AI-Native Reinvention of a Forty-Year-Old Wholesaler: From Traditional Wholesale to Integrated Wholesale and Retail

The client is one of the leading food importers and wholesalers in Hong Kong, with more than forty years in business. It sources and imports food products and supplies a large base of business customers through a central warehouse. Before the engagement, the business had never used AI. It had only an accounting package and no enterprise resource planning (ERP) system, and its operations relied on manual work.

The client’s problems fell into four areas.

(a) Data. Products, inventory, orders and receivables were recorded in separate spreadsheets and files that were not connected. The same product carried different names, specifications and units of measure across departments and documents, leaving the business without unified, reliable product data.

(b) Supply chain. Costs across purchasing, import, warehousing and delivery were not broken down, so the business could not analyze its cost structure, and pricing and purchasing decisions lacked a sound basis. Shelf life was tracked manually, making it difficult to act on near-expiry stock in time. Ordering relied on individual experience, and there was no record of suppliers’ lead times, quality or prices. Food safety traceability depended on paper documents and could not be completed promptly.

(c) Operations. Wholesale orders were received by phone and instant messaging and entered by hand. The price catalog for wholesale customers was compiled manually, which was labor-intensive and slow to update. Credit, reconciliation and collections depended on the experience of individual employees. Staff worked up to 15 hours a day, and their well-being directly affected service quality. Growth in business volume required a matching increase in staff, while the industry faced labor shortages and an aging workforce: according to projections by the Labour and Welfare Bureau, the trading sector will face a manpower shortfall of 11,000 to 16,000 people by 2028, and about 15% of Hong Kong’s employed population was aged 60 or above in 2023.

(d) Market. Under pressure from e-commerce and other emerging channels, the business’s share of the wholesale market declined year after year. With wholesale as its only channel, and without retail outlets reaching consumers directly or a brand of its own, it had no new channel to offset the loss of wholesale business.

The engagement proceeded in four stages.

(a) Initial Diagnostic. We assessed the business’s current operations and determined the role AI could play and where to begin. The diagnostic concluded that the root of the problem was not a lack of tools but the absence of unified data and systems, with operations relying on manual work; adding tools to existing processes would not change this. The decline in wholesale market share also showed that wholesale alone could not sustain growth. On this basis, the business decided to proceed to a full diagnostic.

(b) Full Diagnostic & Proposal. We carried out a full diagnostic based on Ground-Up Reinvention and redesigned its business operating system on AI-native lines. AI-native means deciding at the design stage whether each task is performed by the system or by people, rather than looking for places to use AI after existing processes are in place. The full diagnostic established three directions for reinvention:

(i) Supply chain optimization. With the central warehouse as the hub, the entire process of purchasing, import, warehousing and supply was rebuilt. Import costs are allocated to individual products on defined principles, giving a clear cost structure. Inventory shelf life is managed by batch, with stock shipped on a first-expired, first-out (FEFO) basis. Ordering is based on demand forecasts. Suppliers’ lead times, quality and price history are recorded. Import documents and health certificates are linked to batches to support food safety traceability.

(ii) Digitalized wholesale. Wholesale customers submit orders by instant messaging; the system reads each order and generates a draft, which becomes a confirmed order once the customer approves it. Shipping, reconciliation and collections are carried out by the system according to rules, with disputes referred to staff.

(iii) Extension into retail. The business established its own brand and entered chain retail, operating company-owned and franchised stores alongside an online store, with a membership program connecting stores and online. All downstream channels are supplied by the central warehouse on an outright-purchase basis, and wholesale and retail share the same data and systems. Relationships between channels are governed by rules: franchise supply prices are published, and company-owned stores settle at the same prices; the maximum price difference between stores and online and the minimum gross margin are enforced by the system; stores buy stock outright, so ordering rights and responsibility for shrinkage go together; and member discounts are recorded by source and settled by the party that bears them.

These directions formed the Reinvention & AI Roadmap, and the operating rules and the positioning of the retail brand were set down in writing.

(c) Implementation. Following the roadmap, we built unified product data first, then the wholesale business, then stores and membership and finally the online store. Unified product data was built first, as the common prerequisite for everything else. The store system was designed from the outset to the standard franchisees would use, so the franchise business could be launched through configuration. Each phase was followed by a review, and the scope of the next phase was set on the basis of actual results.

Existing business continued without interruption during the transition. Wholesale customers kept ordering by instant messaging and did not need to install a new app. In the early stage after launch, sales representatives confirmed orders on customers’ behalf, and every choice they made became data from which the system learned. Once recognition accuracy met the required standard, sales representatives moved to handling only the orders the system could not resolve.

The system’s architecture and content system are described below under “AI Architecture” and “Content System.”

