First, let's talk about a charging pile.
Wuxi Bomeixin makes new energy charging piles and is a wholly owned subsidiary of the Dutch BUCOMEC Group. Its products are mainly exported to Europe and sold to more than 50 countries. It sounds like a standard business, but the real trouble is hidden in the details: products of the same power rating spawn several configuration versions because national grid standards differ; the core electronic control is developed in-house, sheet metal structural parts and power modules are outsourced for processing, and imported components and domestic parts are purchased together; every charging pile that leaves the factory must be linked to that specific BOM using a unique SN code.
That's complex enough. Now here comes the question: for a company like this to adopt AI, what should the first step be?
Most people's first reaction is to choose a model, select scenarios, and find vendors. But when I looked at the kickoff meeting materials Yonyou released on November 9, I instead felt the answer lies elsewhere—they set the goals of the Phase 1 project as "orders visible, costs calculable, quality traceable," first centralizing master data management and distributing it on demand, unifying the definitions of materials, BOM, customers, suppliers, and accounts, and then building order-based operations monitoring and MRP linkage.
In other words: first spend half a year solidifying the foundation, then talk about intelligence. This is not conservatism; this is the only viable path.
1. Context is not data volume; it's the ins and outs
Why is the foundation so important? Just look at another message from the same day and you'll understand.
Hitachi began offering TWX-21 Procurement AI on November 9, with the first batch focused on procurement operations. Its confidence does not come from the model, but from the network that TWX-21, an inter-enterprise trading platform, has built up over more than 20 years—according to public figures, about 87,000 companies are connected to it. With this layer of relationships, the Agent knows why this supplier was late last time, in which year that part number last went up in price, and whose quotes have always been on the high side.
This kind of information has a characteristic: it is not in the database; it is in the folds of the process. Emails, faxes, WeChat messages, the purchaser's notes, a return from three years ago. If you throw these things away, even giving the Agent ten large models would be useless, because all it has in hand is a clean quotation sheet and cannot see any anomaly.
Rillet's story is another kind of proof. This agentic ERP, which just raised a $100 million Series C three months ago at a $1 billion valuation, announced on November 9 that it is entering the UK, and currently serves nearly 1,000 customers across 75 countries. Among its customers, Black Forest Labs CFO Dave Arkell said something that I think is more accurate than any product introduction: the tech stack needs AI at its core, "not just features like MCP connections".
Coming from a client-side CFO, those words carry different weight. He doesn't want a chat box that can query data; he wants the same data to be operable both by people and by Agents, with consistent definitions, consistent permissions, and consistent auditing.
Looking at these two things together, so-called "context" is not about stuffing all of a company's data into the model. It is a business record with causes and consequences: why this order was changed, why this supplier was kept, why this material uses a substitute part. These answers are written in the process, not in the documents.
II. Multiple organizations and multiple standards—the most expensive lesson
Returning to the Bomeixin case, it makes the pain points of manufacturing companies going overseas clearer than a proposal PPT:
| Business Characteristics | A question posed to the system | must first be resolved before AI |
|---|---|---|
| Three-country legal entities (Wuxi / Bukomeike / Hong Kong) | Cross-border multi-legal-entity independent accounting + centralized group financial control | Unified account subjects, currency types, and tariff standards |
| ETO/MTO hybrid, customized per project | Frequent design changes mean delivery dates change accordingly. | Order changes can be evaluated and rescheduled |
| In-house development + outsourcing + mixed sourcing | Planning and execution are disconnected; material shortages and excess inventory coexist. | MRP can accurately calculate the linkage among procurement, outsourcing, production, and inventory. |
| Multi-country configuration versions of the same power rating | High pressure in version and engineering change control | BOM version and effective date clearly stated |
| Each pile's unique SN is associated with the BOM | Forward and reverse quality traceability | SN, batch, and storage location run through from incoming material to shipment |
| Quotations rely on experience | Cost quotations are rough, and gross profit cannot be calculated accurately | The four items of material costs, manufacturing costs, labor costs, and selling expenses can be broken down. |
In this table, not a single row is an "AI requirement." They are all ERP work. But they determine whether AI can be connected later—if the SN and BOM versions don't match, you can't even answer "is there a problem with this batch of goods," let alone talk about intelligent quality inspection and intelligent traceability.
There is another detail worth copying: Bomeixin has set up a master data gatekeeper in the project organization. This role is not a software function; it is a person, with a name, who is responsible for the definitions for materials, BOMs, customers, and suppliers. The project advances step by step by module; general issues go to the project manager, while major matters are submitted to the steering group for decision.
In projects, I have increasingly come to believe one thing: in eight out of ten failed master data governance projects, the problem is not a lack of tools but a lack of someone who 'has the final say and is willing to take responsibility.' Before implementing another MDM system, it is better to first designate this person.
