Why AI Hits a Nerve

Why is it hard to run a business? Simply put, because we cannot know the future. From estimating market sizes to production costs to transportation times to price sensitivity and the impact of marketing campaigns, we just do not know exactly how to quantify them.

A Lucky Break

Take Rockefeller's Standard Oil Company. The original plan was to create a lamp oil that would not suddenly combust in the home. To get the cost for transportation down (before he came up with the idea of building cross-country pipelines, this meant using trains to bring the oil from Texas to the Northeast, where most people lived), he needed an economy of scale and consequently bought every oil field he could get his hands on.

Then Edison, backed by Chase, invented the electric light bulb. Ouch.

Without the invention of the internal combustion engine to power cars, Rockefeller would surely have gone bankrupt. Instead, he then kept the combustibles that were waste before and made those his product. When eventually it was forced to split, even the splinters of the company became massive businesses like Exxon, Mobile, BP, Amoco, and Chevron.

Such long-term technological developments are very hard to forecast. But the pain of running a business because of uncertainty is a daily one. How big is the demand? How much will sell at this price? If I run this campaign, what are the short-term and long-term effects?

We just do not know.

The False Prophet

Enter machine learning.

• "We can make the machine learn!"

• "No need for complicated algorithm development, we use data!"

• "We can tell you the future!"

To someone running a business, these naturally sound like the promises of a quack selling snake oil for exactly your pain and ailments. Finally, no more guesswork! Now we are "data driven." We will get much better operational results, increase revenue, and lower costs; the promised land is near.

After a decade of investing in very expensive data infrastructure and even more expensive machine learning, the results are sobering. ML has undoubtedly had an impact and yielded results, especially in anomaly detection and predictive maintenance. But it is far from the silver bullet it was advertised to be.

Instead of carefully analyzing why ML is only a piece in the puzzle but not the be-all and end-all tool, we now see influencers who, in all seriousness, suggest asking large language models to determine when and how much to order for replenishment. I have no doubt that, embedded in carefully crafted protocols, LLMs will soon be able to invoke analytics, but they are not at all equipped to answer such business questions by themselves without orchestrating specialized agents that are suited for the task.

So where is the gap?

"Make Perfect Predictions, Please!"

I have heard of this request a lot lately, most recently from a colleague at a large automotive company. It is time to counter it with a very clear message:

Perfect predictions of the future are impossible.

That is the bottom line. It can be said differently, depending on your audience:

• Technical: "There is not enough signal in the observable features to perfectly determine the dependent variable."

• Behavioral: "People's decisions are influenced by emotions, biases, and social trends that cannot be perfectly quantified and predicted."

• Political: "A sudden tariff, a new environmental law, or a political event can disrupt supply chains or change consumer behavior, making the future impossible to predict."

• Environmental: "We always face natural disasters, pandemics, and other unpredictable global events. A major hurricane can shut down a factory, a pandemic can change consumer priorities, or a crop failure can cause commodity prices to spike. We cannot predict the future perfectly."

Therefore, forget about splitting the arrow. We will never have perfect predictions. Period. What is needed is a tool for making decisions under uncertainty.

Never Ignore Uncertainty

So, is that the end of data-driven decision-making? Absolutely not! But we have to do better than to say, "Make your best forecast, and then we plan for that." This is a bit hard to grasp at first, but the issue is that plans that are optimal for a specific future are frequently a disaster for futures that deviate only slightly.

Take air travel. Say you are looking for an itinerary and you need to switch planes. You have an inbound arrival time and an estimate of how long it will take to switch planes. Obviously, the best outbound flight is the first one that, according to the forecast, you can catch. Or is it?

Well, an optimizer that just uses the forecasts at face value without taking the uncertainty into account will choose that connection because it gives the lowest overall travel time. Alas, sometimes our inbound flight is a bit delayed, and other times the immigration takes longer than average. And boom, not only did we miss our connection, but the next two flights are fully booked as well, and we are stranded at the airport for many hours.

Some optimization vendors advise their clients to just use the best forecast, ignore all uncertainty, and create an optimal plan for that. See this Colab with a critique of a competitor as an example: https://colab.research.google.com/drive/1SHzvxrWdkywjsmfMb1fDpCDBDm-Yn2ZC. In this example, the competitor used forecasts deterministically and lost the client close to $1 Million per week as a result. While claiming their plans were optimal, mind you.

And this is the real reason why optimization has such a hard time in supply chain planning and why its impact is so very limited in practice. Because it assumes that forecasts were 100% accurate, and that is simply not the case.

The Solution

Is there a holy grail? No. It will always be hard to run your business in the face of uncertainty that cannot be magically predicted or defined away. But there is a way for you to run your business and your operations in a data-driven fashion that will give you an edge over your competition.

Here is your recipe.

1. Identify what decisions that you make in your business affect the top and the bottom line the most. Where would making better decisions help you gain market share and be more profitable?

2. Write down what data and information are currently being used to make those decisions. Importantly, note how those predictions frequently differ from what then really happens.

3. Create a model for a solver like InsideOpt Seeker that allows you to optimize your decisions so that the suggestions from the solver will work efficiently and resiliently for a big probability mass of futures and not just the point-predicted future.

4. (Optional): Only now consider where tighter and more accurate predictions would result in better decisions.

5. (Optional): Only now consider what additional data sets will help make tighter and more accurate predictions.

If you follow this workflow, you will recover your investments the fastest, because the first thing you build is the tool that actually helps you make better decisions under uncertainty. And only those decisions can turn data into dollars. Instead of hoping for vague "insights" by building data infrastructure and expensively curating data and making forecasts without a plan for how to use them, you immediately leapfrog to the step that creates the value. If you choose to invest in better forecasts and data later, that is fine, but at least then you have a clear path to value, and you avoid investments that never improve your business.

Most importantly, you get efficiency and resiliency for smooth and profitable business operations. Even without perfect forecasts.