Blue Yonder and WMG, University of Warwick recently released a report exploring the digital readiness of today’s retail supply chains with only 15% of global retailers reporting prescriptive or autonomous supply chains driven by artificial intelligence (AI) and machine learning (ML). However, over half of those surveyed are eager to embrace AI and ML, with 61% indicating plans to implement these technologies.

Investing in supply chain technology is a mandate for all retailers, but when shopping for solutions it’s important to remember that not all AI is created equal.

How can you differentiate between “good” AI and “bad” AI when evaluating what’s powering your organization’s supply chain? At Blue Yonder we believe AI should empower associates and organizations through collaboration with real-time insight then used to orchestrate actions across the ecosystem. In this blog, we’ll share a pragmatic five-stage process called IDEAS for evaluating whether an AI solution is “good” or “bad”, with retail forecasting as an example.

For AI to be considered “good” in today’s day and age, it must be:

Interconnected

AI models are built to mimic reality, but the closer the model can get to the complexity and interconnectedness of reality, the better that model is. Some AI solutions start with a base forecast and use machine learning to iteratively add on factors. However, AI solutions that make fewer baseline assumptions are stronger, as they consider how factors change and impact each other rather than expecting the same things to happen repeatedly.

When it comes to the retail supply chain, forecasting previously used a layered approach, beginning with a base forecast of weekly sales and adjusting that forecast based on factors like a promotion or an event. That base forecast would then be evaluated and altered in completely different silos based on each factor, and the results split by location and day of the week often using simplistic or static profiles. This resulted in interconnectivity between factors not being considered and the time horizon rather stretched. As we all know, human behavior isn’t so simple; it’s more like a network with a myriad of factors simultaneously impacting our decisions and the products we buy. So, would it not be better to have a model that mimics reality? Seemingly impossible a decade ago, Big Data, SaaS and AI have joined forces to make simulating reality…a reality. Good AI models look at all the factors that impact a situation concurrently, resulting in the most accurate insights available.

Dynamic

AI solutions must be dynamic as reality is rapidly changing in complex ways. To effectively deal with uncertainty spurred by the rapid velocity of change day to day, good AI can keep up with changes in real-time and adapt accordingly on its own. Along with responding quickly to changing influences, Good AI also understands that forecasts are not certain. If they were, we would all be rich, but humans are not 100% predictable. Good AI is not only able to limit the unpredictability but also understand it, creating a probabilistic view of the forecast based on the individual uncertainty of each location, item, and day. The probabilities will change as the inputs change, producing a distribution that is information-rich, revealing for example the risks of under or over-ordering, which can then be used to fundamentally improve inventory management processes. When implemented in the retail supply chain, this allows for a deeper level of understanding when it comes to choosing the right times for offering certain products and promotions, as well as the right times for scaling back.

Models like these probability density functions are rich sources of business information that enable sophisticated supply chain decision making by automatically producing distributions that show a range of probabilistic demand over time. These distributions are adjusted by the AI itself based on the changing reality of the interconnected network of factors that impact consumer behavior.

Explainable

One of the biggest obstacles for AI adoption in the real world is trust as people tend to wonder how they can put their trust in something they don’t fully comprehend. If it’s not understood how the machine works, then how can you possibly know what benefit it’s providing?

An example of bad AI is “Black Box” AI, a type of model that does not allow the user to conceptualize the reasoning – it’s essentially an impenetrable system that fails to offer human collaboration capabilities. To thoroughly evaluate and trust an AI solution, the model must follow a “Glass Box” AI composition, so that the machine’s thinking can be observed and understood. Blue Yonder’s AI interface graphically illustrates the impact of influencing factors such as promotions, social media, or weather on predicted outcomes. This accelerates personal and organizational learning, building trust while improving results. Our experience has shown that trust is one of the keys to the speed of AI adoption and success in the supply chain.

Automated

Probabilistic forecasts can be used to set the right levels of stock in every location to meet your business strategy, whether to maximize profit, reduce waste, manage distribution costs, or improve freshness. With hundreds of factors combining with tens of thousands of products over hundreds of locations, it is increasingly efficient to trust the machine to automate and reveal demand forecasts for every type of product at each store location. The depth of information provided by outcome automation solutions can add deeper layers of transparency and understanding that can be applied across the supply chain.

Scalable

AI solutions must be scalable so that they can be utilized widely and effectively. The “AI Chasm” is the void that sits between a demo and an actual, scalable, and self-learning solution that can be successfully deployed at a vastly larger level, interacting meaningfully with users. Blue Yonder’s embedded depth of experience means our process scales reliably and robustly. Blue Yonder supports over 7 billion machine learning transactions per month. When it comes to the supply chain, implementing a robust AI model that crosses the chasm not only helps to keep things running – informing decision making across the business from operations to store, logistics and warehouse operations – AI also delivers real-time optimizations, understanding the latest impacts to optimize tasks and actions, minimizing disruptions, delivering efficiency, profit, service, and competitive advantage. 

Together, the five qualities that make up a good AI offering – Interconnected, Dynamic, Explainable, Automated, and Scalable – serve to reveal the underlying reality of real-world dynamics to inform and automate business decision making with a precision and scale unattainable just a short time ago. This is a truly transformative outcome for the growth of both businesses and supply chain professionals.