On the evening of July 10, the China Academy of Information and Communications Technology released a list. The list was short, containing only a few names. But if you are involved in enterprise software selection or managing enterprise digitalization, this list is worth a careful look.
This is the result of the first batch of "Internet Intelligent Agent Governance ERP Intelligent Agent Capability Assessment" by the China Academy of Information and Communications Technology. Changjietong (a subsidiary of Yonyou)'s Agentic ERP intelligent hub "Xiao Chang" has passed the assessment, becoming one of the first batch of ERP intelligent agent products in China to receive this certification.
In plain language: The AI capabilities in ERP finally have a set of national-level standards to measure them.
Previously, when vendors claimed "our ERP has AI," you would ask "how can you prove it?" — everyone could only watch each other's demos, case studies, and PPTs. Now it's different. The CAICT has provided six dimensions; if you pass, you pass, and if you don't, you don't. When enterprises are selecting solutions, they no longer have to just listen to vendors' stories.
Six dimensions to measure the true capabilities of an ERP agent
This assessment is not a mere formality. I carefully reviewed the evaluation framework from the Academy of Information and Communications Technology, and the six dimensions basically cover what an ERP agent should do.
Assessment Dimensions | Specific Investigation Content | Align with Real Enterprise Needs |
Interactive Understanding | Natural language intent recognition, multi-turn dialogue, anaphora resolution | Employees don't need to study the operation manual |
Business knowledge support | Understand ERP business rules, compliance constraints, and process boundaries | AI does not act recklessly or operate in violation of regulations |
Business Processing | Conversational document generation, process initiation, information completion | Can use natural language to complete actual business |
Data Query and Analysis | Query data with natural language, automatically generate charts, interpret trends | Managers gain business insights with one sentence |
Safety protection | Data encryption, least privilege, audit logs, malicious instruction interception | AI enters core systems, security is guaranteed |
Skills Management | Skill access review, version management, source traceability, fault isolation | Each AI "skill" is manageable and controllable |
If you are selecting an ERP, directly ask vendors about these six dimensions — what level your AI has achieved in these six areas. The CAICT has already prepared the exam questions.
Among these six dimensions, I paid special attention to "business knowledge support" and "Skills management." The former measures whether the AI understands enterprise rules—for example, orders cannot be below cost price, reimbursements must include compliant receipts, and whether the AI can proactively identify these constraints. The latter measures the vendor's full lifecycle management of AI skills—whether there is review before skills go live, whether issues can be isolated when they arise, and whether version updates require re-validation. These two aspects are precisely where problems are most likely to occur after AI enters core enterprise processes.
Details of Changjietong's "Xiao Chang": Multi-layer Memory and Business Closed Loop
The product certified by Chanjet this time is called "Xiao Chang," serving product lines such as Haoyecai and Haoshengyi. Some of its design concepts, I believe, are worth referencing for all teams working on ERP AI.
One is a multi-layer memory system. The L0 layer is real-time conversational memory, allowing the next round to pick up on what was discussed in the previous one. The L1 layer is a conversational semantic summary, where core information is automatically extracted. The L2 layer is long-term memory for enterprise static knowledge. Additionally, there is tenant-level shared memory across agents—when switching between business scenarios, the AI does not "lose memory."
What inspiration does this approach offer for Odoo to do AI?
In Odoo 20's Agentic AI roadmap, AI needs to handle not just one module, but cross-module business processes involving CRM, accounting, inventory, manufacturing, and more. If AI cannot remember which customer's order the user just checked in CRM, it will lose context when switching to the inventory module. Multi-layer memory is not just an enhancement; it is a prerequisite for AI to truly get work done.
The other is business closed-loop capability. When handling business operations, "Xiao Chang" uses conversational step-by-step information collection to automatically match document types and generate compliant documents. When information is missing, it accurately identifies the issue, actively pushes reminders at key nodes, and automatically sends alerts for overdue items. This is not just a Q&A robot, but a "digital colleague" capable of following up on business processes.
"What enterprises truly need is not more AI tools, but a mechanism that enables AI to operate in coordination with the organization, allowing knowledge, experience, and capabilities to be continuously accumulated and reused."
——Jin Die Vice President Liu Zhongwen, July 2026
What the National Standards Mean for the ERP Industry
The CAICT's move, as I understand it, mainly changes three things.
First, the capabilities of ERP agents are no longer determined solely by vendors. In the past, when everyone developed AI, you claimed to have an assistant, and so did I. But there was no unified standard to measure what your assistant could do or to what extent mine could perform. Now there is one. This doesn't mean all vendors must build AI in the same mold, but rather that enterprises will know where the "passing line" is when selecting solutions.
Second, security and control capabilities have become hard thresholds. Among the six dimensions, security protection and Skills management are directly related to whether AI can "work safely." Data encryption, least privilege, audit logs, and malicious instruction interception—these are baseline requirements, not bonus features. Whether it's a commercial version or an open-source version, if you want to build ERP AI, you must pass this checkpoint.
Third, the evaluation criteria will force an industry reshuffle. The CAICT framework is likely to become a reference for enterprise ERP selection. Government procurement, bidding by central and state-owned enterprises, and selection by large private enterprises will probably directly cite it. Those that fail to pass will not even qualify for shortlisting.
This is both an opportunity and a pressure for Odoo. Odoo 20's Agentic AI is already handling core business tasks such as accounting audits, inventory forecasting, and work hour suggestions. If Odoo can proactively benchmark against such standardized evaluations—especially in the Chinese market and among companies going global—it would serve as a solid proof of competitiveness.
Back to the Enterprise Perspective: Three New Questions in ERP Selection
After the standards from the Academy of Information and Communications Technology were released, from the perspective of enterprise selection, I think at least three things need to be asked.
First: Can the vendor's AI capabilities pass these six assessments? There's no need to actually obtain certification, but can the vendor clearly explain what level they have achieved in each dimension? Those who can explain clearly indicate they have made serious plans. Those who cannot explain are likely still "telling stories."
Second: Can AI understand the business rules of your industry? Every industry has its own regulations. Manufacturing has BOM change approvals and quality traceability; trading companies have credit limit controls and multi-currency reconciliation. Whether AI can comprehend these rules and directly help you prevent non-compliant operations in conversations is far more important than "being able to chat."
Third: When AI goes wrong, is there a "circuit breaker mechanism"? This is exactly what the Skills management dimension examines. If a skill malfunctions, can it only affect itself without bringing down the entire system? Has it been re-verified before a version update? If a problem occurs, can it be traced back to a person? Asking these three questions can help you avoid many pitfalls.
Summary
This certification by the CAICT is not large in scale, but it provides a yardstick for the industry. The AI capabilities of ERP are moving from "storytelling" to "having measurable standards." Changjietong's "Xiaochang" is the first to take the plunge, and it certainly won't be the last.
For enterprises currently selecting an ERP system, this ruler is ready to use. Using it to measure vendors can save a lot of time.
For open-source ERP systems like Odoo, this is also an opportunity to prove themselves. Odoo 20's AI capabilities are already ahead in terms of functionality. If it can pass standardized evaluations—especially in the Chinese enterprise market—its competitiveness in the AI era will be more solid.
The matter of integrating AI into ERP can no longer remain at the level of "Can your AI chat?" The real answer lies in "Can it properly and compliantly get things done?"
