Council Post: What Happens When An AI Model Gets Punched In The Face?
Chris Burchett is Senior Vice President Generative AI at Blue Yonder, a leading AI company for end-to-end supply chain transformation.gettyAccording to ex-boxing champ Mike Tyson, everyone’s got a plan until t...
Chris Burchett is Senior Vice President Generative AI at Blue Yonder, a leading AI company for end-to-end supply chain transformation.
getty
According to ex-boxing champ Mike Tyson, everyone’s got a plan until they get punched in the face. That rings true for supply chain leaders who spend all day, every day getting punched in the face by the unexpected—late shipments, breakages, equipment outages, workforce issues and all the other factors that come with running a complex operation with countless moving parts.
That the supply chain functions at all, given all this face-punching, is a tribute to the systems we’ve built and the people who run them. But these systems also complicate efforts to automate supply chain operations. Mechanistic approaches—using “if-then” logic to break warehousing and other critical operations down into rigid rules—are doomed to fail. Things will always go wrong, forcing operators to be creative and improvise solutions not yet enshrined in any rulebook.
Specialists, Not Generalists
The first step is recognizing that you can’t automate the supply chain using generic LLMs. Frontier labs have made remarkable progress in areas like coding, math and textual analysis, but their LLMs can’t run warehouses because the vast datasets on which they’re trained don’t include information on how warehouse managers make decisions.
If you’re training an LLM to write a novel, you can feed it countless books to reverse-engineer the process of authorship. But supply chain data is closely held; you can’t scrape the web for the insights needed to run a warehouse. What’s more, it isn’t enough for LLMs to make good decisions: they must also select and engage the right tools—from Excel spreadsheets to warehouse and transportation management systems—to execute those decisions adaptively across complex, interconnected and endlessly varied facilities and operational processes.
In other words, we need to train LLMs on supply chain data so they can make smart decisions. And use reinforcement learning to iteratively refine models against real-world scenarios, turning decision-making power into operational excellence. Without both data and experience, LLMs will never meet the needs of supply chain organizations.
What’s required are supply chain savants: highly specialized LLMs that do one thing, such as running a warehouse, far more effectively than any generic LLM can manage. Such a model wouldn’t solve Olympiad-level math problems, recommend recipes or write novels. But it would understand warehouses, and its capabilities would generalize across a range of facilities and use cases, enabling flexible deployment with minimal customization.
Partnering For Success
Such LLMs, however, can be built only by organizations immersed in the supply chain, with access to operational data and experience with real-world supply chain challenges. So, supply chain organizations must themselves play a key role in building supply chain AI. But few warehouse operators have the resources to build AI models from scratch: the economics of AI, from GPU runtime to expensive talent, simply don’t make sense for most businesses. Further, most businesses’ IT departments are focused on day-to-day operations rather than innovation.
On the other hand, supply chain technology companies specialize in innovation, have ample tech expertise and experience, but lack the data. It’s a perfect partnership. Supply chain tech companies can aggregate anonymized data and whole-industry insights to build specialized models that can rapidly adapt to individual organizations' needs. Success depends on supply chain operators and trusted vendors partnering to develop models grounded in a deep understanding of both the industry and the specific processes that power each organization.
The resulting specialized models would be far leaner than generic LLMs, reducing training costs while enabling inference with far fewer tokens. Such models could initially be bootstrapped using existing frontier models as trainers, but reinforcement learning and other techniques would enable specialized supply chain models to quickly surpass anything Claude or ChatGPT can manage.
Crucially, such models could also be tailored to specific operational use cases—understanding the nuances of multi-customer facilities or cold-chain warehouses, for instance or mastering the specific distribution needs of individual customers. There would also be opportunities to layer in upstream and downstream data—such as signals from demand-forecasting tools or transportation management systems—to drive further optimization and enable models to anticipate and avoid punches before they land.
The Path Forward
Achieving real-world value with AI is tough for any enterprise: according to McKinsey, only 10% of large enterprises have successfully integrated AI into their operational workflows. But it’s especially challenging in the supply chain, where the environment constantly changes, operations are enormously complex and the cost of errors is sky-high. There are a thousand wrong ways to run a warehouse, so we need LLMs that can reliably identify the right solution and the right tools to get the job done.
Getting there will require activating the data flowing through supply chain organizations, and also capturing the hard-won knowledge siloed in the heads of human experts. There is no shortcut. Solving for the supply chain requires intimate knowledge of real-world processes—and that will have to come from supply chain operators themselves.
To bring transformative AI to the supply chain, we’ll need strong partnerships between industry leaders and trusted tech vendors. Specialized innovators who already understand AI and the supply chain will use aggregated data to build the technological scaffolding. Enterprise operators will then build on those foundations, partnering with vendors to develop customized LLMs that map onto and streamline their operations.
It’s a more challenging path than simply adopting generic LLMs, but also a far more rewarding one. Supply chain leaders who commit to this process and demand that their vendors do the same will be the first to build supply chain AI that really packs a punch.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?