At a roundtable at the China International Import Expo, Kang Kai, Vice President of Dun & Bradstreet China, shared a rather pointed observation: in Yiwu International Trade City, almost every storefront sign reads "Factory Store." Then he asked: What if it isn’t a factory store? If you say it’s a factory store, does that make it one?
This statement was originally about cross-border trade, but what occurred to me after hearing it was something else. Many companies are now preparing to let AI determine 'whether this supplier can be added to the list,' 'how much credit can be granted to this customer,' and 'whether this order carries related-party transaction risk.' But if the system cannot even accurately identify 'which company is actually being analyzed,' then no matter how fast it runs afterward or how beautifully the report is written, it is merely packaging the error more exquisitely.
The several materials released on November 8 and 9 happened to surround this issue from three directions.
1. Identify the wrong entity, and everything that follows is wasted effort.
On November 8, Dun & Bradstreet launched its "Intelligent Open Platform" for the first time in Asia at the 9th China International Import Expo. Founded in 1841, this data service provider does something that, put plainly, is not complicated: it packages trusted business data into services that AI agents can directly invoke, for use in customer development, supplier management, credit assessment, and compliance due diligence.
Its first layer is not the model; it is identity. The platform uses the D-U-N-S Number as the unified identity identifier for enterprises, first matching the same enterprise scattered across different systems—different name spellings, different languages, changed addresses, multiple layers of group hierarchy—these are all longstanding problems in ERP. After matching, it then relies on a global business graph covering more than 200 countries and regions and over 900 million enterprises to supplement, within the authorized scope, group, equity, control, transaction, credit, and risk information.
The second layer consists of two deliverables. MCP services give agents a standardized entry point to call enterprise records, risk information, operational updates, and enterprise scores on demand; business Skills, meanwhile, package data, rules, and analytical logic into reusable task modules, such as “conduct a supplier due diligence” and “run a round of credit review.” This is not the same as the idea recently discussed of “connecting an MCP to ERP and then being able to query data”—the former turns professional judgment into callable capabilities, while the latter merely opens up the query entry point.
| segment | Past | Now (vendor-reported) |
|---|---|---|
| Customer onboarding / due diligence / risk screening | Two or three people, for half a day or even longer | Results in about 10 minutes |
| Process Record | Interpersonal collaboration may not be remembered completely. | Which API was called, what data was viewed, and how conclusions were drawn—all fully documented. |
| Decision Ownership | Human check, human judgment, human signing | AI collects facts, humans make decisions; it does not replace credit policy or manual approval |
That supplementary remark by Product Director Feng Zhikai was crucial: the platform does not replace the company's credit policy and manual approval. He also made an analogy: the model is like a smart, quick-reacting brain, but to make correct business judgments, it still has to correctly understand the real world; "A very smart child, if all he reads are wrong books, may also make wrong judgments."
People who do ERP implementation are no strangers to this scenario: in the supplier master data, three records are actually the same company, only differing slightly in name, tax ID, and address. In the past, this was called "master data not maintained cleanly," and at most affected reconciliation; now that agents are going to use it to determine whether credit can be granted, this bit of dirty data directly becomes decision risk.
There is another detail worth noting: in China, the platform has already partnered with WorkBuddy, KIMI, and others, while some overseas services connect to OpenAI and Microsoft Copilot; domestic connectors are first opened for registration with free credits and pay-as-you-go billing, whereas overseas ones usually require purchasing a data license first before activation. The same data capabilities, when entering different workflows, also require separate discussions on billing and authorization methods—for clients undertaking multi-system integration, this is another account that must be calculated clearly in advance.
II. Keep data in place and let capabilities move
If Dun & Bradstreet addresses “where external facts come from,” then the Oracle and Snowflake partnership reported on November 9 deals with the other end: how to reduce the hassle with all that ERP data inside the enterprise.
At Oracle AI World, the two parties announced that Snowflake AI Data Cloud will be introduced into OCI as a native service. According to the joint statement, enterprises whose data resides in Fusion Cloud Applications, NetSuite, Autonomous AI Database, MySQL HeatWave, and NoSQL can directly access Snowflake's analytics and AI capabilities within OCI, without having to migrate workloads to another environment.
Route A
External trusted data brought in: unified entity identity + business graph + callable business Skills
Route B
Internal ERP data stays within the domain: bring analytics and AI capabilities next to the data, eliminating cross-cloud pipelines
Ashish Chaturvedi of HFS Research laid out the sequence of benefits concretely: large Oracle Applications customers already using Snowflake benefit first, especially those who have to migrate ERP and financial data across clouds every night. Keeping financial and supply chain data within OCI brings the most direct benefit: eliminating outbound data transfer fees. He also singled out NetSuite customers—those who hold valuable Oracle data but lack engineering capabilities; native connectivity exactly saves them the “integration work that would otherwise be difficult to staff.”
Stephanie Walter of HyperFrame Research added from the development team's perspective: spend less time building pipelines to move data over, and more time using that data to build applications. For CIOs, a single contract and a single bill also conveniently lower the procurement threshold.
