Skip to content
Fast Company Executive Board

The Fast Company Executive Board is a private, fee-based network of influential leaders, experts, executives, and entrepreneurs who share their insights with our audience.

Would you trust an AI agent to run your supply chain?

We must develop an understanding of how AI agents should work across the supply chain, and use that to promote the best practices needed to safely accelerate progress.

[Stanislav Komogorov / Adobe Stock]

Profile photo of Chris Burchett

Senior Vice President, AI Transformation · Blue Yonder

You shouldn’t be too quick to put an AI agent in charge of your supply chain.  

That’s the takeaway from a recent study in which researchers tasked top AI models with running a vending machine. While the agents were generally able to monitor inventory, place orders, and restock the machine, every single model also experienced unrecoverable meltdowns. One bot began firing off emails threatening suppliers with “total legal destruction”; another simply pulled the plug on the vending machine. 

Other reports also suggest that agentic AI isn’t yet ready for prime time. A team at Apple found that advanced LLMs suffered a “complete accuracy collapse” when presented with challenging logistical puzzles. More prosaically, ChatGPT’s agentic AI couldn’t find the price of eggs. 

The immaturity of agentic AI is a problem, because supply chain pros urgently need AI systems capable of making consequential decisions with minimal oversight. Machine learning and LLMs are powerful, but focus on manipulating data. To transform supply chains, we’ll need AI tools to act in the world and move physical things from A to B without human involvement.  

We won’t get there by deploying agentic tools before they’re ready. Instead, we need to develop a clear understanding of how AI agents should work across the supply chain, and use that understanding to promote the best practices and technologies needed to safely accelerate our progress.

INTRODUCING THE SADA LOOP

At my organization, we evaluate AI according to its ability to support the SADA loop—that’s See, Analyze, Decide, and Act. Agents need to see, by collecting data from across our organizations (and beyond), then analyze, by figuring out what’s salient and what it signifies. We need them to decide, in transparent and dependable ways that are verifiably grounded in data and analysis. Then, and only then, do we need AI agents to step up and translate approved decisions into action. 

Most of the AI tools in the market check someof those boxes. But to the extent that AI models reason independently, they do so like a hyperintelligent 5-year-old: with poor judgment and a dangerous degree of misplaced confidence in their own abilities.

The SADA loop reveals these gaps and shows us where more work is needed. It’s by doing the hard work of addressing those gaps that we’ll ultimately bring safe, effective agentic AI tools into the supply chain.  

Fortunately, the supply chain is the ideal place to build out these missing links: a highly quantified and data-driven space, full of analytical challenges for which AI is well-suited. It’s also a labor-intensive space, where agentic AI would drive real value. If agentic AI is going to succeed anywhere, in other words, it will succeed in the supply chain.

FIVE RULES FOR AGENTIC AI

To reach that goal safely and effectively, however, we’ll need to follow a few key principles: 

1. Know and work within AI’s limitations.

AI technologies are generally only good at a few classes of problems. An LLM can summarize text or write code incredibly well. It can even reason through small problems—but it can’t reliably reason through a complex series of problems. To build effective AI tools, we need to assign the righttasks—and provide the right context, via well-curated data and carefully tuned prompts—to set our models up for success. 

2. Define tasks narrowly.

Instead of building a single all-encompassing AI agent—which inevitably becomes a single point of failure—it’s better to create constellations of AI tools, each tasked with specific subtasks. By carefully managing agent flow, memory, and context, a process called context engineering, it’s possible to knit models together to solve big, complex problems without relying on a single AI agent working perfectly 100% of the time. 

3. Create clear guardrails.

To prevent AI agents from coloring outside the lines, we need to specify exactly where those lines are. That means setting thresholds for permissible behavior: an agent might be allowed to place orders up to a limit of 100 units or $1,000 in total spending. It’s possible to enforce such rules, but only if we define them clearly in advance. 

4. Trust but verify.

You should never ask AI for answers you can’t find any other way—and that goes double for agentic AI. If an agent is assessing and optimizing my inventory levels, for instance, I might automatically run a separate mathematical formula to benchmark the model’s output. If the AI’s output is off by a designated threshold, an alert should occur, and an automated fail-safe should kick in before any real-world actions are taken. 

5. Customize your AI.

You can’t buy a generic, off-the-shelf AI agent and set it loose on your supply chain. Instead, work with solution providers that understand your industry, articulate your needs and processes, break down data silos, and put appropriate safety measures in place. Customization is the key to safe and effective AI use—especially when you’re working with agentic technologies.

CONNECTING THE DOTS WITH AI AGENTS

These rules are natural extensions of the best practices for non-agentic AI applications, and we’re already finding powerful use cases for such technologies across supply chains.  

The dashboards on which supply chain pros rely, for instance, are perfect places to deploy generative AI to streamline data interpretation and summarize findings. Machine learning, meanwhile, is already solving complex logistical challenges—like optimizing inventory across multiple locations—far faster than any human.  

In the vast majority of cases, it’s still humans making decisions and putting plans into action. However, companies are beginning to cautiously deploy AI agents that make and execute decisions within narrow guardrails. That is set to accelerate.

By validating increasingly sophisticated AI models across increasingly complex use cases, we’re learning to close the SADA loop in effective and dependable ways. If we’re ambitious and thoughtful about integrating new AI technologies into our operations, then it’s just a matter of time before we’ll join the dots between these emerging applications—and usher in a new era of robust and effective AI agents that really canbe trusted to drive value across your supply chain. 

Chris Burchett, SVP Generative AI at Blue Yonder, is a  technology leader in product and corporate strategy, technology, and security.