AI Agents Fail at a High Rate—Here’s What Logistics Leaders Can Do About It

How are logistics leaders supposed to think about the promise of this still-developing technology as they plan their software deployments and work to improve efficiency?

What Agentic AI Really Is

When you implement a chat workflow or a driver empowerment tool that’s powered by AI, the baseline technology that simply generates briefs and responds to customer questions is generally built atop large language models (LLMs). These are the “classic” ChatGPT and Gemini use cases.

But as we start to ask AI to do more complex jobs in logistics, we’re increasingly turning to AI “agents” that can use tools to get a task done within specified boundaries. This is the difference between an LLM rating your carriers versus actually booking them for you.

In logistics, AI agents hold the potential of rescheduling orders for customers based on their conversations, proactively reaching out to suppliers when inventory levels are running low and even scheduling and routing transfers or deliveries. 

Some of these use cases are closer on the horizon than others, but the question for all of them is the same: What can logistics businesses do to ensure a low enough failure rate so these agents are actually useful?

How To Improve The Success Rate Of AI Agents In Logistics

The success rate of an AI agent isn’t just a matter of luck, and it’s not just a matter of incremental advances in AI technology, either. When it comes to logistics, it’s about three things:

1. Properly Scoping The Agent And Its Boundaries

Let’s say you have an AI agent that’s been deployed for rescheduling last-mile delivery exceptions. You could go about scoping that in a few different ways. On the farthest end of the spectrum, your agent has access to most of your last-mile tools and handles the process from end to end, including updating the route on its own.

A more limited scope might have a boundary condition around actually adjusting the route schedule while still enabling the agent to send a scheduling link to the customer to let them reschedule themselves based on your existing capacity. This is a similar experience from the user’s perspective, but it reduces the error risk significantly.

Businesses can use this type of guardrail to build toward more autonomous deployments as agentic AI becomes more broadly reliable.

2. Bringing A Human Into The Loop In The Right Place

There’s a huge difference between an AI agent that can make updates to a delivery schedule and an AI agent that can only make hypothetical changes that have to be approved by a human. The latter is much lower risk since potential failures won’t make their way into production.

Obviously, there’s a balance to strike here. If the human has to spend all their time reviewing the agent’s activity, you start to lose out on the efficiency gains that come from automating these tasks in the first place. Luckily, agents can learn over time, which means that human-in-the-loop processes will produce feedback that improves performance over time.

This also helps to put you in a position where you’re not totally reliant on your vendor and can monitor and interrupt AI processes if needed.

3. Leveraging Deep Logistics And Supply Chain Knowledge

All of this talk of scoping and human feedback loops brings us to perhaps the most important takeaway for logistics leaders thinking about agentic AI: AI in logistics is at least as much a logistics problem as an AI problem. Creating agents that actually work requires an in-depth knowledge of logistics and domain-specific knowledge about the nitty-gritty of thorny industry challenges.

For example, one of the ways that AI technology is improving agent performance is by using “agent swarms” that are actually comprised of multiple smaller-scale, more bounded agents that break down tasks into smaller components. This helps boost the reliability of agentic AI, but it puts more pressure on the architecture—the vendor developing the agents needs to have an extremely sophisticated model of how logistics processes actually happen.

This is an area where logistics leaders need to leverage their own domain knowledge to make sure their AI vendors can speak their language. Agentic AI is complex, but it shouldn’t be a black box. Those developing the technology should be able to provide an explanation rooted in supply chain logic for why the agent actually works.

Risks And Rewards For Logistics Leaders As Technology Evolves

When you evaluate a vendor of AI technology, you should really be evaluating them as a vendor of logistics technology. Keeping that in mind can be key to lowering your risk as you introduce AI-powered capabilities into your last-mile logistics or any other segment of your supply chain.

Avoiding agentic AI entirely because of the high failure rate is potentially shortsighted—but right now, the risk of failure is real, and decision-makers need to take steps to mitigate it. Agentic AI for areas like dispatching and supply chain analytics is likely to become commonplace in the near future, so it's critical to take steps to reduce that failure rate and maximize the power of this emerging technology.