当理解和使用业务知识的主体变成 Agent，企业本体应该如何支撑业务理解与行动？让我们先从企业每天都在面对的问题看起。
本文源自火山引擎知识引擎 VeKnow 业务实践思考，火山引擎知识引擎 VeKnow 作为企业级 AI 知识底座，依托大模型与大数据技术，提供从知识接入、加工处理、知识建模到智能问答的全链路知识管理与应用能力，让企业知识真正流动起来，在业务中持续创造价值。
一、企业本体，表达的是怎样一个业务世界？
两千多年前，亚里士多德在《形而上学》里追问过一个“存在之为存在”的问题，哲学家把这类追问叫作“本体论”（Ontology）：世界上究竟有什么、以何种方式存在、彼此之间又存在怎样的关系。
同样的问题，企业每天都在问。售后要搞清一次报修工单涉及哪台设备、装了哪些零部件、改过什么配置、以前修过几次、这次换件会不会牵连到其他部件；风控要知道“客户”背后挂着哪些账户、这些账户之间的转账有没有异常关联。每个岗位、每套系统，背后都在回答同一组问题：业务世界里有哪些对象、它们怎么联系、受什么规则约束。
这套对业务世界的表达，就是“企业本体”。
比如电商订单号、商品、金额、收件人等数据，以及订单与商品的关联，业务系统里都能存下来。但想要真正让 Agent 理解并用在业务里，还要知道：几件商品是不是同店打包发货、退款对应整单退还是单品退、订单享受的折扣优惠规则是什么，这些跨系统的业务含义、规则适用范围和版本信息，往往分散在数据、代码与文档中。
企业本体要做的，正是把这些对象、关系与约束统一表达，让人和 Agent 都能在一致的业务语义下使用它们。
过去，企业要把业务世界的关系网搭起来，往往需要业务专家和工程师反复拉通，读懂材料、定义对象、补齐关系、跟上变更，构建就难，持续运转更难，很多企业最后留下的是一堆各自为政的数据表，和写在 PPT 里的方法论。
Agent Native 时代的到来，第一次让这件事有了新的解法。大模型驱动的 Agent 承担了大量的阅读、整理与步骤组织，知识工程逐项依赖人工的成本结构被动摇，人的精力可以更多放在目标、标准与关键判断上。
业务世界需要被表达，这个命题没有变；构建和维护这份表达的方式，正在被重新改写。
二、找到一条规则，不代表它适用于当前业务
那么，接入一个通用 Agent，企业本体就能自动建好、持续用起来吗？
比如零售客服场景下，用户问“这笔订单能不能退货？”，通用 Agent 调工具、读文档，回复“符合 7 天无理由退货，可以申请退款”，实际上，这笔订单来自一次直播专场的限时活动，走的是第三方渠道，商品属于定制类目，退换货规则在活动开始前刚刚更新过一版。Agent 确实找到了一条“退换货规则”，但没有同时识别销售渠道、商品类别、活动批次和规则生效时间，最终引用了一条相关、却并不适用于当前订单的规定。
出错的不是模型，而是它背后没有一张能表达“规则—渠道—类目—活动—时效”真实关系的本体可用，看得见条款，看不见适用边界；能生成答案，无法校验。
企业每天跑在 Agent 上的决策越多，本体缺失带来的隐性错误就被放大得越厉害，代价也越高。
火山引擎知识引擎 VeKnow 产品，以 Agent 能持续使用的方式来帮助企业构建本体，提供了一套可交付、可协作、可持续优化的知识编译能力，让业务目标、Agent 执行和人机协作进入同一条链路，也让企业本体从“被构建的结果”逐步变成“持续运行的系统”，具体来说，这套产品化能力落在三个核心点上。
第一个能力，是把产品能力做成 Agent 可组合的工具。 知识编译并不是一条固定页面流程，面对不同业务目标，Agent 可能需要以不同顺序组合资料读取、Schema 生成、实例抽取、资产查询、版本更新和评测等能力。VeKnow 通过 CLI 把这些能力封装为可调用、可组合的工具，让 Agent 能够根据目标、当前状态和反馈，自主规划下一步，而不必为每一种场景重新开发一套固定流程。
CLI 同时为执行过程提供了可控边界：关键动作可以显式确认，任务状态和中间结果可以保留，失败步骤可以局部重试，中断任务可以基于既有上下文续做。开发者负责稳定供给能力，Agent 负责组织工作，人负责目标与关键判断，三者共同构成一套面向真实交付的 Agent Harness。
第二个能力，是自动化建本体。 在传统项目里，业务专家与本体建模工程师需要反复沟通，逐项完成材料解读、对象定义、关系设计、实例填充和规则配置，整个过程既慢又重。VeKnow 把这些工作重新组织成 Agent 可执行的知识编译任务：系统读取权限范围内的企业知识，识别业务对象、属性、关系与约束，生成 Schema 提案，并进一步完成实例抽取与本体构建。
自动化并不意味着跳过业务判断。业务专家负责定义目标和边界，在 Schema、关键关系与冲突项等节点进行审核；Agent 根据修改意见更新上下文并继续执行；技术性建模由系统承担，人集中处理业务语义和例外问题，既降低了跨角色协同成本，也让每一项结果都保留了与源材料的对应关系，方便复核和追溯。
第三个能力，是让每一次优化都能被评测验证。 本体版本发生变化，并不天然代表质量提升。VeKnow 把评测嵌入知识编译闭环：先围绕业务目标确定评价范围、评测集、标准答案和评价维度，再将当前产物与标准对照，定位对象、关系、术语或实例层面的差距，并结合源材料分析原因、生成优化建议。
修订完成后，新版本会回到同一套基准上复测，既检查预期问题是否修复，也观察其他部分是否出现退化。评测集、标准答案、产物版本和结果随资产一起沉淀，质量变化可度量、问题定位可追踪、优化经验可复用，企业本体就能在真实业务反馈中持续演化。
三、让 Action 成为 Agent 的行动语言
理解业务之后，Agent 要开始动手做事。
如果只是给本体对象挂上一堆函数，再让 Agent 按图索骥去调用，解决的仅仅是“入口在哪里”，企业真正关心的是另一组问题：Agent 能不能理解当前能做什么、为什么能做、做完会怎样，并基于这些判断去组织行动。
以“取消订单”为例，未发货时，可以直接取消并退款；已发货时，可能要先拦截物流或进入退货流程。退款失败，订单不能被标记为“已退款”；拦截失败，后续路径也要随之改变。同一个“取消订单”，背后其实是一张随状态而变的决策图。
未来，Action 要做的，就是把一次业务状态变化封装成一份可执行的契约：前置条件决定能不能做，授权边界决定谁能做，预期影响定义会改变什么，失败处理约束异常如何收场。Agent 负责理解目标、规划路径和组合动作；系统负责校验条件、执行操作、验证结果并回写状态。
两者合起来，Agent 才能真正走进企业的业务流程，在规则之内做得更准，在复杂场景里走得更稳。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
How should the enterprise itself support business understanding and action when the subject of understanding and use of business knowledge becomes Agent? Let us begin with the problems that business is facing every day.
