In June 2025, I wrote about JD’s AI supply-chain case. At the time, I saw it as a corporate transformation experience worth learning from, although its scale felt distant from that of many Taiwanese companies.

Looking back now, I care more about how to turn someone else’s case into questions for my own work. Methods that work for a large company still need to be reexamined under different data, processes, and customer conditions.

Start with the Working Paper, and Ask What It Actually Establishes

The Stanford GSB archive page lists AI-driven Supply Chain Transformation, jointly authored by Hau L. Lee, Zuo-Jun Max Shen, Yongzhi Qi, and Zhe Chen. It is dated May 7, 2025, and classified as a working paper. SSRN’s author information also lists two of the authors’ affiliations with JD.

The abstract uses the Triple-A framework to discuss three JD cases: interpretable time-series forecasting, logistics optimization for inventory allocation and fulfillment, and a customer-to-manufacturer system that uses natural language processing to connect consumer demand with product development.

These sources provide a contextual starting point for the case. Experience documented jointly by academics and company researchers is worth considering. Archiving, authorship, or posting on SSRN cannot by themselves be treated as independent validation of the company’s results.

When I read material like this, I separate the methods described by the authors from the question of whether those methods would work at another company. The second question still has to be tested against that company’s own conditions.

Turn the Case into Three Questions

Starting from these cases, I bring three questions back to my own workplace.

Forecasting: What are we trying to predict? Is it demand for one item, replenishment volume, or delivery time? When I see an improvement in accuracy, I first ask which metric was used, what baseline it was compared with, what period it covered, and whether the data resembles the work in front of us. A single percentage cannot answer those questions.

Explanation: What does the frontline need to understand? Users need more than a recommendation. They also need to know its basis, its limits, and when it should be questioned. An explanation can support communication, but when the data is wrong, the objective is unsuitable, or the process cannot accommodate the recommendation, clarity does not replace fixing the problem.

Planning: How does an option enter execution? Comparing alternative configurations when a warehouse is unavailable, for example, requires clear statements of capacity, demand, cost, and constraints. Comparing options and directly changing live operations are different capabilities; the latter also requires authorization, state checks, stopping controls, and validation of results.

These three questions draw my attention not only to the model but also to the work arranged around it.

From Internal Capability to External Service

Organizational changes in JD-related businesses can easily be folded into the same success story. An announcement published by JD through the Hong Kong Stock Exchange confirms that JD Industrials was listed and began trading on December 11, 2025, rather than merely receiving listing approval.

But a listing is an organizational and capital-markets event. The listing itself does not establish how much of the result came from AI, nor does it establish that a particular AI supply-chain service has generated net profit.

The more useful question for me is this: if a company wants to turn capabilities accumulated internally into an external service, what is still missing? Can another company’s data connect to it? Are its rules and permissions different? Who maintains the interface, who responds to exceptions, and how will service outcomes be checked?

Packaging a process as a service is different from making it reliable for different customers. Productization and continuous maintenance still stand between the two. A platform can be one direction, but company scale or a listing event cannot answer those questions by itself.

Which Part Should a Taiwanese Company Examine First?

A Taiwanese company does not need to begin by copying JD’s scale. It can first choose a process with a clear problem, accessible data, and results that can be checked, then examine where a tool can help.

If I were considering an external service, I would ask: which capabilities are suitable for a supplier, and which knowledge should remain internal? How will the data be used, how will fees be calculated, and what happens when the service is interrupted? Connecting to a platform does not remove the company’s own responsibility for validation.

If the company keeps its existing tools, the question is whether it can maintain clear data definitions, versions, and handoffs. Managing inventory in Excel does not automatically mean a company is behind. If the data is difficult to verify and exceptions are not recorded, moving to a more complex system may not solve the problem either.

I prefer to examine competitiveness through concrete questions: can promised delivery dates be confirmed, can the cause of a stockout be traced, and can the effect of a change be compared? These questions are closer to customer needs than simply attaching the label “AI supply chain.”

What I Value More Now Is Trust That Can Be Validated

I still care whether an organization is willing to use AI recommendations. But trust does not mean asking frontline staff to believe first, nor does it mean renaming a technical problem as a cultural problem.

Data quality, system integration, cost, security, permissions, and responsibility all affect whether a recommendation can be adopted. Helping people understand is one part. Giving them ways to question, stop, and confirm results is equally important.

I therefore begin by asking: which decision are we preparing to let the system assist? Who may adopt its recommendation? What standard determines whether it is right or wrong? Can a small, observable process produce enough evidence to support the next step?

The value of the JD case, for me, is that it opens these questions. Only by bringing the questions back to our own operations and revising from results do we have a chance to build supply-chain capabilities that fit us.

Four-panel comic: Looking back now, I care more about how to turn someone else's case into questions for my own work. Archiving, authorship, or posting on SSRN cannot by themselves be treated as independent validation of the company's results. Comparing options and directly changing live operations are different capabilities; the latter also requires authorization, state checks, stopping controls, and validation of results. The listing itself does not establish how much of the result came from AI, nor does it establish that a particular AI supply-chain service has generated net profit. Data quality, system integration, cost, security, permissions, and responsibility all affect whether a recommendation can be adopted. Only by bringing the questions back to our own operations and revising from results do we have a chance to build supply-chain capabilities that fit us.

Revised November 2026: This article discusses how to read the case using the archived abstract of a 2025 working paper and the listing announcement. The three practical questions and the checks proposed for Taiwanese companies are my own framework.