The two paths look like they point in opposite directions, yet they point to the same thing: the ceiling on AI's usability is determined by data reachability. One turns external facts into callable services; the other drives the cost of moving internal data down to a minimum. In neither architecture is the model the protagonist — position and definition are.
The analyst also left a note: integration still faces challenges in data governance and operational complexity. Eliminating migration does not mean eliminating the need to align definitions. Even if cross-cloud pipelines are dismantled, tasks such as field meanings, period locking, and permission boundaries remain just as numerous—they simply shift from 'the ETL engineer's job' to 'something the business side must articulate clearly itself.'
III. Irreversible Matters: Don't Hand Them to It Yet
Once the data is correctly identified and within reach, only then does it come to the question that is most easily skipped: which judgments can be left to the AI itself?
On November 8, EY released the "White Paper on Financial AI Applications and Trusted Governance" at the China International Import Expo, offering a quite useful yardstick. It breaks down the value realization of financial AI into a four-level path—rule automation, intelligent assistance, agent collaboration, and decision support—and then offers two judgments: the degree of AI autonomy should match business risks and the irreversibility of operations; process restructuring, rather than stacking tools, is the key to level transition.
| Level | What is AI doing | Can it be automatically allowed through |
|---|---|---|
| Rule Automation | Trigger, move, and validate according to predefined conditions | Sure, just hardcode the conditions. |
| Intelligent Assistance | Summarize, compare, and provide recommendations | Yes, the conclusion is not saved to the database |
| Agent Collaboration | Organize materials and draft documents across steps | Add a manual gate and keep an evidence chain. |
| Decision support | Affects credit granting, pricing, and payment | Humans sign; AI does not touch the final step. |
Two other pieces of news from the same day can be used to test this ruler. DingTalk QwenWork announced on November 8 that six leading companies had been added in a week: Shagang applies Agents to post-operation analysis for electric furnace smelting positions, transformer condition assessment, and financial and operational analysis, with efficiency in some core scenarios improved by up to hundreds of times; Supcon Technology, after deploying across the entire marketing chain, is extending into procurement, supply chain, finance, HR, and R&D; Jack Technology focuses on cost accounting, supply chain scheduling, and international order processing. These terms share one thing in common—analysis, assessment, processing—all are still on the “giving suggestions” side; no one says payment approval has been handed over.
More directly, the Wanlian Yida case reported by China Economic Times: a non-standard procurement of 100 tons of lithium carbonate worth tens of millions of yuan. Within minutes, the agents of both the supply and demand sides clarified the grade and delivery date, verified supply capacity, and negotiated quotations and payment terms in parallel; then agents for procurement, quality, finance, legal affairs, and others consolidated the plan and pending matters. Throughout the process, the business owner intervened only at the three stages of authorization, decision-making, and acceptance. The platform has connected more than 5,000 enterprise users and runs more than 26,000 agents.
Break "authorization—execution—decision—acceptance" apart: the human's position is designed, not what's left over. This is far more useful than vaguely saying "human in the loop".
EY global survey: 68% of CFOs believe there is an urgent need to reassess how corporate value is measured; 80% expect AI-driven business models to have a significant or moderate presence in their organizations within the next 12 months
80% says “it will come,” not “it has already been made to work end to end.” The white paper spells it out: finance AI has entered the critical stage of “whether it can have a substantive impact.” From recording, control, and reporting to forecasting, allocation, and co-creation, what stands in between is not a generational gap in models, but whether processes need to be reworked.
IV. Three Practical Recommendations for CIOs
- First align the entities; don't rush to launch agents. Map external enterprise identities against the partner, supplier, and customer tables in ERP, and spot-check 50 records that are duplicates or have similar names. If the mapping rate has not reached a defensible level, do not roll out scenarios such as credit granting, due diligence, and related-party identification yet.
- Make “reviewable” the default configuration. For every judgment involving AI, record in the database the APIs called, the data read, the logic organized, and who ultimately approved it, and retain replay capability. Conclusions without an evidence chain must not be connected to the approval workflow. This is far more important than which model is chosen.
- Grade autonomy by irreversibility.Queries, summaries, and comparisons can be opened up; changes to master data, price changes, loan disbursement, order issuance, and letter issuance must all retain a human gate. Translate EY's four-tier path into your own permission matrix and write it into the rules—don't just leave it in the proposal PPT.
This week's material circles back, and it really boils down to one sentence: models will be replaced generation after generation, but master data and the evidence chain will not grow by themselves. First let AI recognize correctly; only then can we talk about letting AI make decisions. If the order is reversed, the bit of time saved will ultimately be paid back in review, rework, and explanation costs.
Sources: Shangguan News / Qianlong.com / Sina Finance / Tencent News (Dun & Bradstreet Intelligent Open Platform, 2026-11-08), Hithink RoyalFlush Finance (Oracle × Snowflake, 11-09), EY's 'Financial AI Applications and Trusted Governance White Paper' (11-08), DingTalk QwenWork official announcement (11-08), and China Economic Times (Wanlian Yida, 11-08). The efficiency and accuracy figures in this article are based on the vendors' or institutions' own claims and have not been independently reviewed by third parties.
Zodioo · Enterprise Digitalization Insights · 2026-11-09