This paper is derived from the Volcengine knowledge engine VeKnow business practice thinking, and the Volcengine knowledge engine VeKnow, as the enterprise-level AI knowledge base, builds on LLM and big data technology to provide a full-link knowledge management and application capability from knowledge access, processing, knowledge modelling to smart questions and answers, allowing for a real flow of business knowledge and continuous value creation in the business.
I. What business world does the business have in mind?
More than 2,000 years ago, Aristotle asked a question about “the existence of the existence”, which philosophers call “Ontology”: what is in the world, how does it exist and how does it relate to each other?
The same question is asked daily by enterprises. A copy of the manifest is to be obtained on sale of which equipment, which spare parts have been installed, which configuration has been modified, how many repairs have been made before, whether the change of office will involve other components, and whether the “client” has an unusual connection to the accounts behind them. Each post, each system is followed by the same set of questions: what are the targets in the world of operations, how they relate, and what rules govern them.
This expression of the world of business is the business nature.
Data such as electrical order numbers, commodities, amounts, recipients, and the connection between orders and goods are stored in the business system. But to really get Agent to understand and use it in business, you know: Whether the goods are packaged from the same shop, the refund of the entire item is a single refund, and what are the discounts for orders, are the cross-system business implications, the scope of application of the rules and the information on the version, often scattered in data, codes and documents.
What the enterprise itself does is to express these objects, relationships and constraints in a uniform way, so that they can be used in consistent business terms.
In the past, companies have often needed business specialists and engineers to connect, read and understand materials, define objects, complete relationships, keep up with changes, make construction difficult and continue to operate, and many end up leaving separate data sheets and methodologies written in PPT.
The advent of the Age Native led to a new solution for the first time. Agent, driven by LLM, carries a great deal of reading, sorting and process organization, and intellectual engineering has been shaken by the artificial cost structure, and human energy can be more focused on objectives, standards and key judgements.
The business world needs to be expressed, and the proposition remains unchanged; the way it is constructed and maintained is being rewritten.
Finding a rule does not mean that it applies to current operations
So, access to a universal Agent can automatically be built and used on a continuous basis?
