据金蝶2026年中期业绩披露，灵基发布后签约26家客户、上线40余个Agent，AI原生产品收入同比增长189%。到本次更新前，灵基付费客户已达3706家。
期间，金蝶集团总裁章勇、金蝶中国副总裁兼研发中心总经理刘仲文围绕产品定位、行业变局等进行分享。其核心思路是将ERP从记录系统升级为Agent的认知底座，让Agent带着企业既有的业务逻辑、权限边界和历史习惯去干活。
金蝶的AI尝试早于灵基的发布。2023年8月，金蝶发布苍穹GPT大模型，彼时的思路是将大模型能力嵌入ERP与财务场景，推出预算规划、代码生成等AI助手。到了2024年，金蝶苍穹升级为新一代企业级AI平台，这一阶段验证了AI在企业管理场景中的可行性，但同时也暴露出一个问题：Agent如果没有企业既有的业务逻辑和上下文，很难真正干活。
如今来看，软件行业里的共识已经在发生变化：过往企业软件以权限为驱动，为满足企业每个角色的使用价值，企业软件只能无限广地开发和组合系统功能点，这种不断扩张的功能会不断强化用户在软件上的使用黏性，从而构成了每套企业软件自身的护城河。而Agent的开发逻辑大大不同，除了Harness工程，另一方面也会强烈需要丰富的数据，包括并不限于企业自身积累的知识数据、实时业务数据，以及第三方外部数据等。
刘仲文表示，过去像ERP系统可以理解为记录系统，在AI原生组织里，AI要成为企业核心生产力，人负责关键判断、对结果负责，就还需要在记录系统之上构建一套行动系统，记录系统把事情记下来，而行动系统则将事情推进下来。记录系统中有金蝶的AI套件，以及星瀚、星空、星辰等云ERP系统，在记录系统之上长出的有灵基Work、灵基Manage，以及灵基Build等AI原生产品。
为此，灵基的设计思路希望通过AI能力发挥出原先ERP系统或SaaS系统无法实现的能力，为企业进一步带来价值。而ERP仍是企业非常重要的数据和知识核心资产。
为调度与原有ERP系统的连接，灵基被嵌入到金蝶客户所有ERP系统首页，企业员工打开ERP系统首页就能看到灵基。在使用场景上，尽管最开始、最容易切入的场景其实是企业数据分析场景，即通过灵基将ERP系统中的数据进行查询、分析和归因，但金蝶认为客户已经走到AI应用场景的深水区。
刘仲文指出，灵基的ERP连接器，最重要的核心功能是开放了数据查询的能力。在他看来，把一个复杂的ERP系统和AI平台结合起来，能够让它完整地去干活，是一件极其复杂的事情。比如金蝶自身在整合灵基和苍穹的探索过程中，就曾意识到，外部第三方平台在没有苍穹的权限体系、管控体系和上下文支撑下很难做到，而查询对客户系统的伤害是最小的。
另外，上下文方面，灵基则是把ERP从记录系统升维成Agent的认知底座，让Agent带着企业既有的业务逻辑和规则去干活，而不是从零开始猜。Agent的能力上限不完全取决于模型本身，而取决于它能拿到多少结构化的企业上下文。ERP系统里沉淀的不只是数据，还有业务对象之间的关系、审批流转的规则、权限边界以及历史处理习惯，这些是Agent查询、表达、行动的基础。
对于任何一款新软件，通常需要12到18个月才能从试用过渡到在关键业务系统中运行。起初，许多Agent产品无法大规模推广到企业场景中去，也在于产品本身仍不够完善。金蝶对灵基的回应并不回避“早期”二字：产品发布仅五个月，技术上成立，但客户价值的可信度、人机交互的边界等，也需要靠工程能力与客户校验共同推进。
章勇对于未来的判断是，当前多项因素叠加，共同构成了金蝶新签约客户的增长动力：一方面，国资委围绕穿透式监管与DRP（全域数字化资源管理平台）建设，对央企提出了一系列明确且相互关联的要求；另一方面，AI作为“十五五”规划的重要发展引擎，也在持续释放政策与市场红利。
（本文首发于钛媒体APP，文 | TechHorizon）
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
According to a 2026 mid-year performance disclosure by Fluttershy, Rin Ki publish has contracted 26 clients and over 40 Agents on the line, and AI primary product income has increased by 189 per cent over the same period. By the time of this update, the number of Rinki paid clients had reached 3706.
