On August 6, at the "Small and Micro Enterprise Digital Intelligence Forum" of the 2026 Global Business Innovation Conference, Chanjet released a new AI architecture. The press release was not long, and industry media did not give it much coverage, but I carefully read through the product logic and felt it was more worth pondering than many grand product launches.
The core boils down to two points: Chanjet has defined an "Agentic ERP" architecture for itself and built two AIs — "Xiao Chang" handles conversations with people, while "Chang Lobster" handles tasks on behalf of people.
"Xiaochang" may look familiar to many — it was the one that passed the CAICT's first batch of ERP intelligent agent evaluations. Its positioning is as an intelligent hub: when you ask in the chat box, "What's the payment status for this customer?", it automatically pulls up related orders, payment terms, and payment records, so you don't have to dig through module by module. This layer is still the work of an "assistant."
The real change lies in "Chang Lobster." According to the official description, it is an "AI agent for enterprise self-evolution"—capable of autonomously planning tasks, invoking business data, and executing actions. Note this detail: it directly calls core data such as product archives, customer information, and financial accounts through the MCP protocol. In other words, it does not just answer questions in a chat window, but reaches into the transaction layer of the ERP system.
I. Between "answering questions" and "getting things done" lies an entire execution chain
The difference between the two has been debated in the industry for a long time, but this year, the actions of various manufacturers have drawn the boundary increasingly clearly.
过去的 ERP AI，形态基本是「聊天机器人」：你问它上个月库存周转率多少，它给你算个数；你让它生成个报表摘要，它给你写段话。看起来挺聪明，但所有动作都停在「输出信息」，下一步操作还是要人来点。
The "Chang Lobster" represents a different form: from instruction reception to task execution, it forms a complete business loop. When you say "This batch of goods won't arrive until next month, send the shortage alert to procurement," it can plan the steps on its own—check inventory, verify in-transit items, generate a shortage list, send reminders—and then carry out the actions after your confirmation.
Changjietong's own statement is quite straightforward: "Let AI evolve from 'answering questions' to 'getting things done.'" In plain terms: for the first time, AI can directly execute business actions, rather than just providing business advice.
The significance of this shift, at a smaller scale, is an additional automated entry point; at a larger scale, it means ERP's AI has transformed from an "advisor" to an "executor"—from only giving advice to being able to take action.
二、这不是畅捷通一家在动，是整条赛道在集体换挡
Zooming out a bit, over the past month, almost all mainstream manufacturers have been crowding in the direction of "execution":
| Manufacturer/Product | AI Form | 执行能力 |
|---|---|---|
| Odoo 19.4 | Native MCP Connector + AI Agent | External AI tools can directly connect to the ERP database for reading and writing, and execute voice commands. |
| Oracle Fusion | 22 Agentic Applications | Agent teams autonomously run processes within permission/approval/audit boundaries, with humans only handling exceptions. |
| Epicor Prism | 18+ 供应链预置智能体 | Interpret MRP logs, validate payables, identify shipping performance, and automatically execute pre-approved actions. |
| Chanjet | Xiaochang + Chang Lobster Dual Engine | MCP calls core data, autonomously plans, executes, and forms a closed business loop. |
Four vendors—international, open-source, and domestic—each take different approaches, but their key terms heavily overlap: MCP, autonomous execution, and human confirmation as a safety net.
再看需求侧的数据，就更明白这股换挡不是厂商自嗨。Futurum Group 上周发布的企业软件决策者调研（n=833）显示：43% 的企业期望生成式 AI 主要以「智能体自主执行任务/流程自动化」的形态交付，而期望「副驾式助手」的只有 31%。也就是说，买家脑子里的预期已经变了——不是让 AI 帮我想，是让 AI 帮我做。
所以「畅龙虾」这种产品的出现不是孤例，是需求倒逼出来的必然。谁的动作慢，谁手里的 AI 就停在演示层面。
III. What truly tests execution-oriented AI is not the model, but three checkpoints
话说到这份上，得泼点冷水。「AI 能干活」听着爽，但真让 AI 在 ERP 里动手，考验的根本不是大模型聪明不聪明，而是下面这三件事：
1. Permission boundaries. What data AI can access and what orders it can modify must strictly fall within the existing role-based permission system. Changjie Tong, Oracle, and Epicor all emphasize "operating within permissions" — this is not empty talk, but a matter of life and death for execution-oriented AI. An AI that can autonomously create orders, if the permission model has vulnerabilities, is more dangerous than ten inefficient employees.
2. Approval as a safety net. All execution-type products retain a "manual confirmation" step: AI completes 90% of the work, and the final step is reviewed by a human. This both provides businesses with a checkpoint and gives auditors something to account for.
3. 数据质量。AI 执行的准确率，上限就是主数据质量的下限。商品编码混乱、客户档案重复、科目口径不一，AI 越勤快，错得越快——这一点在后面那篇关于采用率差距的文章里还会展开。
Among these three barriers, permissions and data are the most easily underestimated. Many enterprises, when deploying an execution-type AI, first react by choosing models and configuring parameters, only to discover after going live that the master data is so messy the AI cannot define standard actions, and the permission boundaries are so vague they dare not let it actually act. After all the fuss, they end up falling back to manual work.
IV. A Pragmatic Assessment for Enterprises
At this point, I'd like to offer companies that are selecting or observing a not-too-complicated decision framework:
First, stop evaluating ERP AI by "whether it can chat." In the next six months, the AI gap between vendors will quickly show up in "what actions it can autonomously execute within permission boundaries." When watching product demos, ask one more question: after this AI takes an action, can it be found in the audit log?
Second, pilot execution-oriented AI starting with scenarios where the cost of mistakes is controllable. For example, tasks like order creation, reconciliation, shortage alerts, and duplicate invoice detection—these have clear value, and even if errors occur, they won't harm the core.
Third, open-source ERP will be the most comfortable testing ground for execution-oriented AI. Systems like Odoo, with native MCP support, open permission models, and transparent audit logs, have significantly lower modification costs and better controllability when integrating an execution-oriented agent compared to closed systems. This is also why Odoo 20 treats Agentic AI as a major focus—only with an open architecture can AI safely reach into the transaction layer.
One final reminder: be wary of any execution-type AI solution that promises "full automation, zero human intervention" right off the bat. Over 40% of Agent projects in the industry are halted, and most fail not because of model capabilities, but because of "unsexy" aspects like permissions, auditing, and exception handling. Execution can be handed over to AI, but the steering wheel must remain with humans.
Finally, I want to say this: from "being able to chat" to "being able to work," this small step is the watershed moment for ERP AI moving from demonstration to production. The significance of Changjietong unveiling "Changlongxia" lies not in how mature this product is, but in the fact that it marks domestic small and micro ERP officially entering the arena to bet on the direction of "execution closed loop." In the next six months, it will come down to who can first guard the three gates well and who can first get AI to truly roll up its sleeves and get to work.