In the context of retail passenger service, for example, users ask, “Can the order be returned?”, the generic Agent modulation tool, reading the document, and replying that “there is no reason to return the goods for seven days, and can apply for a refund”, in fact, the order comes from limited-time, a live broadcast, using a third-party channel, a custom category of goods, and the rules for the return of the goods were updated just before the event began. Agent did find a “replacement rule”, but did not identify both the channels of sale, the category of goods, the batch of activities and the time when the rules became effective, eventually citing a relevant provision that did not apply to current orders.
It is not the model that is wrong, but the body behind it is not one that can express the true relationship between the “rules-channel-category-activity-proscription”, can see the terms, cannot see the applicable boundary; and can generate answers that cannot be verified.
The more companies make decisions on Agent every day, the more hidden mistakes the absence of the body is magnified and the higher the cost.
The Volcengine Knowledge Engine VeKnow product, which helps enterprises build their own core in a sustainable manner, provides a set of deliverable, collaborative, sustainable and optimized knowledge compilation capabilities that allow business objectives, Agent execution and human collaboration to enter the same chain, and allows the enterprise's home to evolve from a “constructed result” to a “continuing system”, specifically at three core points.
The first is to make product capacity a combination of tools. Knowledge compilation is not a fixed page process and, in the face of different business objectives, Agent may require a combination of information reading, Schema generation, sample extraction, asset queries, version updates and assessment capabilities. VeKnow is encapsulating these capabilities through CLI as a callable, combustible tool that allows Agent to plan next step autonomously according to objectives, current status and feedback, without having to redevelop a fixed set of processes for each scenario.
CLI also provides a manageable boundary for the implementation process: key actions can be clearly identified, mission status and intermediate results can be preserved, failure steps can be partially retry, and interruptions can be based on existing context. The developers are responsible for stabilizing supply capacity, Agent is responsible for organizing, people are responsible for objectives and critical judgements, and together they constitute a set of Agent Harnesses for real delivery.
The second capability is automation. In traditional projects, business specialists and home-based modeling engineers need to communicate repeatedly, with material reading, object definition, relationship design, case filling and rule configuration completed on a case-by-case basis, and the process is slow and heavy. VeKnow reorganizes these efforts into Agent's enforceable knowledge compilation tasks: systematic access to business knowledge within permission, identification of business objects, attributes, relationships and constraints, generation of Schema proposals, and further completion of case extraction and home construction.
Automation does not mean skipping business judgement. Operational experts are responsible for defining objectives and boundaries, conducting audits at points such as Schema, Key Relations and Conflicts; Agent updates and continues to implement context on the basis of modifications; and technical modelling is carried out by the system, where people focus on operational semantics and exceptions, both reducing the cost of cross-role synergies and allowing each outcome to retain its correspondence with source material to facilitate review and traceability.
The third capability is to allow each optimization to be evaluated and validated. The change in this version does not represent a natural increase in quality. VeKnow embedding the assessment into the knowledge compilation closed loop: first, setting the scope of the evaluation, the assessment set, the standard answers and the evaluation dimensions around the operational objectives, then comparing the current product with the standard, locating the gaps at the level of the object, relationship, terminologies or examples, and combining the source material to analyse the causes and generate recommendations for optimization.
Once the revision has been completed, the new version will revert to the same set of benchmarks to check whether the expected problem has been repaired and whether the rest has been degraded. Assessments, standard answers, product versions and results are deposited with assets, quality changes are measurable, problem positioning is traceable, optimized experience is reused, and the enterprise itself can evolve in real business feedback.
iii. Make Action the action language of Agent
After understanding business, Agent has to start working.
If it's just putting a bunch of functions on its own object, and then letting Agent call by Tusop, the solution is just "Where's the entrance," and the firm's real concern is another set of questions: Does Agent understand what he can do, why he can do it, how he can do it, how he can do it, and on the basis of these judgements, organize his actions.
In the case of “cancellation of orders”, the goods may be cancelled and refunded directly if they have not been shipped; when they have been delivered, they may have to be intercepted or entered the return process. Refunds failed, orders could not be labelled as “refunded”; interception failed and subsequent paths changed. The same `cancellation orders' is behind a decision-making map that changes with the state.
In the future, Action is about encapsulating a change in business status as an enforceable contract: pre-conditions determine whether or not to do it, empower borders to decide who can do it, anticipate changes in the definition of impact, and how failure to deal with unusual constraints ends. Agent is responsible for understanding objectives, planning paths and grouping actions; the system is responsible for verifying conditions, executing operations, verifying results and writing back.
Together, Agent can truly enter the business processes of the enterprise, do more precisely within the rules and move more steadily in complex scenarios.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.