(d) Operations & Maintenance. The system has kept running since launch. Each AI task has a test set built from the business’s real order messages and supplier documents, and a model may be used only after passing these tests. After launch, the first month of operating data serves as the baseline, with improvement targets set each quarter. Management monitors system performance through recognition accuracy, the rate at which staff override the system and the number of daily exceptions, and uses operating data to calibrate, step by step, the range of tasks the system performs on its own. As the capabilities and prices of models keep changing, the system can switch models for individual tasks without any change to its architecture.

The business operating system was designed with one objective: as business volume grows, the work that requires people does not grow in proportion.

Design principles: connectable, composable and evolvable. An AI-native system does not use AI at a single point; every layer of the system has the following three properties.

(i) Connectable. The capabilities of every layer are exposed through the orchestration layer as standardized operations subject to permissions. The project uses MCP (Model Context Protocol) as its open standard. MCP is an open protocol through which AI calls external systems; it is now managed by the Agentic AI Foundation under the Linux Foundation and has been adopted by the major AI platforms. Through MCP, AI calls these operations and can thus reach any layer, but it connects to no layer directly and never changes the ledgers directly; every operation is checked against permissions and rules. Within the system, order recognition, content generation and demand forecasting are performed by the intelligence layer, which the orchestration layer calls task by task. When more capable AI emerges, it needs only to connect through MCP, without rebuilding the system.

(ii) Composable. Every part of the system can run on its own and can also serve as a component of a larger system. In this project, supply chain, wholesale and retail each run independently while together forming an integrated wholesale and retail operating system. The store system was designed from the outset to the standard franchisees would use, so franchised stores join through configuration.

(iii) Evolvable. Any part of the system can be upgraded or scaled on its own, with the impact on other parts confined to their interfaces. In this project, models are configured by task, so changing a model requires only a configuration change; a change to an underlying system’s interface requires adjusting only the orchestration layer; and the online store holds only copies of data, so it can be rebuilt in its entirety.

Composable does not mean building everything in-house. Proven software is used for standard functions, and custom builds are reserved for those that create an advantage for the business, keeping the system’s complexity and cost under control. With these principles, the system the business builds today can keep expanding as AI capabilities improve and the business grows, without being rebuilt.

The architecture rests on four design decisions.

(a) Four layers. The system is divided into four layers, each with its own role.

No entry point and no AI can change the ledgers directly; they can only submit restricted operations through the orchestration layer. This design allows the underlying systems to be replaced, so the business is not locked into any single software package.

(b) Three single sources of truth. Product data, batch-level inventory and available-to-sell quantities are each kept in one place only across the company. Other systems hold copies, and copies play no part in any decision. Wholesale and retail share the same data, preventing the same stock from being sold twice and preventing wholesale and store margins from both being misstated.

(c) Three distinct kinds of capability. What is usually called an “AI system” contains three kinds of capability with different natures, each designed separately in this project:

(d) Division of work between people and the system. For every type of work, decision rights are assigned in advance:

Every AI judgment and every staff override is recorded as the basis for continuous improvement of the system. The system is deployed on accounts owned by the business; the system and data belong to the business, models can be replaced and the business is tied to no vendor.

The business’s product content is generated by an AI Content Pipeline from its unified product data. AI extracts product information from the various materials suppliers provide, and the information enters the database after staff sign-off; statutory labeling items such as ingredients, allergens, nutrition labeling and country of origin must be signed off by designated staff. Routine marketing copy, images and videos are generated by the system on its own and checked by sampling. The same product data generates the content each channel needs:

(a) the price catalog for wholesale customers;

(b) shelf labels and promotional materials for stores;

(c) product pages for the online store;

(d) posts for social media and member messaging;

(e) short videos such as new product introductions, unboxings and cooking demonstrations.

Content for each channel is derived to that channel’s specifications; when product data is changed once, every channel is updated at the same time.

(a) Fully digitized data. Products, inventory, orders and receivables run in a single system, data is consistent across channels and the state of the business can be queried and analyzed at any time.

(b) Evidence-based pricing. Costs at every stage are allocated to individual products, so wholesale pricing and purchasing decisions are based on true costs.

(c) Higher productivity. The system reads orders, generates drafts and produces the price catalog; staff now focus on confirmation and exceptions, and their workload has eased. Growth in business volume no longer requires a matching increase in staff.

(d) Shrinkage and collections under control. The system flags near-expiry stock for action in advance; reconciliation and collections are carried out by the system, with overdue accounts followed up by tier.

(e) Traceable food safety. In the event of a food safety incident, the business can quickly trace the source of a batch and every downstream customer and store.

(f) A new source of growth. The business has established its own brand and operates chain retail stores through company-owned and franchised outlets, together with an online store and membership program, with wholesale and retail sharing a single operating system.