Introduction: AI Finally Stops Merely "Giving Advice" from the Sidelines
Over the past two years, discussions about AI in enterprises have mostly remained at the level of 'ask it a question and it answers with a paragraph.' When it actually comes to having it enter a purchase order into the ERP for you, match customer names across two systems, or pull back customs tax refund details and check them, AI often gets stuck—because it cannot reach the systems, or the systems have not opened an interface for it.
This bottleneck was broken through simultaneously this month by two companies using two different approaches. One company’s idea is “wrapping,” while the other’s is “native.” Viewed together, they instead make a more pragmatic main path for AI implementation clearer: first get AI into the actual business front line, then discuss whether it is native or not.
1. Insights from Soiawork: AI Can Operate ERP Even Without an API
On November 2, Shizai Intelligence released Soiawork (Shuya Work), an enterprise-grade agent platform positioned as a "dedicated AIOS." What ERP practitioners should most consider is its set of "dual-connection" capabilities.
For systems with open interfaces such as APIs and MCP, Soiawork uses connectors to directly query, create, and update within the authorized scope; for existing systems that have no interfaces and where employees can only open the interface and operate manually, it relies on Workers deployed on execution devices, combined with RPA and Computer Use (computer vision operation) capabilities, to achieve "operate as soon as the screen is recognized"—without needing to modify the original system or requiring the other party to open interfaces, AI can "understand" the screen, retrieve data across systems, and automatically fill out forms. The company claims this capability achieved a 90.2% task success rate on the OSWorld benchmark.
Even more critical is its governance design: permissions, credentials, logs, and human confirmation mechanisms, which constrain the boundaries of 'what AI can do,' making important operations traceable and results verifiable. The recruitment scenario is a good example—after a specialist receives a JD, Soiawork assists with creating talent profiles, screening resumes, and writing greeting scripts, but the specialist can stop at any time and retains final review authority.
The logic behind this is plain: companies don't buy ERP to raise yet another new system, but to let AI take over the manual work of "copying across systems, reconciling, and passing along" within their existing systems. What Soiawork targets is precisely the "last mile" of AI entering the business front line.
II. Sellfox's approach: turning ERP from a “store management tool” into a “digital workforce”
Another path comes from cross-border e-commerce. Ebrun reports that Sellfox ERP upgraded its AI capability system at the 2026 summit, launching a matrix of autonomous agents covering all cross-border operational scenarios and emphasizing a zero-threshold end-to-end intelligent workflow.
It differs from Soiawork’s “parcel” approach and is more like growing AI inside ERP: first build data middleware, unifying more than 800 metric definitions and over 10,000 knowledge nodes, then self-develop a cross-border inference engine and ReAct loop. It enters high-frequency scenarios from compliance, advertising, and creative: compliance connects directly to tax authorities and customs, automatically aggregates export tax rebate details, directly reports 0110 and 9810, and 9810 verification can also automatically aggregate multi-source data and trace in reverse. The report mentions that active sellers grew by 226%.
The difference between these two companies precisely outlines two approaches for AI entering ERP: one is bolt-on, not touching existing systems; the other is embedded, treating AI as a native capability. But for most growing enterprises that already have an ERP in place, the threshold for the first is much lower.
III. What It Means for Enterprises That Are Already Using Odoo
If we shift our perspective back to Odoo, we will find that open-source ERP is actually a natural breeding ground for the "connection camp."
Odoo 19/20 Enterprise Edition comes with an OpenAPI 3.0-compliant REST API, OAuth 2.0, and webhook subscriptions, while the Community Edition also has a mature ORM and a large number of third-party modules. This means: you don’t need to wait for the vendor to feed AI into every menu—you can use connectors or MCP yourself to connect agents to high-frequency areas such as sales, purchasing, and inventory. Combined with ready-made RPA and document automation modules in the community, SMEs can completely adopt a low-cost approach of “one high-frequency scenario + one set of connectors,” first putting digital workers on duty in a certain area, and then expanding after validating effectiveness.
There is an often overlooked sequence: first have clean master data and standardized processes, then let AI take over, and only then will the results be stable. In areas where the data all speak for themselves and the processes themselves have not yet taken shape, bringing AI in will only amplify the chaos, not eliminate it.
Conclusion: Connection Over Rebuilding
Taken together, the two developments this week are actually offering a new answer to an old question: should companies tear everything down and start over for AI? For the vast majority of companies, the answer is no. Connecting existing ERP systems smoothly with connectors, keeping permissions under control, and handing the most painful high-frequency step to AI is far more cost-effective than rebuilding an AI-native system.
The main path for AI deployment is shifting from “building intelligent systems first” to “letting agents enter existing systems first.” Whoever makes this path work first will see their digitalization truly move from demonstrations to the production line.
