LLMs Are Everywhere in Logistics, So Let’s Talk About Doing It Right

Dear ChatGPT, please fix my shipping problems.

That joke isn’t far from reality. Shipping teams are already experimenting with AI to analyze shipping data and troubleshoot operational issues. Need help analyzing a delay? Ask ChatGPT. Trying to write a carrier escalation email? Ask ChatGPT. Want to figure out why your two-day shipment is suddenly five days late? You guessed it—ask ChatGPT.

By the time Manifest rolled around this year, it was obvious these tools had quietly made their way into everyday logistics work. Some companies have official policies. Many don’t. Either way, the usage is already happening.

So during the show, Lori Boyer, Head of Content at EasyPost, sat down with Tom Butt, Director of Platform Analytics, and Tyler Diestel, Senior Product Manager, to talk through what they’re actually seeing with LLMs in logistics.

The conversation moved through a few questions shipping teams are running into right now:

• Why LLM usage is climbing so fast even when companies haven’t fully approved it
• Why the way you ask a question has a huge impact on the answer you get
• And where teams need to slow down and think about security before pasting data into a public tool

Because whether companies are ready for it or not, LLMs are already part of the workflow.

The uncomfortable truth about “shadow AI”

One of the talking points that drew a lot of audience interest showed a gap that pretty much everyone in the room recognized immediately. In logistics today, only about 23% of companies have formally adopted LLMs, yet more than 90% of employees are already using them in some way at work.

That mismatch tells you a lot about where the industry is right now. The tools are super easy to use and really helpful, so people reach for them even when their company hasn’t quite figured out the rules yet.

Tyler summed up the hesitation from leadership in a single sentence: “A lot of companies are still trying to figure out what data can actually go into an LLM.”

That uncertainty is the real issue. Teams want the productivity boost these tools offer, but companies are still working through security policies, governance, and data boundaries. In the meantime, the day-to-day work of running shipping operations doesn’t slow down. When something goes wrong with a shipment, people look for the fastest way to understand what happened—and that’s often through an LLM.

The risk is that this kind of unofficial use can drift into bad habits. Someone pastes information into a public tool that should have stayed internal. Or a confident-looking AI response gets accepted without anyone taking a closer look. Neither of those problems comes from bad intentions. They happen because the tools are new and the guardrails are still being built.

What an LLM actually is (and why shipping teams should care)

LLMs are sometimes talked about as if they’re wise little robots that know the right answer to everything. That framing is misleading, and understanding how shipping AI actually works makes it much easier to use it well.

Tom explained it in straightforward terms during the session. A large language model predicts the most likely next word in a sequence based on patterns in massive training data. The result can sound incredibly knowledgeable, but it is still a prediction rather than a verified fact.

Tom also offered a distinction that resonated with a lot of people in the room: “APIs help us connect data systems… but LLMs help humans connect with our data.”

That shift is important. Companies have had dashboards and analytics platforms for years. The difference now is that someone can ask a question in plain language and receive an explanation that feels more like a conversation than a spreadsheet. For many teams, that makes data far more accessible.

At the same time, the way shipping AI works creates a few limitations. Tyler Diestel pointed out that LLMs are particularly strong at working with language. They can summarize messy information, organize scattered ideas, and help people draft documents or analyze patterns in large bodies of text. Those strengths make them useful for tasks like writing emails, reviewing reports, or making sense of operational notes.

Where they struggle is computation. Tyler explained it this way: “They’re not really good at doing computation… they’re trying to predict what the equation equals.” In other words, the model is generating the most likely answer rather than actually performing the calculation.

That distinction matters in logistics because much of the industry depends on precise numbers and clear cause-and-effect relationships. Delivery commitments, service levels, and routing decisions all rely on accurate math and real-world constraints. An LLM can help analyze information and suggest possibilities, but it should not replace the judgment of someone who understands the operation.

Tom captured the right mindset during the discussion: “An LLM can be a great copilot… always do the ‘trust but verify’ method.”

Used thoughtfully, these tools can help teams explore data and surface insights faster than before. The key is remembering what they are doing under the hood and treating their responses as a starting point for investigation rather than the final answer.

Prompt engineering that doesn’t waste your time

One of the biggest takeaways from the session was how much the quality of an LLM’s response depends on the way the question is asked.

When people first start using tools like ChatGPT or Claude, they often type a quick question and hope for the best. If the response feels generic or unhelpful, the conclusion is usually that the tool itself is overrated. In reality, the model is only working with the information it receives. The clearer the request, the more useful the answer tends to be.

During the session, Lori shared a simple framework she uses to structure prompts when working with LLMs. The approach is called ROCKS, and it gives the model a clearer understanding of the role it should play, the type of answer expected, and the context surrounding the problem. In practice, frameworks like this are becoming some of the most effective AI prompts for supply chain analytics.

The ROCKS framework

R – Role
Start by telling the model who it should act as. That might be a warehouse operations manager, a supply chain analyst, or a senior product manager reviewing a document. Assigning a role often leads to responses that are more specific and relevant.

O – Output
Describe how you want the answer delivered. Some situations call for a ranked list, others for a short explanation or a draft email.

