NextAI+ Praxis--:----UTCIf the previous issue’s Home Depot case represented how enterprises with relatively strong digital foundations can advance AI deployment by building on existing strengths, Chow Tai Fook is closer to the real starting point of most traditional enterprises. It relies heavily on store networks, sales-associate service, jewelry craftsmanship, product assortment, and customer trust, while its online and offline links are not naturally integrated. When such enterprises discuss AI, they often treat a complete data platform, mature online channels, stable inventory systems, and cross-store coordination as prerequisites for starting transformation. Chow Tai Fook’s experience shows that traditional jewelry retail can also begin with infrastructure, online–offline connectivity, digital customization, and employee workflows, gradually completing its foundation before extending AI into customer experience, internal operations, and retail planning.
This article breaks down public announcements and customer cases from enterprises, cloud vendors, and software providers, and selects, from a third-party perspective, the parts that can be explained through business logic. The core questions we care about are these: for traditional retail enterprises represented by Chow Tai Fook, what capabilities were treated as problems that had to be solved first at different stages; which business departments are involved in a traditional retail enterprise; where the data structure becomes complex; and where AI ultimately lands. These lessons apply to any retail sub-sector. What we are inevitably dissecting are “old cases”: technology patterns update quickly, and what is truly transferable is not a particular technology stack, but experience. For that reason, we want to use the Chow Tai Fook case to dig deeper both horizontally and vertically.
Dissecting a large enterprise’s AI path over the past several years has two clear purposes for consulting.
First, it is to read the enterprise’s starting point and sequence of entry. When a large enterprise advances intelligent upgrading, the places it moves first are often the areas it considers most foundational, easiest to show results in, or most in need of early investment. Arranging these choices chronologically makes it possible to abstract a directional path and a set of lessons that other enterprises can reference. But one point must be emphasized: a path is not a universal sequence. In real projects, what comes first and what comes later often varies by enterprise, depending on its own data maturity, departmental demands, and management expectations. What we want to extract is a set of referenceable actions and the judgment logic behind them; the concrete sequence of implementation must still return to the actual conditions of each enterprise.
Second, it is to read the evolution of suppliers. The changes in Chow Tai Fook’s partners over several years are themselves a signal. It first chose a traditional SaaS giant such as SAP in the early stage (2023), then introduced cloud vendors and office productivity platforms, and later turned to specialized AI vendors for retail planning. This evolutionary line shows that, in certain links, traditional service providers are no longer sufficient on their own to support capability-building for the new AI era. When a new participant enters, it often means that the enterprise has reached a stage where a need has emerged that the old supplier cannot cover. We therefore study each supplier: what kind of company it is, what its core business is, and which comparable players exist. We do not judge “why Chow Tai Fook chose A rather than B.” But when a giant enterprise chooses A, we can infer that enterprises in similar positions may be choosing among B and C. As consultants, what we need to do is lay out these alternatives and provide more flexible solution options.
Below, we look at Chow Tai Fook along these two lines: what is being dissected on the business side, and what is changing on the supplier side.
Chow Tai Fook’s digital foundation can be broken down into two stacked layers.
The first layer is the data management foundation, undertaken by SAP. In FY2023, Chow Tai Fook listed “accelerating digitalization” among the Group’s five strategic directions. In October of the same year, SAP announced its cooperation with Chow Tai Fook in a customer case titled “Chow Tai Fook Jewellery Goes Live with SAP SuccessFactors to Accelerate HR Digital Transformation.” Chow Tai Fook deployed the Employee Central and Recruiting modules of SAP SuccessFactors to unify HR processes and employee records across Hong Kong, Macau, Japan, South Korea, and Southeast Asia. Later, according to SAP’s customer case “Chow Tai Fook Jewellery: Driving Global Expansion for a Century-Old Industry Leader with Data-Driven Talent Transformation,” Chow Tai Fook further evaluated SAP S/4HANA Cloud Private Edition to centralize finance systems scattered across different regions, supporting financial management across multiple companies, currencies, and accounting standards.
What this layer supplements is the unification of talent, organization, finance, and cross-regional management standards. For a retail enterprise with a large store and franchise network and continued overseas expansion, this foundation determines whether subsequent AI can obtain stable and trustworthy data for employee efficiency, business analysis, budget forecasting, performance, and supply-chain decisions.
