From November 4 to 5, Science and Technology Daily, China Daily Online, and others successively reported on "IDC MarketScape: China Intelligent Agent Platform Vendor Assessment, 2026" (No. CHC54110526; figure caption: October 2026; author: Sun Zhenya, Research Manager, IDC China). For the first time, the report placed 28 mainstream domestic vendors on the same capability–strategy coordinate chart.
What is more worth examining than "who is in the Leaders Quadrant" is the procurement motivation data: enterprise demands are shifting from "whether it can be developed" to "whether it can be managed well, calculated clearly, and used reliably", and the focus is moving from "whether a Demo can be produced quickly" to agents in production environments. The procurement question has changed.
IDC survey: surveyed enterprises ranked "strengthening agent governance and compliance" as the primary goal for adopting the platform (highest-ranked multiple-choice item, 2026-10 report / cited in November)
1. Three figures changed the procurement question: first buy “can manage,” then buy “can do.”
IDC completed its assessment along the two dimensions of capability and strategy, combining vendor interviews, user research, and expert evaluations. Among the three top priorities, none is "a stronger model" or "a faster demo":
| Primary goal | Proportion | Corresponding pain points to be addressed |
|---|---|---|
| Strengthen agent governance and compliance | 58.7% | Blurred permission boundaries and untraceable execution chains |
| Improve development and delivery efficiency | 52.5% | Duplicate construction across multiple departments, long delivery cycles |
| Support large-scale replication across scenarios | 49.2% | Single-point tools are difficult to reuse, and Token costs are out of control |
The three items are multiple-select; it is normal for the total to exceed 100%. IDC judges: what enterprises need is no longer a single tool, but a foundation that can uniformly host, connect, run, and govern various agents, turning scattered Agents into governable, reusable, and continuously iterable digital assets; platforms, together with MaaS and data knowledge platforms, collectively form the enterprise IT infrastructure of the AI era.
The market-side figures were repeatedly cited in the same week: the scale increased from 21.2 billion yuan in 2025 to 44.9 billion yuan in 2026 (roughly +111.8% year-on-year based on this set of data); by IDC's measure, it is expected to exceed 332 billion yuan in 2029 (China Economic Net, 2026-11-04, third-party citation; boundaries are subject to the original report).
A remark by Sun Zhenya can serve as the acceptance criterion for this round of platform procurement: "The larger the scale of enterprise agent applications, the more the platform needs to uniformly manage agents scattered across different departments and business systems. The platform must not only support development and deployment, but also ensure reliable business execution, control risks and costs, and continuously iterate and optimize based on operational feedback."
According to a publicly cited MarketScape diagram, cloud vendors, ERP vendors, and the open-source project Dify are placed in the same chart (quadrant positions are based on the original report). Open source and commercial offerings appearing together shows that the "platform layer" has become an independent procurement category.
II. "Large-scale popularization" does not equal "entering core processes": four sets of gaps need to be examined separately
The report states that Chinese enterprises' agent construction has "moved beyond proof of concept and entered a cycle of large-scale popularization." But another set of publicly cited data from the same week warns: there is still a wide gap between popularization and deep use (People's Political Consultative Conference Network, 2026-11-05; the original text notes "publicly cited basis" and "statistical boundaries differ among sources"):
73% → 19%
IDC figures: 73% of global Top500 enterprises have launched agent pilots, while only 19% have achieved cross-departmental scale.
18%
Cited in CAICT's "Enterprise-Level Agent Technology and Application Research Report (2026)": only 18% of enterprises have integrated agents into their core business processes
| Data | Statistical object | Nature |
|---|---|---|
| 73% pilot / 19% scaling | Global Top500 Enterprises | Institutional Research (Secondary Citation) |
| 18% into main business processes | Chinese enterprises, main business processes | Quoted from CAICT report |
| >40% of projects will be suspended or canceled (by the end of 2027) | Agentic AI Project | Gartner forecast value |
| 95.4% of enterprises have deployed an average of 11 production workflows | IDC Resilience and Spending Survey, as of 2026-07 | Institutional research (independent basis) |
These four groups of figures use different definitions—some count enterprises, some count projects, and some count workflows; the geographic scope and the definition of "scale" all differ, so they cannot be added together to produce a penetration rate or conversion rate. The only judgment they can jointly support is this: deployment is fast, while entering core business processes and remaining governable is much slower. Writing the definitions into the selection minutes is more useful than citing impressive percentages.
The government and enterprise side put it more bluntly. Xu Liang, vice president of Sugon Cloud, used "a world of ice and fire" to describe customer polarization: some are fully AI-enabled, while more medium and large organizations and government departments are concerned about five factors—security, data, environment, control, and cost; Wu Xingjie, director of solutions, attributed the obstacles to mechanisms—"The biggest obstacle is that enterprises have not yet established a trust mechanism for incorporating agents into real business."
