On August 20, 2025, Zionex was invited to participate in the AI Café Tainan session, organized by the Institute for Information Industry (III) and the Artificial Intelligence Innovation Application Promotion Association (AIF). As a speaker and industry representative, we shared insights on the applications of AI in manufacturing and brand operations.

The topic of the presentation was “Breaking the Five Major Myths of AI in Enterprise Applications.” Through real-world case studies and interactive discussions, we guided attendees in overcoming common misconceptions and gaining a renewed understanding of the true value of AI in sales forecasting and operational improvement.

🔍 Five Major Myths and Case Studies Sharing

Myth 1: Is AI / ML Really Necessary?

This question doesn’t really have a single “standard” answer. Traditional statistical models are indeed useful — they rely on fixed formulas and trend extrapolations, which can provide good results under stable conditions. But the strength of machine learning models lies in their ability to continuously learn from data and dynamically adjust, especially when demand is complex and influenced by multiple factors.

A real-world example:
We worked with a globally renowned food company (its main products are tofu and related food items). Their demand patterns were highly complex — not only affected by seasonality and promotions, but also by health and dietary trends as well as channel strategies. In the past, relying only on historical data and simple statistical methods often led to overstocking or stockouts.

After introducing an AI-based forecasting model, the system was able to incorporate both internal sales records and external market factors, while selecting different models for different product lines. The results were significant: forecast accuracy improved, production plans became more closely aligned with actual demand, and inventory pressure was greatly reduced.

So, going back to the question: “If we only have historical data, do we still need machine learning?”

The answer is: it depends on the complexity of demand.

If the market is relatively stable with little variation, traditional methods are sufficient. But if demand fluctuates greatly and is influenced by many factors, that’s where AI / ML can truly demonstrate its value.

Myth 2: If AI Isn’t Accurate Enough, Is It Not Worth Using?

Let’s start with a sharp question:

“If AI’s accuracy is only 85% or 90%, and it can’t reach 95% or even 98%, does that mean it’s not worth using?”

This concern is actually quite common among enterprises.
But the truth is — no model can ever be 100% accurate, because demand itself is inherently uncertain.

The value of AI does not lie in delivering a perfect answer, but in providing early, directional signals.

For example:
If AI can improve forecast accuracy by just 10%, for a company managing thousands of SKUs across multiple channels, that difference could translate into tens of millions in business value.

That 10% improvement brings more than just a number — it enables:

• More precise inventory allocation
• Fewer stockouts and less waste
• More stable production planning

Myth 3: Can a Single Model Forecast All Products?

At the event, many attendees also asked: “Can we just use one model to forecast all of our products?”
This is a common expectation among enterprises, but in practice, it’s very difficult to achieve.

The reason is simple: different products have very different demand patterns.
Some products have high and stable sales, some fall into the low-frequency long-tail category, while others fluctuate wildly. If you force a single model to handle all products, the outcome is usually that some products forecast well, but others are completely off the mark.

Our approach is to first help clients classify their products.
For example, through ABC-XYZ analysis, we group products based on sales value, volatility, and data completeness:

• Some products are suitable for machine learning models.
• Some can be handled well enough with statistical methods.
• Others require customized forecasting strategies.

Next, we design a multi-model strategy, such as:

• Prophet: good for seasonal and elastic demand.
• Croston: effective for intermittent demand.
• XGBoost: well-suited for highly volatile, high-dimensional data.

The truly effective method is not “one model rules them all,” but dynamically selecting the best combination to match different product characteristics.

In addition, we emphasize hierarchical forecasting. This ensures that:

• Headquarters can view aggregate numbers,
• Marketing can focus on brands,
• Supply chain teams can look at SKUs,

all starting from the same forecasting logic.

This design guarantees that forecasts across levels do not contradict each other, and the final results can actually be collaboratively used, executed, and trusted in production and scheduling.

Myth 4: Once AI Software Is Implemented, Will Forecasts Run Automatically?

Many clients, when first introduced to the system, often ask: “Once the AI system goes live, does it run completely on its own without human intervention?”

Here’s how we usually respond, supported by actual practices:

• Step 1: The ML model generates an initial forecast (Baseline Forecast).
• Step 2: Different departments collaboratively adjust the forecast to ensure it aligns with real business situations.
• Step 3: We continuously monitor the gap between forecasts and actual outcomes, identify discrepancies, and refine the model through a PDCA cycle (Plan–Do–Check–Act).

AI is not a plug-and-play magic box. It creates value, but only when embedded into a company’s daily processes. Through ongoing collaboration and optimization, forecasts become increasingly accurate and practical, effectively supporting production, scheduling, and decision-making.

In short, AI is not a “set-it-and-forget-it” system — it is an intelligent tool that requires continuous engagement and adjustment in order to become a reliable foundation for business operations.

Myth 5: After Implementing AI Forecasting, Should We Only Focus on Accuracy?

Accuracy is certainly important — but the true value lies in improving resource allocation and operational outcomes.

Take a high-tech electronics company as an example:
Their product portfolio was very broad, ranging from high-frequency standard items to highly customized, long lead-time products.

In the past, they applied a single approach to all products, which often led to either material shortages or excessive inventory.

Zionex helped them by implementing product segmentation and differentiated forecasting strategies:

• Stable products: Used baseline forecasts combined with inventory stocking.
• Volatile products: Applied demand planning with rolling adjustment mechanisms.
• Long-tail or customized products: Relied entirely on orders, enabling rapid response through a make-to-order (MTO) strategy.

Results achieved:

• On-time delivery rate improved to 97%
• Unsellable inventory reduced to 3%
• Inventory turnover days improved by 45%
• Delivery lead time shortened by 65%

This proves that implementing AI forecasting isn’t just about chasing accuracy. The key is combining it with product characteristics and operational strategies so that AI can truly transform into tangible business value.

If you’d like to learn more, please fill out your contact information, and our team will reach out to you!