From the supplier perspective, SAP represents traditional enterprise software: SuccessFactors belongs to the human capital management (HCM) track, while S/4HANA belongs to core ERP. In other words, what Chow Tai Fook bought at this layer was a mature “system of record.” There are many comparable options: in HCM, Workday and Oracle Fusion HCM; in ERP, Oracle, Microsoft Dynamics 365, Infor, and others. For enterprises in similar situations, the real question is not “whether to use SAP,” but how to choose among these mature SaaS systems according to their regional structure, compliance requirements, and pace of expansion. It is worth noting that this layer solves the “unification and trustworthiness” of data, but it does not directly generate AI capabilities by itself. It is the foundation for everything that follows.
The second layer is the process operating foundation, undertaken by Alibaba Cloud and the D-ONE digital jewelry customization platform. According to Alibaba Cloud’s public customer case, Chow Tai Fook adopted Alibaba Cloud infrastructure to support online back-end operations, business process management, online and offline transactions, and major promotional scenarios. It also connected the D-ONE digital jewelry customization platform with AI and automation engines, enabling customized jewelry to be prepared for delivery within 24 hours after an order is received. Chow Tai Fook’s connection with the Alibaba ecosystem in fact predates this round of HR and finance system transformation. As a jewelry brand that entered online retail relatively early—through channels such as its Tmall flagship store and new retail initiatives—its cooperation with Alibaba Cloud in e-commerce back-end systems and cloud infrastructure began earlier. The integration of D-ONE with AI and automation engines further digitalized the customized delivery process on this basis. The exact starting time of the cooperation should be based on Alibaba Cloud’s customer case and the two parties’ announcements.
What this layer supplements is the compression of a traditional jewelry customization process—previously dependent on store communication, manual coordination, craftsmanship handoff, and production scheduling—into a digital process that can run, be monitored, and be delivered: customer needs enter the system, orders and customization actions become process-based, and production response and delivery cycles are managed digitally.
From the supplier perspective, Alibaba Cloud is a public cloud vendor, providing something closer to a “carrying layer for process operations,” often delivered through elastic methods such as serverless architecture. Comparable options include AWS, Microsoft Azure, Google Cloud, and domestic Chinese cloud vendors such as Tencent Cloud and Huawei Cloud. This is where a selection issue that traditional enterprises often underestimate emerges: choosing a foundation supplier is, in essence, a balancing problem. A public-cloud serverless model offers strong elasticity and a fast starting point, but industries with higher confidentiality requirements may need private deployment or dedicated servers in exchange for stronger control. Enterprises need to retain their own degree of control over this layer and make trade-offs around three things: the supplier’s capability boundaries, the corresponding price range, and the current state of business development. Placing these items side by side for comparison is precisely one of the values that consultants can provide to enterprises—not deciding on behalf of the enterprise which vendor to use, but clarifying the options, costs, and conditions for fit.
Taken together, the two layers show that SAP gradually unifies enterprise management data, while Alibaba Cloud and D-ONE make key business processes recordable by systems. The former answers the question of whether the enterprise can possess unified, trustworthy, and callable data; the latter answers whether business actions can be recorded by systems and generate verifiable process outcomes. For traditional enterprises whose digital foundations are not yet mature, these two layers jointly form the precondition for subsequent AI. Only after the foundation has been completed can AI extend into digital customization, employee collaboration, and retail planning.
After the foundation was completed, Chow Tai Fook placed AI into three types of business scenarios. It should again be emphasized that writing them together is meant to dissect “where AI can be embedded,” not to prescribe that they “must be done in this order.” In different enterprises, the sequence of these three types of scenarios may well be reversed. Priority depends on which type of scenario has the highest value, the most complete data, and the lowest resistance.
First, customer-side digital customization. This was one of the earlier scenarios in which Chow Tai Fook placed AI into a real delivery chain. The D-ONE platform connected AI and automation engines to Alibaba Cloud, enabling customized jewelry to be prepared for delivery within 24 hours after an order is received. Jewelry customization is naturally suited as an entry point: the average order value is high, customer needs are relatively clear, the process from design to ordering, production, and delivery is clear, and outcomes can be directly verified through delivery cycle and customer experience. Its closest business logic is to compress a service chain that depends on manual communication and store coordination into a shorter and more monitorable digital process, while continuously accumulating customer preferences, customization needs, and production response data to provide real data sources for later recommendation and planning optimization.