During this period, Zhang Yong, Managing Director of the Golden Fluttershy Group, Liu Jianwen, Vice-President of China and Director-General of the Centre for Research and Development, shared around product positioning, industry changes, etc. The central idea is to upgrade the ERP from the records system to the cognitive base of Agent, where Agent works with established business logic, permission boundaries and historical habits.
Butterfly's AI tried earlier than the publish of the Ring. In August 2023, publish GPTLLM, Golden Fluttershy, was based on the idea of embedding LLM ' s capabilities in the ERP and finance scenes, introducing budget planning, code generation, etc. By 2024, the golden aloft had been upgraded to a new-generation enterprise-level AI platform, a phase that validated the viability of AI in the business management landscape, but also revealed a problem: Agent had difficulty working without the established business logic and context.
Today, the consensus in the software industry is changing: past-business software, driven by permission, can only be developed and packaged to meet the utility value of each enterprise ' s role. This expanding functionality will continue to reinforce the user-friendliness of the software and thus constitute the moat of each enterprise ' s software. The development logic of Agent, on the other hand, in addition to the Harness Project, is very different, and there is a strong demand for rich data, including and not limited to the knowledge accumulated by the enterprise itself, real-time business data, and third-party external data.
According to Liu, in the past, an ERP system could be understood as a recording system, in which AI would be the core productivity of the enterprise, with people taking responsibility for critical judgement and for results, there would also be a need to build an operational system over the recording system, which would record things and move things forward. The record system includes the Golden Fluttershy ' s AI suite, as well as cloud ERP systems such as stars, stars, stars, and so forth, which are based on the record system and are based on the elixir Work, the elixir Manage, and the initial products such as the Lingki Build.
To this end, the spirit-based design approach seeks to add further value to the enterprise through AI ' s ability to develop capabilities that could not be achieved by the previous ERP or SaaS systems. ERP, on the other hand, remains an important data and knowledge core asset for enterprises.
In order to move the connection to the original ERP system, the cell is embedded on the front page of the all-ERP system of the Fluttershy client, which is visible when an enterprise employee opens the first page of the ERP system. In terms of the use of the scenes, although the first and most accessible scene is the business data analysis scene, i.e., the search, analysis and attribution of data from the ERP system through the matrix, the Fluttershy believes that the client has reached the depth of the AI application.
According to Liu, the most important core function of the Rinki ERP connection is the ability to open up data queries. In his view, it was extremely complex to combine a complex ERP system with an AI platform that would allow it to work in its entirety. For example, in its own exploration of the matrix and the sky, the butterflies themselves realized that external third-party platforms were difficult to achieve without the support of the cascading permission system, the control system and context, and that the least harm to customer systems was caused by queries.
In addition, for context, the linguist was to move the ERP from the record system to the cognitive base of Agent, where Agent worked with business logic and established rules, rather than from scratch. Agent's cap on capacity does not depend entirely on the model itself, but rather on the extent to which it has access to structured enterprises context. It is not just data that settles in the ERP system, but also the relationship between the business audience, the rules for approving the flow, the permission boundary and historical processing habits, which form the basis of Agent ' s inquiries, representations and actions.
For any new software, it usually takes between 12 and 18 months to move from trial to operation in a critical business system. Initially, many Agent products could not be scaled up to the business scene, and the product itself was still incomplete. The Golden Fluttershy's response to the linguist did not shy away from the word “early”: the product publish was technically valid for only five months, but the credibility of the value of the client, the boundaries of human interaction, etc., also required a joint advance of engineering capability and client validation.
Zhang Yong's assessment of the future is that there are a number of factors that, together, constitute the driving force behind the growth of the newly contracted clients: On the one hand, the Commission has put in place a set of clear and interrelated requirements for Zhang, built around penetrating regulation and the DRP (Global Digital Resource Management Platform), while on the other hand, AI, as an important development engine for the “Fifty-Fifty-Five” programme, has continued to unleash policy and market dividends.
(This post was launched on titanium medium APP, Man, TechHorizon)
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.