Retail Sales Forecasting: AI and Human Expertise

26/09/2026

Lessons from a Misaligned Sales Forecasting Project

We once took on a retail chain analytics project for a major beverage company. The goal was to accurately forecast sales for the summer season. Initially, the model performed exceptionally well on historical data. However, during real-world deployment, the results were significantly off. The issue was not the algorithm but the oversight of micro-level factors: a minor local event caused a sudden spike in demand that the system could not anticipate. That experience taught us that technology is merely a tool; understanding the real-world context is the decisive factor.

The Current State of the Forecasting Process

To understand what AI can contribute, we need to examine the current process. The first step is data collection. Accounting staff manually enter sales data into Excel. Next, the sales team analyzes trends based on intuition and experience. Finally, the chain director makes purchasing decisions. This process typically takes three to five days per week. The biggest bottleneck lies in data aggregation. Sales data resides in POS systems, weather data is on the web, and event data is scattered across social media. Combining these sources is a time-consuming task prone to errors.

Where AI Fits into the Sales Forecasting Process

Rather than replacing humans, AI should be positioned to support data processing and analysis. The system automatically collects daily sales data from POS. Simultaneously, it connects to weather APIs and local event calendars. Machine learning algorithms analyze the relationships between these variables. For example, as temperatures rise, beverage sales increase at a specific rate. When a sporting event occurs, demand surges significantly above normal levels. The result is a detailed forecast for each store and product. Processing time drops from several days to a few hours. Staff only need to review and adjust based on their own insights.

The Irreplaceable Role of Humans

Many believe AI will completely replace humans. In reality, that is not the case. Humans still play a crucial role in identifying unstructured factors. A local employee knows that a festival is coming to the area, even if there is no official announcement yet. A sales manager notices that customers are beginning to change their consumption habits. This information is difficult to encode into data. Therefore, the best process combines machine-generated data with human judgment.

Scaling from a Single Line to an Entire Distribution Chain

Successful deployment in a single store or production line is just the first step. When scaling to an entire chain, challenges become more complex. Each store has its own characteristics. A store in the city center has a different customer base than one in the suburbs. A store near a school has different demand patterns than one near an industrial zone. The system must be capable of segmentation and model adjustment for each store group. We have partnered with Masan and Meat Deli on similar projects. Experience shows that standardizing data and processes is the key to successful scaling.

Real-World Deployment in Other Markets

Experience in Vietnam can be applied to other markets such as Thailand and the Philippines. Retail chains there face similar challenges. Weather and local culture significantly impact demand. However, each market has its own unique characteristics. The system needs to be adjusted to fit the specific context. In Mexico, we also see sales data analytics models being widely adopted. Flexibility in system design is the key to success across multiple markets.

What AI Still Cannot Solve

Despite significant progress, AI still has certain limitations. Sales forecasting is not a problem with a single correct answer. The market is always in flux. A minor economic crisis, a change in tax policy, or the emergence of a new competitor can completely alter the landscape. Forecasting systems only indicate probabilities based on historical data. They cannot predict sudden, unprecedented events. Therefore, humans must still play a supervisory and adjustment role. We acknowledge that achieving absolute accuracy is impossible. The practical goal is to minimize error and help humans make faster, more accurate decisions. To achieve this, continuous investment in both technology and people is required.

There is more here than one article can hold. Keep reading on the AIVISION blog, look at our display scoring solution, or get in touch.