If an enterprise wants to build such a customization process itself, it can be broken down technically into a sequence of several actions: understanding a customer’s personalized needs, retrieving craft and product knowledge, generating a solution, confirming the order, triggering production, and following up on delivery. In the past, these actions often had to be stitched together with a series of independent tools—workflow orchestration, knowledge retrieval, model access, permission control, and call observability each covering one segment. This is also why the tool lists in cases such as Chow Tai Fook appear especially long. At the end of this article, we will consolidate these referenceable open-source components into a list for teams that hope to build on their own. But a judgment should be made first: more “AI-native” integration methods have already emerged today, bringing steps that previously required multiple toolchains into fewer links. Having more tools is not the objective; clarifying the business process and then supporting it in as unified a way as possible is the real point. We will return to this later.
Second, employee-side workflows. A case disclosed by Microsoft in 2026, titled “Hyper-Intelligence: Chow Tai Fook and Microsoft Join Hands to Redefine the Future of Global Luxury Retail,” shows that Chow Tai Fook deployed more than 400 customized AI agents based on Microsoft 365 E5, supporting more than 24,000 employees, and stated that efficiency in core business processes improved by more than 70%.
There is only one truly informative signal: Chow Tai Fook’s AI has moved from front-end customer experience into employees’ daily collaboration and internal business processing, with usage frequency approaching the level of enterprise productivity tools.
Why is the employee side suitable for expansion after organizational and data foundations are in place? Because employee workflows cover store operations, customer service, product management, training, marketing activities, back-office approvals, and management reporting. Information is highly dispersed and strongly dependent on identity, permissions, documents, and organizational data. After Chow Tai Fook first used SAP to unify organizational and talent data, employee agents could more easily access real work contexts and take on high-frequency tasks such as knowledge queries, content generation, process follow-up, material summarization, and assisted decision-making. Their value can be translated into language familiar to traditional retail: shorter process handling time, less repetitive work, faster cross-department response, and more efficient training and knowledge acquisition.
From the supplier perspective, what Microsoft provides at this layer is a combination of productivity platform and agents. Its direct counterpart is Google Workspace plus Gemini. Which office suite an enterprise chooses usually follows its existing installed base. For us as AI consultants, the key is to understand the interfaces of these office software platforms, embrace the third-party ecosystem as much as possible, and complete data integration, unification, and cleansing on top of it, so that AI capabilities can be connected smoothly.
Third, retail planning and supply-chain decision-making. In 2026, o9 announced that Chow Tai Fook had selected o9 Digital Brain, in a customer case titled “o9 Selected by Chow Tai Fook Jewellery Group to Optimize Data-Driven Retail Planning,” to unify category planning, merchandise financial planning, production planning, allocation, and replenishment. The announcement mentioned goals including reducing stockouts of high-demand products, strengthening allocation mechanisms, providing a more suitable product assortment for each store, and accelerating replenishment cycles.
This scenario enters the operating core of traditional jewelry retail and is also the area most worth digging into vertically. The complexity of the jewelry industry is concentrated in the matching of products, inventory, stores, and regional demand: the number of stock keeping units (SKUs) is large, the value of each item is high, store distribution is broad, regional preferences differ significantly, and festivals, weddings, and fluctuations in gold prices continuously disturb demand rhythms. If digital customization solves the efficiency of a single delivery chain, and employee agents solve the efficiency of organizational collaboration, then retail planning handles the efficiency of resource allocation among products, inventory, production, finance, and stores—that is, which stores should carry which products, which products should be produced first, which stores should be replenished first, how inventory should be reduced, and how regional demand should be forecast.
