GenAI is ready for more autonomy, but supply chains are not
Disruptions, vulnerabilities, late deliveries and customer interactions are among the supply chain concerns that GenAI could address autonomously -- if businesses can trust it.
The chief procurement officer at a specialty chemicals manufacturer didn't trust AI to do things right two years ago. But that has changed. Hexion's Gael De Martelaere now says if businesses wait until AI is fully vetted "somebody else is way ahead of you."
From heavy manufacturing to healthcare logistics to a mattress factory, businesses are deploying AI systems and seeing real gains in specific use cases, while learning the limits of AI deployments.
De Martelaere's measure of success isn't based on how many AI tools are deployed. "I'd rather implement two tools that get full adoption than implement 20 tools and have no adoption at all," he said.
Hexion plans to run six AI procurement use cases over the next 18 months, one built in-house and five purchased from vendors. Among them is an in-house agentic system that extracts insights from news and market subscriptions and converts them into faster procurement decisions.
The real promise of AI in logistics isn't just doing the same work faster, but spotting trouble before it strikes, said Shane Curtis, CTO at defense intelligence company Artorias. For businesses that constantly scan public information, such as social media, traffic data and shipping reports, in addition to their own data, AI can flag vulnerabilities in the supply chain early and shorten the time between a warning sign and a company's response. "It's a manpower multiplier more than anything," Curtis said.
Chris Burchett, senior vice president at supply chain software provider Blue Yonder concurred. "You need to understand the threat surface," he said, "and you need to know what the common vulnerabilities are and protect against those." A Blue Yonder team is focused on published vulnerability frameworks for large language models, including red teaming its own AI agents to test for exploits in the supply chain.
Narrowing AI's capabilities increases reliability
Josh Medow, CEO of Mercury, a healthcare and life sciences shipping company that transports temperature-controlled medical devices, diagnostic kits and biological samples, said he built reliability into his AI tools by narrowing what they're allowed to do. "We've really limited it," he said, resulting in greater reliability.
Mercury's internal chatbot helps logistics coordinators answer questions. That reliability, Medow said, came only after he disconnected the system from the company's broader data and rebuilt the training set from scratch using verified current documents. The first version of the system failed because it pulled data from outdated files that contradicted newer ones.
Medow doesn't plan to use AI autonomously in situations involving too many parties and complex judgments. "I can't see that in the near future," he said. Correcting a flight delay in a shipment bound from Boston to London, for example, might include finding alternative carriers, weighing options for the customer, updating the customs broker and notifying the recipient.
GenAI can autonomously reduce costs, improve customer service and speed decision-making along the supply chain.
Late deliveries are also a supply chain concern that's slowing adoption of AI in areas of operation requiring complex decision-making. "If we get it wrong [using AI], we're not shipping, we're not stocking the shelves, we're not getting inventory to our customers," said Brad Forester, CEO of JBF Consulting, which advises manufacturers and logistics companies on AI-enabled supply chain strategies. Missed shipment agreements can result in penalties. "Those come with top line, bottom line impacts to the P&L [profit and loss statement]," Forester added.
Jack Dell'Accio didn't wait for a vendor to solve his procurement problem. The founder and CEO of manufacturer Essentia Organic Mattress built his own AI tools and integrated them into his open source Odoo ERP system. Every sale, purchase order and shift in freight costs is fed into Dell'Accio's procurement projections in real time. "The projection that was there yesterday may be different today because of the results of the day," he said. "It's never been so real time for us."
AI's real-time updating of information has enabled Dell'Accio to stop overbuying raw materials, freeing up cash from his old spreadsheet-driven approach. "For cashflow purposes, the use has been fantastic," he said. "Like a new employee, eventually trust [in AI] is built."
When stakes are high, trust is low
Businesses have long hesitated to use AI autonomously in their supply chains due to AI's inability to adapt to changing situations, Burchett said. Supply chain platforms were historically designed to support human decision-making rather than making choices autonomously. Highly structured workflows, such as optimizing how trucks are loaded and routed, can be automated reliably, but AI agents are limited when conditions deviate from what the agent was programmed to do. "It's still early days for those reasoning algorithms," Burchett said.
The technology right now is probably more mature than the customer trust to accept it.
David VallejoVice president of digital supply chain, SAP
But even in situations where AI is capable of executing autonomously on high-stakes supply chain decisions, businesses aren't ready for it, acknowledged David Vallejo, vice president and global head of digital supply chain at SAP. "The technology right now is probably more mature than the customer trust to accept it," he said.
Businesses still want a human in the loop when supply chain decisions involve multiple companies and affect customer commitments. But human involvement is more about the approval process than hands-on intervention, Vallejo explained. For many lower-impact decisions, however, businesses can let the autonomous AI system run "lights out" without human involvement, he said.
De Martelaere pointed to what could become an industrywide challenge: As AI tools clean and categorize data, the data feeding AI decisions is increasingly generated by AI itself. "How do you trust that the output is good when more and more is AI-generated?" he conjectured. "I don't have the answer to that."
The path toward greater AI autonomy depends on transparency, Vallejo said. AI must explain its reasoning and cite its sources. Without that, pilots fail and trust erodes. And "even if you have trust," he noted, "it's trust but verify."
Eventually, more powerful computing, including quantum systems, will overcome the resistance to using autonomous AI in supply chains, Curtis predicted. Quantum AI will be able to test out millions of "what if" scenarios at once, he explained, and show how a disruption could ripple through supply chains and which responses would best keep goods moving.
But De Martelaere is skeptical that ever-more-powerful computing like quantum is the answer. For a function like protein folding, "you probably need that computing power," he said. "If you do business -- business is not that complex." The roadblock to acceptance isn't AI, he argued, "It's the people and how we feel comfortable adopting and adapting to it."
Patrick Thibodeau has worked for several decades as an enterprise reporter, focusing on IT and workforce management, ERP, high-skills immigration, tech policy, and high-performance computing.
