This article is an excerpt from the report AI for Supply Chain: Debunking the Myths.
The full report can be downloaded here.
Introduction
We have all heard claims about how AI and Machine Learning will make us smarter and more productive. But specifics are often hard to come by. Our Supply Chain Planning systems have gotten so much better over the years, so, practically speaking, why do we need AI in the mix now?1
In addition to the big claims, there are blogs with lots of intimidating terminology such as random forests, multi-objective optimization, perceptron learning, and Naïve Bayes Classifiers,2 along with lots of AI categories such as Artificial Neural Networks, Cognitive Computing, and so on.
Then there is all that data — big data, web data, and all that data with a funny abbreviation — SNEW — social, news, events and weather! The experts3 say we need all of that in our AI and machine learning tool kit. And what about those newly minted data scientists? The experts say we will have to hire them, too.
These would seem to be major impediments to implementing AI/Machine Learning — expensive and resource draining. But the fact is, those blogs and technical publications contain a lot of myths.
AI/machine Learning is somewhat new for supply chain, so we need to get the facts.
This series, AI for the Supply Chain, is geared towards supply chain leaders, managers, and planners who want to get the facts and explore what AI/machine learning can do for their supply chain.
Why get discouraged by the exaggerated complexities before you even start? In this first installment, we want to dispense with some of the myths that surround AI/ML. We will also give you a quick overview of the options in the AI/machine learning solution landscape and what you might need to add to your portfolio.4
Why AI? INTRACTABLE SUPPLY CHAIN CHALLENGES
So, why do we need AI/machine learning?
You have heard it before. And you’re living it. Markets are volatile, the workforce is changing, the customers are changing, and the world of data is vast and, often, unnavigable. Most of the material about our future sounds like so much noise and marketing about how great our future is going to be. However, in supply chain, we are supposed to make quick, accurate and actionable sense of this — not just make general observations.
Yes, things have changed. Customers, both B2B and B2C, now expect an instant, 24/7 response. Yet, in volatile markets, succeeding at that is a huge challenge. And with the web’s infinite catalogue, tomorrow’s customer is a search away — and maybe a sea away.
Any executive has to be asking:
Is my window to the world information rich?
Can my organization respond to the challenges we face?
Do we even know what challenges are ahead?
We have been grappling with huge disruptions in our markets and supply chains which we did not anticipate. How can we avoid that in the future? We need better insights!
Yes, there are new problems to solve due to this changing world, yet we still have intractable ones that supply chain professionals have been grappling with for decades that seem to be always slightly beyond the reach of even the best planning systems.
Why?
Firstly, we didn’t have the computing power and speed provided by high-performance mega-servers to make processing vast data streams and unstructured data practical. Secondly, we just didn’t have the data in the past, often depending on not so dependable one-up/one-down trading-partner data. Today, we can get high fidelity, almost-like-you-are-there data about our world markets, consumers, supply chains, and the environments they operate in. But a lot of the math that is used in those traditional forecasting systems is based on concepts developed centuries ago. Yes. Centuries. Joseph Fourier, of the Fourier series fame, was born in the 1700s. George Box and Gwilym Jenkins, creators of the Box Jenkins Method, are more recent history, having developed their methods in the 1970s.
Of course, over time, we did get better math and planning systems. Yet, we still have those lingering intractable issues. Here are some examples:
New Product Introduction. Whether it be a whole new product or a new product feature, NPI continues to be problematic. What attributes are customers looking for? Are we targeting the right market sector with our features, color palettes, and designs? What should be the initial volume produced and of what assortment per feature? What price should we offer it at? What will be the impact on existing products, on our competitiveness? If we lower the price on the existing model, will customers consume our old inventory or just opt for the new version?
Not answering these questions well means we have lost even before the game begins. And we often lose, since, today, there are over 80,000 new products introduced each year, with a failure rate of 95%!5 Even successful launches often suffer from ill-fitting inventory or logistics strategies. It appears from these misses that many organizations are still practicing NPI with a dartboard.6 That is intractable problem number one.
Stock-outs.
We know they happen. But knowing why and being able to avoid them can elude us. Today, we mostly rely on history and weekly forecast data. Although the week’s numbers seem to balance out, we can actually be losing sales because on any given day we may be short. Our fallback position is generally to pour on more buffer stock. But we just can’t afford that with global competition eating away at unit pricing.7 That eats away at our profit. We need a better approach.