These problems are not new. Retail operations and management scholars have studied them for many years, and this research provides a deeper foundation for our analysis. A classic judgment in the field of supply chain and retail operations can be traced to Marshall Fisher’s 1997 article in Harvard Business Review, “What Is the Right Supply Chain for Your Product?” It distinguishes between “functional” and “innovative” products, arguing that the harder demand is to predict and the shorter a product’s life cycle is, the more it needs a supply chain oriented toward rapid response. Seasonal styles, collaborations, and highly fashion-oriented jewelry products fall precisely at the “innovative / hard-to-predict” end, which helps explain why Chow Tai Fook needed to introduce stronger forecasting and rapid replenishment capabilities on the planning side. Fisher and Ananth Raman’s 2010 book The New Science of Retailing, published by Harvard Business Review Press, further applied this line of thinking to data-driven inventory and merchandise decisions. On the specific question of “which items should a store carry,” the review “Assortment Planning: Review of Literature and Industry Practice,” coauthored by A. Gürhan Kök, Marshall Fisher, and Ramnath Vaidyanathan, systematically organizes the methodology of assortment planning. As for the demand structure of luxury goods and jewelry, the annual Luxury Goods Worldwide Market Study jointly published by Bain & Company and Altagamma, as well as The State of Fashion, published annually by McKinsey and Business of Fashion, are commonly used public sources for observing regional preferences and channel changes. When these studies are overlaid on Chow Tai Fook’s planning scenario, the picture becomes clearer: what platforms such as o9 truly solve is not “installing another piece of software,” but using a unified data model to answer retail allocation problems that have been studied for more than twenty years.
From the supplier perspective, o9 is a specialized AI platform for supply-chain and retail planning. It uses a graph-based data model—what it calls an Enterprise Knowledge Graph—to connect product attributes, customer signals, financial targets, and operational constraints, linking category decisions with financial plans and inventory movement. Comparable players include Blue Yonder, Kinaxis, RELEX Solutions, Anaplan, and SAP’s own IBP. What is intriguing is precisely the supplier evolution itself: Chow Tai Fook used SAP at the data foundation layer, but turned to a specialized vendor such as o9 in the more “AI-intensive” retail planning link, rather than continuing with SAP’s planning module. This supplier switch shows that, for planning capabilities in the new AI era, traditional ERP giants may not be the default option. For enterprises in similar situations, this is exactly a decision point where alternatives such as Blue Yonder, Kinaxis, RELEX, and o9 need to be laid out and compared.
Taken together, the three scenarios reveal a map of “where traditional retail places AI”: customized delivery with high average order value and clear processes; employee collaboration with high frequency and strong permission requirements; and product and supply-chain planning closest to the operating core. Their dependence on data and process foundations deepens in sequence, but this does not mean every enterprise should follow the same order. The value of Chow Tai Fook lies in its complete demonstration of which business departments are involved, which data matters, and where AI lands. The sequencing must always return to the specific enterprise’s maturity and demands.
Viewed along this path, the first budgets that become real in traditional enterprise AI transformation do not land on a particular front-end AI application, but on the digital foundation. Many enterprises’ AI efforts remain stuck at the proof-of-concept stage for a core reason that often has little to do with the absence of models or tools: underlying operating data has not yet entered a unified system, and business processes are still dispersed across different teams and stores. Without connectable data and processes that can absorb AI, AI cannot enter the real business chain, nor can it form measurable outcomes.
The first type of budget lands on the data management foundation. It corresponds to systems such as SAP SuccessFactors and S/4HANA, which gradually unify organization, human resources, finance, regional operations, and compliance standards. In the past, such spending was often treated as an HR or finance system upgrade. Placed in the AI context, it plays a more fundamental role: providing a unified and trustworthy data source for subsequent AI in employee efficiency, budget forecasting, performance analysis, and supply-chain collaboration. What the enterprise buys is not just a system, but a management foundation that allows operating data to be continuously accumulated, called, and analyzed.
The second type of budget lands on the process operating foundation. It corresponds to investments in cloud infrastructure, e-commerce and business process platforms, and customization platforms, allowing key business actions to be digitally recorded and process-based: demand enters the system, orders become traceable, production becomes monitorable, and delivery becomes measurable. Only after processes are first absorbed by systems does AI have the opportunity to participate in demand understanding, recommendation generation, process advancement, and delivery optimization.
These two types of budget are easier to approve inside traditional enterprises because they can be explained in familiar business language: the data foundation corresponds to organizational efficiency, financial compliance, cross-regional visibility, and decision-response speed; the process foundation corresponds to order-processing efficiency, customized delivery cycles, store collaboration, and customer experience. Compared with “spending money on AI,” “first strengthening the data and process foundation in order to improve transparency, reduce friction, and shorten delivery cycles” is clearly easier to accept. What traditional enterprises truly need to budget for first is often precisely those underlying capabilities that do not look like AI, but determine whether AI can land.
