Enterprise Finance AI Needs More Than RAG
企业财务 AI，需要的不只是 RAG
What should remain structured, what belongs in embeddings, and how ontology can give an AI agent a governed enterprise world model.
什么必须保留为结构化事实，什么适合进入 Embedding，以及 Ontology 如何为 AI Agent 建立一个受治理的企业世界模型。
If an enterprise wants an AI agent to reason about finance, the first design question is not “Which model should we use?” It is: what kind of evidence is this, and what is the agent allowed to do with it?
A ledger entry, a contract clause, a forecast note, and a historical accounting memo can all be relevant to the same decision. But they should not be stored, retrieved, or trusted in the same way.
One financial world, four representations
The choice is not “database or vector database.” Most enterprise finance knowledge needs several representations, each with a different job.
当企业希望 AI Agent 能够理解财务问题时，第一个设计问题不应该是“选哪个模型”，而应该是：这是什么类型的证据，Agent 被允许如何使用它？
一条总账记录、一项合同条款、一段 Forecast 说明和一份历史会计备忘录，可能同时与同一个决策相关。但它们不应该以相同方式保存、检索和信任。
同一个财务世界，四种表示
这不是“数据库还是向量数据库”的二选一。大多数企业财务知识需要多种表示，每一种承担不同职责。
Compute and audit
- Transactions, balances, journals
- Actuals, budgets, forecasts
- Entities, accounts, periods
- Approvals and control status
Interpret and retrieve
- Policies and contract clauses
- Accounting memos and explanations
- Forecast assumptions
- Meeting notes and precedents
用于计算与审计
- 交易、余额与分录
- 实际、预算与预测
- 实体、科目与期间
- 审批记录与控制状态
用于理解与检索
- 政策与合同条款
- 会计备忘录与解释
- 预测假设
- 会议记录与历史先例
Structured data answers what happened and how much. Natural language helps explain why it may have happened and what context matters. A knowledge graph expresses what the item is connected to. Embeddings help retrieve semantically similar language.
An embedding should not decide the balance, test whether debit equals credit, calculate a variance, or enforce an approval threshold. Those jobs belong to governed data, deterministic code, and rule engines.
Ontology is the enterprise meaning layer
A model can retrieve the words “professional services,” but ontology tells the agent what that concept means inside the company: which accounts compose it, which legal entity owns it, which vendor and contract are involved, which policy governs it, and which actions are permitted.
结构化数据回答发生了什么、金额是多少；自然语言帮助解释为什么可能发生、哪些上下文重要；知识图谱表达它与谁、什么项目、合同和规则有关；Embedding 用于找出语义相似的文本。
Embedding 不应该决定余额、检查借贷是否相等、计算差异或执行审批阈值。这些职责属于受治理的数据、确定性代码与规则引擎。
Ontology 是企业意义层
模型可以检索到“Professional Services”这几个词，但 Ontology 告诉 Agent：这个概念在企业里具体代表什么，由哪些科目构成，属于哪个法人实体，涉及哪个供应商与合同，受哪项政策约束，以及允许采取什么行动。
A practical example: why did spend exceed budget?
Suppose someone asks: “Why is professional-services expense 35% over budget this month?” The agent should not answer from a document summary alone.
- It calculates actual, budget, and variance from governed facts.
- It follows graph relationships to vendors, purchase orders, projects, contracts, and departments.
- It retrieves forecast notes and contract scope to explain likely causes.
- It applies deterministic controls for approvals, duplicate invoices, or capitalization review.
- It separates verified facts from inference and recommended action.
一个可落地的例子：为什么费用超预算？
假设有人问：“为什么本月 Professional Services 费用比预算高 35%？” Agent 不应该只依据一段文档摘要给出答案。
- 从受治理的结构化事实中计算 Actual、Budget 与 Variance。
- 沿知识图谱追踪供应商、采购订单、项目、合同与部门。
- 检索 Forecast notes 与合同范围，解释可能原因。
- 通过确定性控制检查审批、重复发票或资本化条件。
- 明确区分已验证事实、推断与建议动作。
The architecture I want to test in practice
The point is not to let an LLM replace the ERP. It is to let an agent reason on top of a governed enterprise world model—understanding facts, meaning, relationships, policies, and permitted actions without collapsing them into one undifferentiated answer.
This is where I believe enterprise AI becomes useful: not when it sounds confident, but when its reasoning is traceable, its evidence is typed, and its actions are controlled.
我真正希望进入企业验证的架构
目标不是让 LLM 取代 ERP，而是让 Agent 在一个受治理的企业世界模型之上进行推理：理解事实、意义、关系、政策与允许的动作，同时不把它们混成一个无法追溯的答案。
我相信，企业 AI 真正产生价值的时刻，不是它听起来足够自信，而是它的推理可以追踪、证据类型清晰、行动受到控制。