From November 2 to 5, the 2026 World Internet Conference Wuzhen Summit was held in Wuzhen, Zhejiang, with the theme "Open Source and Open, Co-construction and Sharing". At booth A4-02, Hall A of the "Light of the World" Expo Center, Shizai Intelligence appeared with its enterprise-grade agent platform Soiawork, raising a proposition that is practical for every enterprise: when AI enters business systems, what is missing is not the model, but the "last mile"—how to enable agents to understand screens, operate software, retrieve data and fill forms across systems, and still keep boundaries under control, without modifying legacy systems or requiring open interfaces.
This precisely hits the real pain point of enterprise digitalization today: ERP, MES, OA, finance, and supply chain each manage their own segment, and much of the work still relies on people copying, checking, and passing data between different interfaces. Soiawork sums up the answer in one phrase—"enterprise AIOS" (enterprise-grade AI operating system).
Soiawork's Worker achieved a 90.2% task success rate on the OSWorld benchmark, and can adapt to interface-operation tasks without modifying the original system (Shizai Intelligence, 2026-11-02)
I. Why "Enterprise AIOS": The Last Mile for AI to Enter the Business Frontline
Sun Linjun, founder and CEO of Shizai Intelligence, put it bluntly: "The value of AI ultimately comes down to whether it can get things done and whether it can be accountable for the results." What Soiawork targets is exactly the "last mile" of AI entering business operations. Users state goals in natural language; the platform calls on agents, workflows, knowledge bases, and connectors, covering web backends, local files, desktop software, and business systems, completes tasks, and consolidates the process into assets that teams can reuse.
Its underlying confidence comes from three things: the self-developed TARS large model, ISSUT screen semantic understanding technology, and RPA execution capabilities. With "screen recognition is operation," the platform can, without modifying existing systems or requiring open APIs, enable AI to "understand" the screen, operate software, and retrieve and fill data across systems. For medium and large enterprises with severe "legacy burdens" in IT systems, this is a low-intrusion, fast-validation path.
"Bringing AI into the business is not about removing people from the process, but about reducing repetitive cross-checking and giving time back to judgment." — This is precisely the most pragmatic stance for putting enterprise AI into practice.
2. Three-tier architecture + full-stack connectors: turning “experience” into “reusable assets”
Soiawork has designed a three-tier structure for experience reuse: Knowledge Base enables AI to find enterprise materials, Skill solidifies validated methods into reusable skills, and Workflow organizes the steps and conditions of multi-stage tasks into repeatable processes, driving tasks that an individual has successfully run through to become capabilities that the team can reuse. At the management level, a shared workspace integrates task context, files, knowledge base, skills, workflows, and connectors, and configures permissions according to members' responsibilities.
Connectivity is key to this architecture. For systems with open interfaces such as APIs and MCP, connectors invoke query, create, and update operations within the authorized scope; for systems that have no available interfaces and can only be operated through the UI, Workers deployed on execution devices, combined with RPA and Computer Use Agent, complete tasks within the adaptation scope. Currently, connectors already cover commonly used tools such as MES, ERP, CRM, SRM, office collaboration, and enterprise information query—this is precisely the entry point through which AI truly 'catches' enterprise core systems.
| Business Scenarios | What does Soiawork do | Manual confirmation point |
|---|---|---|
| Manufacturing industry material kitting / work reporting verification | Based on work orders, BOM, inventory, and in-transit purchases, list work-start readiness, shortage items and replenishment completion time, and flag under-reporting and duplicate work reporting. | Write into the authorization system after personnel confirmation |
| E-commerce Product Listing / After-Sales Verification | Document check, review status readback, and consolidation and verification of after-sales orders/amounts/logistics/vouchers | Manual review of abnormal items |
| Cross-border e-commerce settlement reconciliation | Reconcile settlements, orders, and provisional estimates by store/order number/currency/settlement period, and list missing orders, refunds, and differences | Financial Review |
| Pharmaceutical Retail Prescription Orders | Verify authorized prescription orders against inventory, query the review status, and only return results that have passed review. | Prescribing/reviewing/dispensing are still handled by qualified personnel |
III. Governance Is Boundary: The Four Essentials of Permissions, Approvals, Audits, and Human Confirmation
Agents that can "operate software" are most afraid of "operating recklessly." Soiawork's governance logic is worth CIOs learning from: through permissions, approvals, auditing, and usage and cost management, it keeps AI use "within boundaries and with recorded inputs"; at the same time, it provides connectors, scheduled triggers, human takeover, usage management, and audit traceability. All the above scenarios retain a human confirmation step—AI handles information gathering, comparison, and anomaly identification, while business judgment is still made by humans.