The bottleneck most worth watching on this path is the rhythm mismatch between foundation-building and AI scenario implementation. Foundation-building naturally takes a long cycle and involves historical data cleansing, standard unification, process reconstruction, permission boundaries, user habits, and cross-department collaboration. Business departments, meanwhile, hope to see results quickly: whether delivery cycles have shortened, whether efficiency has improved, and whether replenishment judgments have become more accurate. One side requires long-term foundation-building, while the other demands short-term value. The two rhythms can easily conflict: if the foundation moves too slowly, AI remains in trial use; if scenarios move too quickly, they are held back by fragmented data and broken processes.
This kind of friction exists almost equally in every project, and we face it ourselves as well. Here is our approach, stated directly:
Determine priority indicators. We combine the nature of the enterprise, the demands of each department, and its past level of digitalization and AI to define several indicators for judging project priority—for example, urgency and data completeness. These indicators vary by enterprise and must be tailored to the specific situation; there is no universal weighting system.
Score and rank demands. After the indicators are determined, we use them to score and rank all current demands, and move the higher-scoring ones forward.
Validate through a small-scope demo, then iterate quickly. We first build a small and interactive demo within the enterprise and test it together. We run value through in a feasible small scenario first, then quickly replicate it to more departments and scenarios.
The second layer of friction comes from organizational absorption after a system goes live, and this is especially important. In any enterprise, even after a full AI system has been built, it must still be ensured that subsequent business actions continue to take place within that system so that data can keep accumulating. Otherwise, the foundation will gradually idle. Behind this are issues of employee training and organizational structure adjustment. As a third party, we participate throughout the project, accumulate an understanding of the enterprise’s real situation, and provide supporting services: necessary technical training, organization-level training, and workshops to improve personnel AI capabilities.
There is another layer, returning to the repeatedly mentioned issue that “tools are too scattered.” When a case like Chow Tai Fook is broken down to the technical layer, it produces a fairly fragmented picture: data collection, mapping, organization, and storage are distributed across a series of tools, each managing one segment. This picture itself illustrates the difficulty of the work. We are working with technical teams to converge the chain from “data collection to mapping, organization, and storage” into a more unified solution. This part will be discussed separately later. For now, we leave one judgment here: open-source components can provide orchestration, retrieval, permissions, observability, modeling, forecasting, and optimization capabilities, making them useful references for teams that build in-house. But what determines implementation quality is never the number of tools; it is whether the enterprise can arrange the correct sequence according to its own maturity and build the data foundation in as unified a way as possible.
It needs to be emphasized that traditional enterprises neither need to wait until digitalization is fully mature before starting AI, nor can they skip data and process construction and simply buy tools. What is truly transferable from the old Chow Tai Fook case is not which suppliers it used, but a method of judgment: identifying the foundation that most needs to be supplemented now, the scenario most suitable for initial validation, and the implementation sequence that can be best explained by business results. When data standards are not unified, start by strengthening data management; when processes remain dispersed, first build process absorption; when employee collaboration costs are high, introduce knowledge assistants and workflow agents; when product and supply-chain complexity rises, continue building data-driven planning systems. Only when these capabilities are placed into a route that can absorb, verify, and scale them can AI begin from the foundational links that traditional enterprises are able to catch, and gradually settle into long-term digital capability.
Below, by capability link, we consolidate the open-source components that may be used in the “self-built route” discussed above into a list for teams that want to build on their own. It is only a set of candidates, not a recommended configuration. The same capability usually has multiple replaceable options, and the specific choice must still return to the enterprise’s data maturity and team situation. We ourselves are more inclined to converge these links in a more unified and more AI-native way; this part will be discussed separately later.
The public research and cases cited or referenced in this analysis are listed below for further reading. Publication years and links should be based on official sources:
In the next article, we may follow the clues left at the end of this article and develop two issues: first, clarifying the unified solution from “data collection to storage” in combination with the actual conditions of the enterprises we serve; and second, selecting one or two studies that are closer to retail strategy, such as Harvard Business Review discussions on omnichannel and category management, to see whether in-depth research can add sharper judgments to traditional enterprises’ AI planning. Feel free to send a private message or leave a comment to tell us which enterprise AI practice topics you are curious about. You are also welcome to leave a message on the NextAIPraxis.com website or contact us at [email protected] with real problems your enterprise needs to solve. We will contact you within two business days and invite you to join our practice-sharing community.
Cite as · Enterprise AI Deployment Signals · 3 June 2026
If you want both columns delivered together, four times a year, in one quiet email — leave an address. Otherwise just bookmark this page.