III. What Does Platform-Based Governance Buy: Four-Layer Architecture, Open-Source Model Management, and "Data Does Not Leave the Domain"
Break down the platform sample released publicly in the same week; 'manage well, calculate clearly, and use steadily' correspond to four layers of specific engineering capabilities.
| Capability Layer | Public samples and practices | What to solve |
|---|---|---|
| Model Layer (Clear Calculation) | Shenzhou Wenxue MaaS provides unified management of open-source and commercial models, and the model marketplace offers evaluation reports, open-source model deployment, and multi-API Key load balancing. | Models are interchangeable, costs are comparable |
| Semantic layer (understands business) | The knowledge governance layer performs parsing and graph modeling; Authine Yunshu pre-structures business objects, relationships, and rules, and calibrates whether the "contract counterparty" and the "supplier" are the same entity. | Prevent the Agent from guessing business meaning after receiving raw fields. |
| Process and Permission Layer (Managed Well) | Workflow orchestration; the AI resource operations layer provides unified access to models / Embedding / OCR, with authorization based on organizational scenarios and measurable, traceable consumption; Marketingforce GenAIOS + AI-Agentforce + KnowForce | Three Ledgers: Permissions, Links, and Token Costs |
| Deployment and Trust Layer (Stable in Use) | Sugon Cloud OneBox deploys models, knowledge bases, and applications in government and enterprise own environments, with data not leaving the domain; permissions are set by role, and sensitive operations are logged; "Secure Session" filters external network models, and "Tree Bridge" sets permissions by table structure/fields | Data Sovereignty and Compliance Red Lines |
Authine Yunshu's cases are concentrated in scenarios such as the full contract lifecycle and review of construction technical plans—long processes with intertwined permissions (Accelink Technologies, Gezhouba, China Economic Net, 2026-11-04). The hardest part for an agent is not answering, but knowing, within an existing approval flow, whom to contact next, which segment to return to when an error occurs, and under whose name usage is recorded.
The timing and nature of OneBox must be clearly stated: It was first launched at the Wuzhen "Internet Light" Expo on 2026-11-03 and selected as a "New Light" product; "the first government-enterprise trusted desktop-level AI agent" is a vendor's self-described positioning. Product manager Ma Yilong gave very specific red lines: according to policy, core data from public security, discipline inspection, the military industry, and large state-owned enterprises and central enterprises must not be connected to the Internet. The Expo's public figures are 54 countries and regions, 671 institutions, and more than 1,000 new products.
Don't equate "private deployment" with "trustworthy." Self-hosted deployment only addresses data residency; you still need to ask: does the tool inherit existing identities and data permissions, can actions be replayed, and can anomalies be stopped with one click? This is the same set of requirements as the three hard constraints of the mandatory national standard for AI agent security in China in October: "controllability / full-process traceability / digital passport."
IV. Practical Advice for CIOs / Digital Leaders: Replace the Vendor Selection Scorecard with Six Questions
1. Whether permissions are inherited rather than a separate set being created.Use existing identity, organization, and data permissions for authorization, and demonstrate item by item "which table a certain position can view and whether it can write"; newly added business units should inherit permissions and billing rules.
2. Can costs be calculated down to the action level.Models/Embedding/OCR are integrated uniformly and allocated by scenario, providing visibility into each agent's tasks, consumption, and ROI calculations, with upper limits and alerts. "Token cost out of control" is a widespread pain point specifically cited by IDC and should be included in the acceptance checklist.
3. Whether the chain can be replayed and exceptions can be rolled back.Decision logs and manual approval points are linked by business documents; payments, postings, bulk master data changes, and external sending retain manual confirmation and one-click deactivation.
4. Has business semantics been structured? Live demo: whether "supplier" is the same entity in contracts, procurement, and finance, and which action is triggered after approval. If you can't answer, no matter how strong the model is, it is still prone to fail in long processes.
5. Can models and deployments be switched? Verify open-source weight self-hosting, multi-model comparison, external model filtering, and log export on exit. Governing open-source models is not just about saving money—it is about an executable right to exit.
6. How long does it take to replicate a scenario, and how many people are needed? Replicate a scenario that has already been proven in a second business unit, and record configuration hours, permission changes, and regression testing volume. This predicts three-year operational costs better than a case collection.
Sequence matters more than the checklist: first establish the three ledgers of permissions, costs, and audit trails, then discuss the number of use cases. Platform-based governance is not a bonus; it is remedial work—what is being compared is no longer who can build a Demo, but who can make agents work in the system over the long term and in a verifiable way.
Limitations and decision conditions: 58.7%／52.5%／49.2% come from IDC multi-select surveys; RMB 21.2 billion／44.9 billion／332 billion and 73%／19%／18% are public secondary citations, with different boundaries and regions; "large-scale adoption" is an IDC judgment, and the >40% project cancellation is a Gartner forecast; OneBox's "first" and the capabilities of various platforms are mostly vendor claims and lack independent testing. The recommendations apply to organizations that are "preparing to introduce or have already introduced a platform and need unified governance"; if still conducting single-point validation, they should first address business semantics and data quality. This article provides decision support and does not constitute a selection conclusion.
What is worth remembering this week is not who entered which quadrant, but that the subject of the procurement question has changed: from "what can this AI do" to "can this system be governed, accounted for, and used reliably." Only requirements that can be accepted and held accountable can actually land.
Sources: “IDC MarketScape: China Agent Platform Vendor Assessment, 2026” (IDC #CHC54110526, 2026-10, Sun Zhenya; Science and Technology Daily / China Daily website, 2026-11-05; China Economic Net, 2026-11-04; cited from a graphic on the Digital China WeChat official account); People's Political Consultative Conference Network (2026-11-05, including IDC 73%/19%, CAICT 18%, Gartner >40%, and the note that “each party's statistical boundaries differ”); Huanqiu.com / CCIDnet report on OneBox's Wuzhen debut (2026-11-03/04); China Economic Net, “2026 China Enterprise Agent Vendor Implementation Standards” (2026-11-04). Vendor self-statements, third-party citations, institutional research, and forecast values are listed separately, without merging definitions. As of 2026-11-08.
Zhongdi Information · Specializing in Odoo and digital solutions for enterprise management
