，业务知识就能成为共享基础设施。
从 Pipeline 走向 System of Trust
真正的变化，不只是从数仓走向 Lakehouse，也不只是从 Dashboard 走向 Agent，而是从“搬运数据”走向“交付受治理的数据产品”，从“生成答案”走向“支持决策”。
- Trusted Data：事实可信。
- Trusted Metrics：定义一致。
- Trusted Context：含义、血缘与权限清晰。
- Trusted Reasoning：证据路径与不确定性可见。
- Trusted Action：建议有边界、行动可追责。
企业 AI 转型的起点，不是购买更多 AI 工具，而是让数据能够被信任，让业务语言能够被共享，让机器理解每一个决策背后的实体、关系、规则与证据。