C – Context
Provide the details that shape the problem. Shipping timelines, service levels, and operational constraints all influence the answer.

K – Know
Ask the model what additional information it would need to improve the analysis. This step often reveals missing pieces of the problem that might otherwise go unnoticed.

S – Show assumptions
Request that the model explain what assumptions it is making and how confident it is in the result. Seeing those assumptions helps determine whether the response is reasonable or needs further validation.

When those elements are included, the responses tend to become far more useful. The model has a clearer role, understands the situation better, and can explain its reasoning rather than simply generating a guess.

For more examples walking you through how to use LLMS in a variety of scenarios, download our free ebook here. 

Security rules for public LLMs

Security came up repeatedly during the session because this is where well-intentioned people can run into trouble without realizing it. As more teams experiment with AI tools for shipping analytics, understanding the AI security risks for shipping companies becomes critical. Lori shared a simple rule that tends to resonate with teams right away: 

If you wouldn’t put something on your company’s website, don’t paste it into a public LLM.

A slightly more practical version of the same idea works just as well. If you would hesitate to email the information, drop it into Slack, or put it in a shared document with your name attached, it probably does not belong in a public AI tool either.

Most of the time the distinction is fairly straightforward. Some types of information can be safely used to experiment with prompts and analysis, while others should never leave internal systems.

Safe to share (generally)

• Redacted samples with identifying details removed
• Aggregated trends such as counts by zone or average delivery times
• Historical summaries that are not tied to identifiable customers
• Public policy documents or SOPs, if company policy allows it
• Hypothetical or “made up” scenarios created for training purposes

Do not share in public LLMs

• Customer data, including PII such as names, addresses, or account details
• Contracts or confidential pricing agreements
• API keys, credentials, or authentication tokens
• Anything clearly marked or understood as internal-only information
• Unique identifiers that could be traced back to a specific shipment or person

Tyler highlighted API keys in particular, partly because people really do paste them into prompts more often than anyone would like. As he put it during the session, “You definitely don’t want to have API keys being put up on there.”

Some organizations are starting to deploy internal LLM environments where those risks are easier to control. If your company has one of those systems, that can open the door to more advanced use cases. If not, it helps to assume that anything placed into a public tool should be treated as public information.

Why shipping-specific LLMs exist at all

At this point in the conversation, someone usually asks a very reasonable question: if tools like ChatGPT already exist, why would a logistics company need something more specialized?

During the session, two themes kept surfacing. The first was security, which we have already discussed. The second was context. General-purpose LLMs are trained on a broad mix of information. They can explain concepts and generate text, but they don’t understand the operational details that shape shipping decisions. They don’t know a company’s carrier mix, its service level commitments, the performance of specific lanes, or the tradeoffs that teams manage every day. And even if they could help analyze that data, most organizations wouldn’t be comfortable pasting sensitive operational information into a public system.

That gap is why companies are starting to experiment with shipping-specific LLM environments. One example is Luma AI Advisor, a shipping-focused language model designed to operate inside a secure environment and work directly with shipping data. Tools like this are emerging as a safer way of using AI with shipping analytics inside company-controlled environments.

Tom described the foundation of the approach during the session: “We put that into a large language model… within a safe and secure environment.”

The idea grew out of conversations with customers. Over the past several years, EasyPost has built increasingly sophisticated analytics and dashboards to help teams understand their shipping performance. Those tools answered a lot of questions, but they also created a new one. As Tyler explained, customers often came back with a simple follow-up: What should I do with this information?

Luma AI Advisor sits on top of that analytics layer and allows teams to interact with their data through conversation. Instead of digging through dashboards, someone can ask a question and explore the operational implications directly.

What kinds of questions does a shipping team ask?

Tyler gave one that comes up constantly: “What would happen if I added a new carrier to my carrier mix, and how would that affect my shipping costs?”

That is a perfect example of why context matters. The answer depends on your volume, lanes, zones, SLAs, service levels, and how you define success. A generic LLM can give you a generic answer. A shipping-aware system can point you to the metrics that actually move the needle for your operation.

Why tool calls matter

We also talked about a nerdy but important point: computation. Tom explained that part of making a shipping-specific LLM useful is engineering “tool calls,” meaning if the model needs to do math, it can call the right tool instead of guessing. That helps close a real gap in general-purpose LLM behavior.

In logistics, “close enough” math is how you accidentally blow up a promise date.

How to start using LLMs safely this week

If you do nothing else, do these three things:

• Adopt ROCKS prompts for any operational question you care about
• Put a simple security rule in writing and share it with your team
• Normalize “trust but verify” as part of the culture, not a personal quirk

And if you’re at the point where your teams are already using LLMs daily, but security and accuracy are becoming real concerns, that’s when it’s worth looking at shipping-specific approaches like Luma AI Advisor.

General-purpose LLMs still have their place, but shipping decisions carry real consequences. When operational data is involved, many teams prefer tools designed to work directly with that data instead of pasting it into a public prompt.

If you want to see what a shipping-aware LLM looks like in practice, get in touch with our team, and we’ll show you how Luma AI Advisor works with shipping analytics in a secure environment.