In Part One of this series, we looked at pricing- and promotion-related use cases of AI. Here we discuss demand and inventory management use cases for AI.
Demand Planning and Forecasting
The closer we can get to the mind of the customer, the better our demand planning will get. Unfortunately, the mind is not an entity that fits neatly into digital data buckets with fixed field lengths. Demand signals coming from our customers adhere to the customer’s prerogatives and we need to translate and ensure the data quality. In addition, there are so many streams of data that all have nuggets to inform our forecasting, such as SNEW (Social, News, Events, Weather) data and other environmental impacts (such as pandemics). Again, we are confronted with the same challenge — how to organize data so we can process it.
Other elements of critical importance are the forecast methods/algorithms we use against the data.
As conditions change, they necessitate changing the algorithm. The algorithms deployed need to reflect the reality of that product for that channel/customer. Today, most companies are ill-equipped to do this sticking with the same formula for years.
One timely example is that products that had a steady state — a replenishment-type product — may now have lumpy demand or worse. Due to weather or other crises, they may have huge spikes in demand. Or one channel comes to a screeching halt and demand shifts dramatically to other channels, requiring vastly different packaging and logistics. As a result, the forecast method used last week may not fit the need today. Not only do we need the ability to change to different algorithms to forecast, we need to sense quickly, or predictively, that those changes are needed before the crisis hits.
How AI/Machine Learning Helps Demand Planning and Forecasting:1
Many companies have done various scenario plans of sorts — changes in supply, weather, and so on, events that had occurred in the past. To identify various models from external sources to gain some accuracy about the impact of a crisis by geography, for example, companies need to gather these data sources and swiftly apply them within a forecasting model. AI/machine learning is being applied fairly well in these areas.
This leads us to the actual forecasting. AI can evaluate various parameters for their inclusion in the forecast process, and machine learning can be used to select the best algorithm to fit the conditions of each product.
New Product Introduction continues to be an issue that plagues even the smartest companies. Often, we are not clear on what drove demand for our products, especially those with a rich variety of features/attributes. Discerning that is part of the learning process. Thus, we need to begin to create a data model and demand profiles so we can track sales at that next level of detail. Rather than looking at the product family, we can examine how products sell by these relevant attributes.
Attribute-based models can actually be leveraged for many demand challenges. For new product introduction, it is key for product designers to see what actually is creating interest with the customer, unencumbered by markdowns and so on. Initially, the profile for a launch inventory and distribution plan is created by leveraging history. Once the product is launched, these profiles can be changed based on early data coming in from the market. This allows, if needed, lots of changes in the configuration, production cycle, and procurement, which can increase sales and reduce excess inventory. It then informs the next product development cycle early on (which is probably already in process). This fine-tuning of product demand in the current sales cycle can also allow marketing and sales to adjust promotional plans.
Applying these methods historical context now becomes more information-rich. This gets really interesting when mapped against other demand elements such as seasonality, specific holidays (back-to-school, Halloween, Chanukah, and Christmas, for example), geography, customer groupings, or other demand characteristics.
Demand sensing methods are designed to detect changes in the day-to-day, week-to-week replenishment models. Many suppliers apply, not only the weekly forecast, but POS and customer warehouse receipts to modulate their build/ship plans. With AI/ML, demand sensing can take a broader and more microscopic look at those weekly sensings. For example, the week launches with plans in place, and then, icy conditions on the major highways obstruct and delay shipments. AI/ML-based demand sensing can spot these and other short-term events to replan allocations, fulfillment, or just provide assurance that there is still time in the schedule to recover and get back to normal. No action — or additional expense — required.
In volatile demand scenarios, there may not initially be a discernable pattern to apply. A few companies told us that in the early stages of the pandemic, they were just putting thumbs in dikes without a clear view into how demand was shaping up. AI/ML will keep evaluating, seeking a usable pattern to help identify appropriate parameter changes to enable a forecast. As data comes in, the system will keep learning to adjust and build more accuracy into the process.
In building a consensus, machine learning can also evaluate/validate the accuracy of various forecast scenarios, forecasters (sales, marketing, supply chain) as well as the types of data and their sources.
