When supply chain leaders embrace AI, just what problems are they trying to solve? Shana Wray, principal solutions architect and supply chain analyst with FourKites, explains the right way to go about automating workflows.

Companies are turning to agentic artificial intelligence for multiple reasons, Wray says. Some want to relieve employees of the burden of dealing with the data that creates “noise” in their day-to-day lives, or assign “low-impact, redundant” tasks such as email messaging to an automated system. They might carefully create specific agents for that purpose. Others would “just love to automate everything.”

The common goal is to eliminate the pain points that arise from dealing with suppliers, many of which lack true electronic data interchange capabilities for the transmission and confirmation of key documents such as bills of lading, packing slips and commercial invoices.

At the outset, the move to automate with agentic AI begins with the question of whether to build the model in-house, or buy it from an outside vendor. The “build your own” option might seem attractive to a company that believes it needs workflows and standard operating procedures that are highly specific to its organization. The downside, Wray says, is that such an effort typically requires extensive lead times, perhaps 18 to 24 months, during which period the company in question might be losing competitive advantage because it can’t move quickly enough.

Under the “buy” scenario, Wray says, companies gain access to pre-built large language models that are already trained in logistics terminology. The time involved in implementing that model could be just days or weeks. “It’s a quicker ROI,” she says.

Regardless of how a company sets forth to adopt agentic AI in its supply chain, humans will remain in the loop, Wray says. Experience shows that daily tasks can be handled by AI around 86% of the time, but for the remaining 14%, people are still needed.

“It’s important to set really good guardrails,” Wray says, “and escalation paths when the AI agent doesn’t understand what the task is.”