About a year ago, Amcor started to experiment with EazyML, a platform that helps optimize the forecast for both customer demand and the supply side. They trained the tool using three years of data from ERP to look for patterns in fluctuations. The system tries to find categories of change and which events correlate with different kinds of change. For example, it checks for seasonal fluctuations, and whether two or more types of change occur together, or if they’re mutually exclusive.

"The early results we have are quite promising; much more than we expected," says Ranchin. "If you can predict change, you can better anticipate your raw material needs and supplement them ahead of time if necessary."

This is no surprise to Bapat, who says forecasting is an area AI has significantly improved. "In the past, many organizations relied on consensus forecasting, where weighted input from different experts was used to come up with an average prediction," he says. "Studies have shown that stat forecasting, where statistical techniques are used to extrapolate from historic data, consistently outperforms consensus methods. And machine intelligence does even better than stat forecasting. But the trick is to make sure you use the right data."

Inspection

Another example of how AI is being used can be found at Intel, where several chips are printed on a single wafer using lithography. Those nearest the center of the wafer tend to have the best power performance profile. Those near the outer ring, while still reliable, tend to have reduced performance. Intel has a quality threshold against which chips are measured to determine whether they should be kept or thrown out. Having a human inspect wafers would be a time consuming and fault ridden process.

"We use AI to select the right high-quality chips, and that makes us much faster at producing chips and getting them to the market with better quality," says Greg Lavender, SVP and CTO of Intel. "Of course, that's not the only thing we do with artificial intelligence. I have a couple hundred AI software engineers who report into my organization. Some of what they do is used in our fabs for inspection and testing, but sometimes they develop AI that’s delivered inside our products, without anybody necessarily knowing about it."

A case in point is how Intel helps their OEM customers by providing software tools that test for malware. One such tool is the Intel Threat Detection Technology that runs on Intel laptops. When code is executed in Windows, the Intel code examines the instruction stream in the CPU. Using adaptive learning signature algorithms, it looks for anomalies in the code that match a malware signature. If a match is found, the tool intercepts or blocks the malware and alerts Windows Defender to an infection on the device.

"The Threat Detection Technology is built into all of our client CPUs," says Lavender. "Those infections sneak in through the supply chain—and by the time the end product is put together, the only way to find them is with this tool. We've been delivering this and other AI tools for the last few years, but now with all the talk about large language models, more people are talking about it."

According to Cisco’s Bapat, inspection is a huge part of supply chain management, and it becomes a lot easier if the right steps are taken during product design. "You can save a lot of cost if, during product design, you embed instrumentation into the equipment that can generate data to help monitor the flow," he says. "If you take the bill of materials for any product and look at the labor burden costs, you'll see they’re very high. Burden is basically product quality and supervisory overhead. AI is already helping to minimize that cost today."

Optimization

Forecasting and inspection are both important, but the biggest impact will come when supply chains can be tailored to specific customer needs. Bapat draws from an important lesson he learned when he designed one of his best AI algorithms. It took nine months to develop and deploy—and in the end, it still took a surprisingly long time to make it work. Thinking back on what went wrong, he realized that no matter how good the technology was, it wouldn't produce the desired outcome if he didn't first take the time to understand who the end customer was and how they planned to use the application. He also noted that while they generally have the loudest voice, senior management is not the end customer.

"Since then, I’ve made it a point to always start out with a good understanding of the underlying business, whether it involves sales or supply chain management," he says. "Once I have a solid understanding of the requirement, I work my way back into data and AI."

Bapat thinks this philosophy should be applied to supply chain management: "If you truly take a look at the end consumer, AI can help by segmenting and targeting consumers and their environment. And then when you work your way back through the supply chain, look at different costs: labor, production, tax, inventory, and optimize them together."

Once the supply chain is optimized for flow, he adds, you can then start installing and executing on predictive quality and maintenance. From there, you can work your way back into procurement for supply management.

“This supports the notion that suppliers are partners, not adversaries," he says.

So here lies the age-old challenge that supply chains are by their very nature composed of separate companies with at least three reasons not to share data. First, they may have a line of business that competes with one or more of the other partners. Second, they might be part of one or more competing supply chains. And third, they keep information to themselves to strengthen their hand at the negotiating table.

The current generation of AI can optimize supply chains—and even tailor them to deliver the right product to the right customer at the right price. However, doing so would require a level of data sharing that very few companies are ready for.

"What's missing are techniques that allow organizations to share some part of their data in full confidence that they haven't given away too much," says Bapat. "We're still five or 10 years away from that."