Better demand forecasts mean better inventory management. Right? This basic premise fuels billions of dollars in AI technology investments. Companies are in a race to implement machine learning technology that promises unprecedented demand forecast accuracy.
What if this entire premise is backwards? What if the entire focus on improving demand forecast accuracy, with or without the help of AI, represents a misplaced opportunity?
Demystifying the Technology
What do we mean by AI, and how does it apply to inventory management? We mean machine learning technology that analyzes historical data to create forecasts. Machine learning technology analyzes sales history, promotional activity, pricing data, competitor activity, weather, and countless other data points.
The technology is impressive. Machine learning technology can analyze data that would take a human analyst months, if not years, to analyze. It can identify relationships between data points that would never have been possible to analyze manually.
This produces tangible benefits in certain circumstances. If demand patterns remain stable and historic data can be used to forecast future behavior, then machine learning can be used to enhance forecast accuracy by a tangible degree.
However, there remains a basic constraint that cannot be overcome by any machine learning technique: the ability to forecast the future depends on extrapolating from the past. If the future looks like the past, then machine learning techniques can be used to forecast future events. However, if changes have been made to the marketplace, competition, or unexpected trends have emerged, then even the most advanced system will be projecting the past into the future.
The system doesn’t understand your marketplace. It can identify patterns in data without any comprehension of the business reality those patterns represent.
Dangerous Illusion of Certainty
This is what makes the AI forecasting process so tricky. With traditional forecasting methods, the numbers were generated, but everyone knew that the numbers were just an approximation. The element of human judgment was clear. Everyone remained healthy skeptics when it came to the accuracy of the forecast.
But with the AI forecasting system, the presentation of the data is different. The advanced algorithms, the sophisticated presentation, and the precision of the numbers all create an illusion of certainty. The teams that work with the AI system assume that the advanced algorithms used to generate the forecast make them more accurate than they really are.
This is the problem with the AI system. The system could be projecting historical trends that no longer apply, but the presentation does not show that. The presentation appears to be definitive, even when the entire process is highly speculative.
You get precisely wrong forecasts presented with absolute certainty. The system will project the demand with incredible precision down to the decimal point, yet the actual results will vary by wide margins.
Examining the Real Problem
The majority of problems with a business’s inventory can be attributed to structural problems. Long supplier lead times mean that your business is investing capital weeks or months before it is actually used. Rigid ordering systems prevent your business from making frequent changes. Rigid supplier relationships prevent your business from making quick changes. Product portfolio decisions, also made months before, have turned out to be bad decisions.
Then there are the demand volatility factors. Competitor actions that change the market overnight. Social media marketing that appears out of nowhere. Consumer behavior changes that occur before your business can react. Changes in the economy that affect consumer purchasing behavior.
And what does improving your forecasting really help with? Your business is now 80% accurate instead of 65%. That is a real improvement. However, there is still 20% uncertainty, and this is what causes most of your business’s problems.
Better forecasts don’t eliminate demand volatility. Better forecasts don’t shorten lead times. Better forecasts don’t make your suppliers more flexible. Better forecasts don’t eliminate competitor disruptions. Better forecasts incrementally reduce one form of uncertainty while leaving all the underlying issues unchanged.
When AI Confidence Meets Market Reality
Consider the last few years. How many organizations have found themselves in a situation that the AI could never have predicted? Supply chain disruptions are beyond the historic data used to create lead time predictions. Demand shifts too fast for a quarterly forecast model to keep pace. New competitors whose impact on the marketplace happens in an instant.
In those situations, sophisticated algorithms become a hindrance. AI makes predictions based on historic trends that no longer apply. It looks so official and so confident, so we believe it longer than we should. We’re making decisions regarding inventory, locking up working capital, using precious warehouse space, and incurring all the costs associated with carrying inventory based on predictions made by your AI system, predictions based on yesterday’s reality. By the time the AI system has a chance to learn from new trends and adapt its predictions, we made costly decisions.
This is especially a problem in fashion-sensitive categories, trend-oriented products, and markets with a lot of competitive activity. The environments where inventory management is hardest are the same environments where, ironically, forecasting of any kind, including AI-powered forecasting, is least capable.
Escaping the Forecast Trap
What if we can manage inventory without relying on forecasts of demand into the distant future? What if we can react to what is happening instead of trying to predict what will happen?
That’s exactly what Dynamic Buffer Management promises to do.
The beauty of DBM lies in its simplicity. Assign a buffer to each item, not a specific optimal buffer, but a buffer that will suffice. Then, track how much of that item is used over time. If the item tends to run critically low, boost the buffer. If it tends to have plenty of inventory left over, cut back on the buffer.
This removes the forecast dependency, which is a major source of problems. You’re not committing capital based on a forecast of uncertain future demand. You’re committing capital based on a direct reading of what has actually been consumed.
This system learns from reality rather than history. It reacts to change immediately, rather than waiting for new patterns to emerge in history and then be reflected in a new forecast. It adjusts continuously based on what is happening, rather than periodically based on what we think will happen.
Discovering AI’s Real Value
This doesn’t mean that AI has no place in inventory management. It means we’ve been using it on the wrong problem.
AI is great for strategic analysis, where historical pattern recognition can add real value. What kinds of products are growing or declining? Where are regional demand patterns significantly different? What kinds of supplier relationships are associated with top performance? What kinds of product portfolios generate top returns?
These strategic questions can only benefit from the capabilities provided by machine learning’s pattern recognition abilities. “You’re working with complex relationships over enormous sets of data.” “You’re seeking answers that would take a human analyst months to discover.” “You’re making decisions where a 15 percent improvement in strategic direction can yield significant value.”
Use AI for assortment management. Use AI for long-term portfolio strategy. Use AI for supplier and category analysis. Use AI for capacity planning decisions. These types of decisions play to the machine learning model’s actual strengths.
For order management, daily replenishment decisions, and quantity calculations for individual items, use responsive systems. This means using systems that react to the present reality as opposed to predictions about the future. Dynamic Buffer Management should be used for the operational decisions.