Standardizing Product Classification & Automated SKU Management

10/09/2026

Last month, I sat across from the operations director of a fast-moving consumer goods (FMCG) brand in Ho Chi Minh City. Their AI-based product recognition project was three weeks behind schedule. The reason wasn't that the algorithm failed to run, but that the input data was too chaotic. The same 4L motor oil was recorded as "Oil 4L" at Branch A, "Motor Oil 4 liters" at Branch B, and "Oil 4L" at the central warehouse. I wouldn't dare say what the AI system actually learned. The lesson here is simple: if you don't clean your data first, AI just helps you replicate the mess faster. We need to start by standardizing the product classification process, rather than jumping straight into technology.

Step 1: Redefine the Product Classification Framework

Before feeding any data into the system, you need to sit down and agree on a common language. This is the most time-consuming but most important step. Your inputs should include a list of all current SKUs, the category sets currently used in your ERP or WMS, and a list of common error codes. The task is to eliminate duplicate names and establish a clear hierarchical structure. For example, instead of just writing "Milk," define the standard as "Category: Food > Group: Dairy & Beverages > Type: Fresh Milk > Variant: 180ml." The visible output is a single Excel sheet or CSV file where each product has only one name and one unique code. You know this step is complete when two warehouse staff members at different branches look at the same product and call it by the exact same name, without needing to ask anyone else.

Step 2: Clean and Standardize Existing Data

Many businesses skip this step, assuming old data will disappear on its own. It won't. Legacy data is the root of all future conflicts. The input is the entire inbound history from the last 12–24 months. The task is to use logical rules to consolidate name variants into standard codes. If you have large datasets, manual processing is impossible. Here, natural language processing (NLP) tools can help detect typos or near-duplicate names. I once worked with Masan and several instant noodle companies, where just the flavor names had dozens of variations. We used approximate matching algorithms to consolidate them. The output is a "clean" dataset where each SKU corresponds to exactly one standard description. The completion metric is a reduction in duplicate or ambiguous records to an acceptable level, typically under 5%.

Step 3: Build an Automated Code Assignment Model

Now we get to the technology. The inputs are the clean dataset from Step 2 and new product photos or descriptions. The task is to train a machine learning model, either text-based or computer vision-based, to recognize new products and assign the correct classification code. You don't need a complex model from the start. Begin with a simple model, such as classification based on product names and a few basic attributes. Then, expand to barcode reading or packaging image recognition. The visible output is a system capable of suggesting a SKU code for each new product with a certain level of confidence. You know this step is done when the system can correctly assign codes to at least 90% of familiar products and provide clear warnings when it encounters an unfamiliar product it is uncertain about.

Step 4: Integrate into Multi-Branch Inbound Processes

A model that runs well in a lab is one thing, but running well in a warehouse at 5 a.m. is another. The input is your current WMS or ERP system. The task is to embed the AI model into the inbound process. When staff scan a code or photograph a new product, the system automatically checks and assigns the classification code. If confidence is high, it accepts automatically. If low, it routes to a queue for human review. It is crucial to ensure consistency across the entire chain. A product coded at a central warehouse in Binh Duong must have the same code and category when it reaches a distribution warehouse in Hanoi or a partner in Thailand. I have seen this work effectively when working with lubricant companies, where data inconsistencies previously cost them weeks in inventory reconciliation. The output is a seamless inbound flow with minimized manual data entry. Completion is marked by a significant reduction in processing time for new batches and the elimination of instances where two warehouses record different codes for the same product.

Step 5: Monitor, Feedback, and Scale

The work doesn't end when the system goes live. Data always changes. New products launch, packaging designs change, and new data entry errors emerge. The inputs are daily activity logs and exceptions handled by humans. The task is to establish a feedback loop. Every time a warehouse staff member corrects a code incorrectly assigned by the AI, the system must log that error and use it to retrain the model. This is how AI becomes smarter in real time. You also need to track accuracy metrics and manual intervention rates. If intervention rates rise, it means input data is becoming polluted again or the model is becoming outdated. Scaling from one line, one factory, to an entire retail chain is a cautious process. Don't try to apply it to the entire system immediately. Start with one warehouse and a specific product category. Once stable, replicate the process to other branches. To conclude, let's return to the operations director I mentioned at the beginning. After we applied the correct process from Step 1 to Step 5, the project not only caught up on schedule but exceeded expectations. The AI system is no longer a worrying black box, but a virtual warehouse employee that is always consistent. Clean data, standard SKUs, and peace of mind in management have returned. That is the true outcome of proper automation.

This series comes out of projects that actually shipped. More on the AIVISION blog, details on face recognition and the rest of our solutions, or reach out to us.