Engaging the voice of the customer. In a world where consumers are constantly voicing their opinions, we need a way to include this in our plans. Natural language processing helps interpret unstructured data — social, text, voice, graphical data — and helps derive insights.2 Over time, some of these sources can actually inform forecasts, as well as be used in product lifecycle plans.
Inventory management suffers from similar issues as the above-mentioned tasks, as there is product across the chain in various stages of production — components or ingredients, finished goods, packaged or staged, with the inventory requirement at that stage often governed by different systems with different rules. As well, there is inventory data from various sources — suppliers in transit with carriers, 3rd party warehouses, in the channel, as well as in use at a customer site. It seems the data issues persist!
Within the organization, there are also stakeholders who have various opinions about what constitutes the “right level” of inventory, from financial and lean-practitioner minimalist to jittery sales teams who want to make sure there is always more than enough stock. This leads to debates about the right rules and planning methods, as justifications for carrying cost expenses are debated.
Also, inventory planning needs to change at each stage of the product lifecycle. For example, at a new product launch, we have no history of how the product will be accepted and at what pace. In this instance in the past, we adopted historical patterns of product sales and closely monitored how things were unfolding. Hopefully, we would have set things in place so we could up or downshift, given actual results. This is true for end-of-life or special promotions. In the recent past, we did have the ability to receive more timely data from the market to respond more quickly. But again, we were basically using the same math. And this is probably more than most companies are actually practicing.
How AI/Machine Learning Can Help Inventory Management:
Learning what impacted demand over time and how well inventory met those requirements is foundational to better inventory optimization. Most multi-echelon inventory optimization (MEIO) systems use efficiency frontier concepts3 to decide on a good inventory number per echelon (location). More modern MEIO, driven by machine learning, differentiates each node’s requirements,4 seeing the interrelationship between the nodes. Then, it chooses the right math to plan inventory for that location (not based just on past data) to look at near-term demand and supply, statistically evaluating the subtlety in consumption curves — steepness, speed, and so on, as well as the external factors that might influence demand and fulfillment lead times to that node.
Then, we need to build a consensus on the right level of safety stock. As mentioned above, creating a consensus is vital to a healthy learning organization. Who gets to make the call on the build or buy plan is critically important. With machine learning, we can look at historical trends and validate a source over time to see how closely it matched outcomes.5 In addition, companies can use external, curated data services and feeds (POS; weather forecasting; logistics performance data such as rail, air, and shipping schedules and delays) and the millions of unstructured sources. The sheer volume and potential complexity of this data makes it nigh impossible for people to visualize and understand all this without some help.6
In the Third and final installment of this series, we consider a ‘rethinking of AI’, looking at data first, establishing a hierarchy of data stores, ‘best fit’ capabilities, and reducing nervousness in supply chain decision making.
1 Here, we want to avoid being too technical so that managers can read and lead without getting bogged down in the nitty-gritty math. These techniques should be provided by an AI/ML-rich supply-chain solution provider. For some excellent detail on specific algorithmic approaches, read 2021 Planning Tip: Eight Methods to Improve Forecast Accuracy. — Return to article text above
2 There is a lot of work involved in leveraging social data within the process of categories, and validating the data. But the rewards are there for the diligent. Over time, this data can be validated against sales, for example, and then be further refined. — Return to article text above
3 such as service level vs. inventory investment — Return to article text above
4 Modern inventory planning systems are augmented with very specific weather and social patterns, and other real-time events to anticipate requirements. This approach, when added to a solid inventory optimizing foundation will note, for instance, more shoppers going online; or that since there is no snow in Colorado, more vacationers will flock to warmer climates for vacation; or that other factors will increase traffic to any one specific location. — Return to article text above
5 for example, source data from other functional groups such as marketing — Return to article text above
6 One analysis was done for a consumer product that appeared to have a seasonal demand curve, but never turned out right — the company experienced large out-of-stock variability over the season. This was easily solved once various web-calendar sources were included which revealed that key events such as local sporting events or large gatherings (conventions, local events/celebrations) were occurring around many of these out-of-stocks. Since each year these event schedules varied future forecasting plans included these major events, and out-of-stocks significantly dropped. — Return to article text above
To view other articles from this issue of the brief, click here.