Business Simulation: Test Promotions Before Spending Real Money

20/09/2026

Nearly half of the promotional budgets for retail and F&B chains in Vietnam are wasted due to poor audience targeting or mistimed campaigns. While this may sound shocking, it is the average figure we have observed after years working alongside business teams. The issue is not a lack of ideas, but a lack of the ability to validate those ideas before committing real capital.

The old habit was to 'test' by pushing products to the market and waiting for feedback. The cost of each such 'test' is not just the discount itself, but also logistics, personnel, and most importantly, brand reputation if customers are dissatisfied. We began changing this approach by integrating business simulation models into the decision-making process. Not to replace intuition, but to give intuition data to rely on.

Starting with a Single Production Line

It usually starts with a very specific problem: a factory producing instant noodles or lubricants wants to launch a 'buy 10 get 1 free' promotion for first-tier distributors. Instead of relying solely on the sales team's experience, we build a small model to simulate the impact of this promotion on cash flow over the next 30 days.

This model does not predict exact figures down to the cent, but it identifies the 'safe zone' and the 'risk area'. For example, if a 10% discount is applied, the simulation shows that revenue might increase by 15%, but net profit margins will drop by approximately 20% due to incurred logistics costs. When these figures are presented on a dashboard, the meeting is no longer a debate based on gut feelings, but a discussion of variables: What if we increase the free item to two boxes? What if we apply it only to a specific region?

At this stage, input data primarily consists of 12 months of sales history and a few exogenous variables such as weather or holidays. As a result, we can eliminate about one-third of promotional options that 'sound good' but are financially unfeasible right from the start.

Expanding to the Factory and Supply Chain

Once the model performs well at a small scale, the next challenge is complexity. A promotional campaign does not only affect revenue; it also impacts inventory, machine capacity, and even the upstream supply chain. This is when we begin integrating data from ERP and WMS (Warehouse Management System) into the model.

We have deployed this with several major consumer goods corporations, where every promotional decision must balance dozens of SKUs and hundreds of retail points. The simulation becomes more complex at this stage: it must account for supply chain latency. If you run a major promotion, does the factory have enough raw materials for the next 48 hours? Will the transit warehouse be overloaded?

This is where traditional data analytics tools often 'fail' because the data is too fragmented and unsynchronized. We use Agentic AI techniques to automatically collect, clean, and cross-reference data from various sources, creating a 'digital twin' of the factory and distribution network. This twin allows us to run continuous 'what-if' scenarios. For instance, if a factory in Thailand experiences a disruption, the simulation immediately indicates that goods need to be redirected from the factory in Vietnam, and how promotional levels in the receiving region should be adjusted to avoid local stockouts.

This process requires close coordination between technical and business teams. Often, the biggest barrier is not technology, but departments not sharing data with each other. We act as translators, helping them understand that inventory data is just as important as sales data when evaluating campaign performance.

Full-Chain Scale: From Domestic to Multinational

At the largest scale, business simulation is no longer just a support tool for a single factory, but a decision-making platform for the entire corporation. We have partnered with several clients with cross-border operations, including in Mexico and the Philippines, where variables such as exchange rates, tariffs, and local consumer behavior change constantly.

In this context, the model must be flexible enough to handle differences between markets. A promotional package that works well in Vietnam may completely fail in Mexico due to cultural and purchasing power differences. Machine learning models are trained specifically for each geographic region, learning from that region's own historical data.

Interestingly, as the scale expands, the value of simulation lies not in absolute predictive accuracy, but in risk management. Executives do not need to know exactly whether revenue will be 10.5 billion or 10.7 billion. They need to know that the risk of a loss exceeding 20% is very low if they adhere to the constraints proposed by the model. This gives them more confidence when facing the board of directors, rather than just saying 'we feel this campaign will be good'.

We also find that at this scale, model transparency is more important than complexity. If the model is a 'black box' that no one understands, leadership will not trust it. Therefore, we always prioritize explainable solutions, where every forecasted number can be traced back to its input data source.

A Small Action for This Week

You do not need to build a complex AI system right now to start benefiting from simulation. Start with a very small, specific problem that is currently giving you headaches.

Choose a key product. Get sales data for the last six months. Ask yourself: 'If I reduce the price by 5% next month, how will revenue change?'. You can start with Excel if the data is clean enough, or use basic data analytics tools. The goal is not to have an exact number, but to create a habit: asking 'what if' before acting. Once you are used to asking this question and have enough data to answer it, that is when you are ready to talk to technology partners like AIVISION about building a more professional simulation system. Start with a single cell. Do not let perfection become the enemy of action.

Every company hits this differently, and the hard part is usually the data rather than the model. To pressure-test your case quickly, talk to AIVISION - or first see how we deploy and what we have written before.