R&D, marketing, production, parts sales, and Hong Kong finance are brought onto the same foundation: master data is centrally managed and distributed on demand, ensuring business-finance consistency from the source
III. What are the realistic options for doing AI on legacy systems?
Not every company can rebuild a Rillet. Most clients have an ERP that has been running for five to ten years—it can’t be changed and can’t be replaced. At this point, how should the phrase “AI grows in the architecture” be put into practice?
In my experience, there are three criteria, and they have nothing to do with which model you choose:
- Whether the data is in a single database. The "real-time" pieced together across three databases and two intermediate tables is actually "synchronized once per hour." The inventory an Agent sees when judging delivery dates may already be two hours old.
- Are the rules readable? If business rules exist only in a consultant's head, or are sealed in a black box whose source code cannot be seen, you cannot explain to auditors "why this order went through special approval."
- Can permissions be refined down to the record level?It is not "this user can view the sales module," but "this user can only view customers in their own region." With coarse-grained permissions, as soon as an Agent is connected, it becomes an entry point for data leakage.
Take Odoo as an example. Its natural advantages in these three points are: a single PostgreSQL database, business logic readable and modifiable in Python source code, and permissions assigned at model level and record level (ir.model.access + ir.rule). SN and batch tracking, MRP, and subcontracting—these manufacturing scenarios are native modules and do not require third-party add-ons.
But it's also important to be clear about its shortcomings: the community edition has no built-in agent capabilities—you have to build them yourself. This means that tasks like context injection, tool wrapping, and audit logging have to be implemented one by one by the implementer, rather than just installing an application and being done with it.
I've fallen into this trap before: to meet the schedule, I wrote the Agent's query action to run with super-admin permissions. The system ran smoothly, until one day the business team asked "Why can East China sales see South China prices?" and we realized the permission table had long existed in name only. The account used by the Agent must have the same permissions as the real person it serves, not one bit more.
There is another thing that is often underestimated: the volume of historical data migrated directly determines how far back the Agent can look. I have seen a project where, when the new system went live, only opening balances and open documents were migrated, while shipment records from three years ago were left in the old system. As a result, when the procurement Agent assessed supplier delivery reliability, it had only three months of samples and failed when seasonal fluctuations occurred. Later, two years of historical documents were backfilled, and the workload was exactly double that of the originally planned "data cleansing".
So during the planning stage, you have to ask: how long a historical window does this scenario need? Procurement judging delivery date deviations usually requires 12 to 24 months; quality traceability needs to look at the complete BOM version change chain; cost analysis must cover at least one complete quotation cycle. Only when the window is clearly calculated can the migration scope and schedule have a basis; otherwise, it will definitely only be discovered during UAT.
IV. Three Points for the Implementation Lead
- First draw a "context map," then define the AI scenario. Take an A3 sheet of paper and draw out all the documents, master data, historical records, and approval chains involved in this scenario. The parts you cannot draw are exactly where the Agent will definitely answer incorrectly. This diagram is more useful than a requirements list.
- Treat unifying definitions as phase-one work, not a prerequisite. Many projects put master data cleansing under "post-launch optimization," and as a result they are still catching up the following year. Bomeixin's approach is to include the master data distribution mechanism in the phase-one scope and assign a gatekeeper; this sequencing is worth learning from.
- Choose a path that lets you read the source code. It doesn't have to be Odoo, but it must be able to answer "where is this rule written." In the AI era, the most valuable capability is not getting the system running, but being able to explain why when problems occur.
Delight.ai's approach to customer agents also corroborates this. It puts shopping, booking, support, and retention into the same agent, sharing customer identity, preferences, and history, and covering chat, voice, SMS, email, and WhatsApp. The selling point is not "many channels," but that pre-purchase recommendations and post-purchase service are connected to the same context. The same context means customers don't have to be asked repeatedly, and there won't be self-contradiction.
So the conclusion of this article is actually quite plain: enterprises do not need to first buy stronger intelligence; they need to first articulate their own processes clearly. How SNs are assigned, how BOM versions take effect, how outsourced work is inspected upon return, who approves changes, and according to what basis costs are allocated—the answers to these questions are the true boundary of an Agent's capabilities.
Models will change generation after generation, but processes and standards won't grow on their own.Whoever finishes laying the foundation first will have less trouble when the next round of model upgrades comes.
Sources: Yonyou Network's 'Wuxi Bomeixin Chooses Yonyou BIP Super Edition: Making Multi-Organization Operations “Orders Visible, Costs Calculable, Quality Traceable”' (2026-11-09), Hitachi TWX-21 Procurement AI launch (11-09), Rillet UK expansion announcement and customer quotes (GlobeNewswire, 11-09), and Delight.ai Customer Agent launch (11-09). The customer size, customer count, and deployment scope mentioned in the article are all based on the vendor's public disclosures.
Zhongdi Information (Zodioo) · Enterprise Digitalization Observation · 2026-11-10
