How can AI help boost your company’s bottom line?

How can we use artificial intelligence to improve our business? Can we use it to free up resources and spend less time on trivial routine tasks, can we use it to become better at predicting the future, and can we use it to develop new products and bypass the invention process ourselves?

There are plenty of examples of companies that have used machine learning with great results. But critical voices have begun to surface about the returns on AI investments.

A recent survey made by McKinsey concluded that only 8% of AI projects had made it past the pilot stage.

That might explain why you’ll have to google hard to come across a case study of a traditional brick-and-mortar company that uses AI to save cost or increase revenue.

So what does it take to leverage AI in your business successfully? In this article, we’ll discuss five essential questions to consider; your answers will reveal whether AI is right for you. The questions also make it clear why some companies have great success with AI while others struggle.

Data culture – ready to be data-driven?

There is no doubt that AI has been the main star on the consultancy runway for the past couple of years. The exponential growth in information and AI applications is sometimes seen as a cure-all just like computers were back in the 70s. Not surprisingly, the web is sizzling with juicy concepts where all planning in a not so distant future is augmented by AI. These visions build on the premise that because supply chain, by nature, is big on data, planning tasks are primed for AI-supported decision making. I’m pro-science and I’m upbeat about the future of AI but based on my experience I think these views are badly mistaken.

The concept of being data-driven has existed for decades and yet in my experience, I have often seen that many potential improvements in supply chain projects are in fact related to psychology. AI will not change that.

So why is it so hard being data-driven? In inventory planing for example, the goal is to strike the perfect balance of not ruining the bank with capital tied to products while at the same time being able to meet customers’ expectations of availability. So on paper, everyone from sales to logistics seems to agree that stocking out is a necessity.

The reality is somewhat different. I have yet to meet a planner who is comfortable when stocking out occurs.

Why? Maybe because having high inventory is less nerve-wracking than running the risk of stocking out. Maybe because the CFO who looks out for working capital is unable to reach the same decibels as the sales manager in stock out situations?

The emotional imbalance between over and under stocking is from my experience what leads to overstocking 9 out of 10 times – despite the fortunes spent on IT.

To take full advantage of AI data capability, the human element has to be on board and more often than not, companies lack the processes to make that happen.

Resources – can you explain AI?

Assume you work as a data analyst. When you train and test your AI model for predicting sales you get good accuracy and a convincing positive predictive value. You bring it to operations, and they agree it seems to be a powerful model. But they will be hesitant to use it because you (or the model) cannot answer the simple question:

“Why did you predict increasing sales when the baseline seems flat?”

The field of explainable AI has grown in recent years, and this trend looks set to continue. There are exciting and innovative avenues AI experts are exploring in their search for models which not only perform well but can tell you why they make the choices they do. Some consultancies have started to emerge in this field and serve their clients like extended planning departments providing the data crunching resources that are often difficult to hire and keep.

So, if you are to invest in AI make sure you have the resources in-house or consider teaming up with an expert who can translate AI to plain language.

Data quality – is it suited for making predictions?

What you are looking for, what you are looking at, what you buy is all turned into digital footprints that Google maps and uses to predict your next move with AdWords. Traditional companies created in the physical world, however, are subject to a different reality.

Take a traditional company like Coca Cola that sells goods to supermarkets, which then sell to the end consumers. Unlike Google, Coca Cola has a filter between their product and the end consumer. If there is a Coca Cola campaign in a supermarket then consumer sales increase, but it is not a guarantee that Coca Cola at the same time sells more products to the supermarkets – it depends on how much supermarkets have on inventory.

So there can be a time lag between Coca Cola sales and what is actually sold in the supermarkets. That makes it difficult to draw a strong correlation between price and sales.

It may also be that a supermarket is starting to increase its orders of Coca Cola, which makes it look like sales are growing but in reality, it is because Pepsi’s shelves are empty due to delivery problems. This situation is temporary, but it makes it look like sales are rising.

While some supermarkets, especially large ones, have made it a business to sell these data back to their suppliers, it is still a significant challenge for companies like Coca Cola to get a complete dataset from all their customers.

The problem with not having a high-quality dataset is that Coca Cola cannot see the full context of their sales. They can see data in motion, they can detect whether sales are rising or falling, but they have no clear picture of the cause.

Google is the diametric opposite.

Google predicts your actions based on context; they know where you’ve been, what you’ve been looking at, and can look ahead to build a complete picture of who you are. They have thousands of data records from the purest data source: you.

Data access – is yours locked up?

Deep inside the engine room of Manhattan’s financial district, Goldman Sachs has a Formula 1 team of mathematical experts to build models that have so far secured solid returns. The ingredients are a combination of stock prices, exchange rates, historical indexes of market indices. It is information that is accessible to everyone, facts recorded in real time, so no data cleaning is required – and even better, it’s all then imported directly from the net without any kind of intermediary. It’s probably the closest we can get to AI nirvana, a place where Mozart can compose without thinking about laundry.

Coca Cola spends billions on promotions.

Can Coca Cola apply the same methods as Goldman Sachs to predict how price impacts sales?

First, they need access to historical prices and sales to consumers. This is held by the supermarkets – so here Coca Cola needs an intermediary to get that information.

Also, movements in Coca Cola sales cannot be isolated to price alone. What their competitors do also have a significant impact. So if the Coca Cola prediction model is to be robust, they need to obtain data on Pepsi’s price points, their campaigns, and advertisements.

There are analysis institutes that sell this kind of knowledge, but again, it’s a third party source – no direct data capture like Google.

Data volume – how much data do you have to work with?

As a starting point, machine learning requires large amounts of data. There’s no magic minimum, but most successful machine learning projects start off with thousands of observations to work with.

Let’s say you want to predict sales for the next 12 months. You have three years of sales history, so 36 observations of monthly sales to work with. This is insufficient for machine learning.

But if you change the prerequisites for your calculation and instead of months use daily sales, then you have increased data volumes from 36 to just under 1000 observations. And if you also start to include temperature and daily price – now it’s starting to look like something we can feed an AI algorithm.

Our challenge has not changed, but we have re-framed the data set to solve it and made it more suitable for AI.

But before transforming a data set the critical factor to consider is relevancy. For an ice cream maker breaking down sales from months to days and adding temperature is likely to increase the predictive powers of AI. On the other hand, a spare parts manufacturer with highly erratic sales patterns might not have a lot to gain doing so.

My aim in writing this article was to cut through the noise and hype about AI – to demystify AI and offer an overview of some practical applications for traditional businesses.

Hopefully, it has helped you gain some clarity about how useful AI may or may not be to your company at this stage. Feel free to reach out with any comments or questions!