It’s no secret that some enterprises struggle to generate value from their AI investments. Now, a growing number are betting on developing AI agents capable of acting independently and rapidly turning predictive insights into real-world actions.
Take my own industry, supply chain. According to Gartner, just 7% of supply chain teams currently make decisions in real time—and that means less agility, less resilience, and more disruption when things go wrong. Agents could enable enterprises to respond faster (and smarter!) to challenges, delivering benefits across the entire global economy.
Building an agentic world is not easy. In June 2025, Anthropic gave its flagship AI model an identity crisis by asking it to run a vending machine. The troubled bot gave huge discounts, sold products at a loss, and invented conversations with fictional suppliers. Clearly, handing current models the keys to real-world supply networks—or any other critical business function—would be asking for trouble.
However, it’s still important to plan for an agentic future. These technologies are coming soon, and the time to develop best practices is now, before AI models start making decisions for our organizations.
4 BEST PRACTICES FOR AGENTIC AI
The supply chain is a key space in which to consider these challenges, because every supply chain problem—from routing fleets of trucks to creating retail planograms—is ultimately a data problem. Supply networks are full of the kinds of problems at which AI excels—and full of scenarios, from warehouse management to inventory optimization, where agents could do some of the work.
To succeed, however, agents need to work effectively at every step of what my company calls the SADA Loop. That means agents need to see, by monitoring the environment and identifying changes; to analyze, using pattern recognition and other tools to make sense of new information; to decide, based on a deep understanding of your business, what steps should be taken; and only then, act.
That might sound obvious, but most conversations around agentic AI rush straight to action and overlook the other (arguably more important) steps in the SADA loop. It’s important to pause, and ensure you’re building smart, safe, and effective solutions to the problems you’re trying to solve.
Keep these four key best practices in mind:
Focus on the right decisions
Gartner estimates that AI agents will automate or augment 50% of business decisions for decision intelligence by 2027. The bigger question, though, is which decisions those will be. Should AI agents be telling warehouse workers where to park their forklift, or issuing million-unit production orders, or doing something else entirely? To figure that out, organizations should start small, with low-stakes experiments, then iterate, scale, and replicate their successes.
That might mean initially using agents not to act, but to explain actions. The supply chain can be a black box; it’s often hard to understand why a particular process failed or a specific strategy worked. Embedding agents into supply chain software, and letting them surface problems and articulate solutions, will build trust in their capabilities and lay the groundwork for a future where they’re able to act independently.
Put guardrails in place
We need guardrails. I always advise people to view AI agents as enthusiastic but inexperienced interns; they mean well, but you should require them to ask permission before acting. Once a model shows it’s capable of making good decisions, you might give it more leeway to act independently—up to a $10,000 threshold, say—before gradually lengthening its leash.
Over time, we’ll see more technologically advanced guardrails, with agents monitoring agents and notifying humans if they spot anything untoward. For now, enforcing clear limits and keeping humans in the loop is the key to managing risk exposure.
Manage the technological burden
Everyone frets about AI going haywire, but don’t overlook the more quotidian concerns. Agents can strain data and computational resources, so be sure to put technical guardrails in place. When human users are starting their day or your planning solver is running, you might want your agent to cool its heels and save resources.
Remember, too, that agentic AI is just one tool among many. Deterministic solutions are cheaper to build and run—so if that’s all you need to solve a problem, there’s no point getting agents involved. Using the correct tool for the job is always important, especially as enterprises look to right-size their cloud and AI budgets.
Demand domain expertise
Agentic AI is only as good as the data it’s trained on—and even then, it needs to be developed by people who know your industry. (Anthropic’s vending machine bot was a reminder that general-purpose AI isn’t well-suited to complex tasks.) Agents must deeply understand not just the theoretical dynamics of supply networks, but also the nuances and priorities of the business they’re serving.
Such insights can be baked into agents via methods such as context engineering. That’s both an art and a science, and it requires AI innovators to really know the sectors they’re building for. Specialization is the only viable path forward; only operators who intimately know their space will be able to develop effective agentic solutions.
THE POWER OF AGENTIC AI
As agentic technologies mature, we can truly reimagine our operations—breaking down silos, unlocking faster and more accurate decision making, and making enterprises more resilient, productive, and profitable.
To win in the era of agentic AI, however, enterprises will need a well-articulated strategy anchored in clear KPIs. The idea that there’s a “right amount” of agentic AI is a distraction; the key questions are what business metrics your agents will impact, and how you’ll use agentic tools to drive real value for your business.
Above all, organizations will need to combine optimism about the transformative potential of agentic AI with a clear-eyed perspective on the challenges involved. Instead of rushing AI agents into service, organizations should incrementally build trust in these powerful new technologies and unlock sustainable paths to value.
Chris Burchett is SVP of generative AI at Blue Yonder.