Supply chain leaders have been using technology to design and optimize their supply chains for decades. These days, they’re trying to understand the best way to integrate artificial intelligence and large language models (LLMs) into their existing set of tools.

A typical network design cycle consists of the following steps:

• Define objectives
• Gather the data
• Model the network
• Analyze the output
• Summarize the recommendations

When business leaders think about modeling the supply chain, they often have a very good idea about the first step. However, steps two and three are typically the failure points.

Every company these days has plenty of data, but too often it’s incomplete or inconsistent. A freight history table could contain both actual costs and benchmarking data provided by a third party. An order history table could contain non-standardized product nomenclature or canceled orders. A production schedule table could contain trial runs on the production lines with outlier run times and waste percentages.

Feeding this kind of data into a traditional, non-AI supply chain design tool without doing expensive data cleaning beforehand will not result in a useful model. The baseline will be flawed, so creating an optimization will not be possible. That’s why supply chain network design typically relies on heavily curated data. But this leads to a different kind of problem: a “brittle” baseline, one where users can’t flexibly create scenarios around new or unknown conditions.

On the other hand, a network design tool with built-in AI capabilities can effectively process raw data, with all of its irregularities and oddities. Modelers can input many years’ worth of historical data such as shipments, production performance, customer ordering patterns, freight seasonality, commodity pricing and labor statistics. AI is best at recognizing patterns, and there’s a significant efficiency to be gained by going just that. A simple nomenclature consolidation (for example, St. Louis, StLouis and Saint-Louis) could significantly accelerate the model building process.

Now we have a pool of data that can be cleansed and normalized into a “pliable” baseline model. Rather than forcing the user to build and maintain rigid freight tables with thousands of rows, the model can quickly come up with freight approximations for any lane.

If we’re to take this one step further, the AI capabilities would be able to ask yes/no questions, adding a layer of supervision to the LLM and augmenting its capabilities. Some of those questions could be related to product cannibalization and substitution, change in suppliers or unprecedented events like factory shutdowns.

Finally, the AI can make recommendations about aggregating individual products for modeling purposes, based on the demand volume, product cost, sell price and production rate, thus eliminating human bias.

Building the model is only a stepping stone for supply chain leaders; they need to design a future optimized state that they can trust. They’re seeking answers to pragmatic, business-centric questions such as:

• How should my flows be optimized over the next three months?
• When do I need to consider expanding my distribution network?
• What would it take to get the CO­­­₂ emissions down by 5%?

Traditionally, a modeler would undertake a whole exercise to add data and build individual scenarios to answer those questions. A traditional scenario-building approach relies heavily on specific inputs by the user. So in order to answer seemingly simple questions, one could end up making hundreds of scenarios — for example, to determine the tipping point at which it becomes more cost-effective to change a raw materials supplier.

A proper application of AI techniques can completely turn this process around. The language of business can be interpreted by the network design engine to perform the following tasks:

• Automatically detect bottlenecks and simultaneously come up with potential solutions,
• Build in a sensitivity into every analysis, and
• Ask clarifying questions and interact with the user.

The integration of AI and large language models into the supply chain isn’t just an add-on; it's a revolution in the way we approach data processing, model building and scenario management. The key benefits — the ability to handle inconsistent data, recognize complex patterns, and translate business-centric queries into real insights — make AI an essential component of modern supply chain network design.

Marianna Vydrevich is manager of operations research and network optimization at GAF, and Steve Johanson is senior vice president and industry principal with Logility.