We need to know where and why these events are happening. This means we need to take a much broader, yet more granular, view, looking across the supply chain while it is executing. We have to create a cycle of continuous planning so we can determine which factors are affecting performance at this moment, for this product, for this channel and this location. All this needs to be locally granular and the source data needs to be accurate. What kind of data are we talking about? Data could include environmental data — supply chain disruptions,8 weather, holidays and other events that may divert people from going to the store; logistics — warehouse and transportation issues; supplier performance; and consumer data. Then we need to examine our assumptions, such as lead times or standard safety stock. So much of the assumption data we use is often fixed rather than dynamic (responding to the current realities). In other words, the model is not big enough or responsive enough.
To understand why, then, we have to broaden the model to include views about the store and its shoppers and our supply-side dynamics. We need analytic tools (algorithms) to fine-tune safety stock calculations of our traditional systems, and be able to do this automatically.9
Optimizing Inbound Through Outbound Fulfillment
This is another challenge. The irony is, in our quest for supply chain visibility, while we clearly have more data, this visibility has highlighted just how complex something like a “simple” purchase can actually be. AI/ML can take the next step of absorbing complex data, which becomes extremely difficult for users to analyze on their own. For example, solving a multi-mode, multi-leg, multi-stage fulfillment challenge has very many variables. Within this we are looking for a lot of the “rights” — right cost, right timing, right route. Which of those do I want to solve for in a circumstance? The parameters for this circumstance should represent what should work for this one order or shipment, not an averaged approach.
Modern AI/ML systems fuse different types of advanced algorithms to fine-tune and broaden optimization, and can deepen learning or pose new questions.
Making better margins. This is an example of deeper analytics and learning. Suppliers cater to their biggest customers, who often represent a higher percentage of their sales. But do they yield better margins?10 What if the next tier’s sales yield greater profits? Maybe I ought to be focusing more on those customers to grow my sales with them. As well, I ought to be analyzing where there are opportunities to increase margin with the tier-one customers. This type of analysis requires understanding patterns of behaviors — work tasks in logistics, additional services, inventory agreements (holding/carrying costs), special packaging or labeling, and financial issues such payment terms, which can eat away at the margin.
This type of challenge is also present in planning product quantities and assortments for new or seasonal products. Unfortunately, most organizations use history as their only guide here. However, consumer needs change and market dynamics are constantly changing (in case you did not notice).11 Machine learning can process vast quantities of data from many sources to uncover important trends and insights about consumer sentiment and their evolving needs, as well as market volatility. It also can consume real-time sources to discover urgent patterns that may only show up between the orchestrated planning cycles (weekly, monthly, or longer) that set inventory strategy. Machine learning should ultimately move us from orchestrated forecasting exercises to continuous planning.12
In the next installment of this series we look at several AI myths around how AI is different than the human brain, AI autonomy, the importance of data quality, and challenges of building an AI system from scratch.
_________________________________________________________________
1 AI/ML is not exactly new. Some supply-chain solution providers have had AI in the mix for years. However, it is also true that there are many new developments in the AI world and, now, the resources to successfully deploy these technologies. — Return to article text above
2 They classify a web page, document, or text. — Return to article text above
3 consultants, media, and bloggers, who are often established technical professionals in business or tech companies — Return to article text above
4 In subsequent installments, we will go into use cases. — Return to article text above
5 Estimates vary, but we can all agree the scale is huge. And those market flops include all that design, launch, distribution, and excess inventory expense. — Return to article text above
6 generally, in the domain of spreadsheets — Return to article text above
7 And for certain items that buffer might not even be available. — Return to article text above
8 such as current trade wars or COVID-19 — Return to article text above
9 Fact is, we probably can’t rely on the standard replenishment models of the past. Read: Demand Planning in Uncertain Times. — Return to article text above
10 Many demanding customers put huge burdens on our supply chain operating costs; the perceived margins may not be there. — Return to article text above
11 Global trade and risk have been highly uncertain in the last two years due to systemic (changes in trade agreements, Brexit, sanctions, etc.) and geo/social/environmental factors (such as the coronavirus pandemic). — Return to article text above
12 S&OP and formal reviews must continue; but we should also have a continuous view into demand and supply to take advantage of market dynamics and mitigate risks. — Return to article text above
To view other articles from this issue of the brief, click here.