Knows how to work
Connect ERP/MES/CRM, retrieve data across systems to fill forms, automatically check for anomalies, and free people from repetitive work.
Can be controlled
Tiered permissions, approval gates, full-process auditing, and visibility into usage costs; important operations are traceable and results verifiable.
This is structurally aligned with the “enterprise context foundation” approach we previously tracked: Kingdee Lingji uses the “Enterprise Context Platform” to uniformly connect ERP + knowledge base + memory + ecosystem MCP, enabling agents to “understand the enterprise”; Soiawork, by contrast, uses connectors + RPA + screen semantic understanding to enable agents to “enter the field.” The two paths converge despite differing routes—the decisive factor for enterprise AI is shifting from “how powerful the model is” to “whether it can be securely, controllably, and auditably integrated into real business.” This also echoes the three hard constraints of controllability, full-process traceability, and digital passport in the mandatory national standard for agent security implemented in October (plan no. 20263116-Q-252).
Three pitfalls that enterprises must avoid when implementing AIOS:
① Focus on models, neglect connections: You buy a powerful model but cannot connect it to ERP/MES; AI can only do Q&A and cannot access business data, so its value drops to zero.
② Execution without approval: Smart agents can write to systems without human confirmation and auditing; once a misoperation occurs, there is no traceability, and compliance risks surge.
③ Individuals tinker alone, teams scatter: Employees each get their tasks running independently but do not codify them into Skill/Workflow, so experience cannot be reused and scale cannot grow.
IV. Practical Recommendations for CIOs / Digital Leaders
1. Start with "a high-frequency repetitive task". There is no need to refactor the system right from the start. Select a scenario involving cross-system reconciliation and high manual time cost (such as cross-border settlement reconciliation or material kitting checks), use "connectors + manual confirmation" for low-risk validation, and expand after achieving ROI.
2. Put the governance framework upfront in the selection criteria.When evaluating any enterprise AIOS / agent platform, the four items—permission tiering, approval gates, audit traceability, and usage cost—must be hard requirements, not optional.
3. Prioritize a “light-intrusion” approach that does not modify the original system. For enterprises with heavy legacy burdens, the combination of RPA + screen semantic understanding + connectors is more stable, faster, and more economical than tearing everything down and rebuilding.
4. Consolidate "individual run-through" into "organizational assets". Mandate that experience be solidified into reusable Skills and Workflows, incorporated into a shared workspace for unified management, and prevent the fragmentation of AI capabilities.
It is important to be clear: Soiawork represents the "connection + governance" route. It solves the last mile of AI entering business operations, but it cannot replace the data sovereignty and process core of ERP/MES itself. Enterprises still need to keep their core ledgers, master data, and compliance logic in their own hands—AIOS is an amplifier, not a foundation.
Soiawork at the Wuzhen Summit served as a reminder to all companies struggling with "how AI can truly enter the business": the enterprise AI race is not about whose model has more parameters, but about who can safely connect agents into ERP, MES, and CRM, control the boundaries, and turn experience into organizational assets. Only when the last mile is connected does AI move from "demo" to "productivity".
Data sources for this article: Tencent News, “Shizai Intelligence Brings Enterprise-Grade Agent Soiawork to the World Internet Conference” (2026-11-02); China Daily Website, “Zhejiang Anji: Advancing ‘AI + Manufacturing’ from ‘Single-Point Scenarios’ to a ‘Platform Foundation’” (2026-11-02); Official website of the 2026 World Internet Conference Wuzhen Summit (theme: “Open Source and Openness, Co-building and Sharing”, 2026-11-02 to 11-05). Data as of 2026-11-05.
Zhongdi Information · Specializing in Odoo and digital solutions for enterprise management
