Agentic AI for Freight Brokers Part 2: Inside the Operating System

If you run a brokerage, you’ve probably sat through enough AI demos to notice the pattern. Nearly every product now calls itself agentic, and every carefully controlled demo makes the software look like it can run half the floor by itself. 

Dig into how it works, though, and plenty of those “agents” turn out to be chatbots dressed up for the sales meeting.

That doesn’t mean agentic AI for freight brokers is another empty software trend, though. Gartner expects spending in the category to approach $53 billion by 2030. But it also estimates that only about 130 of the thousands of vendors using the agentic label have built something that genuinely qualifies. 

The market is real. So is the amount of noise around it.

A label won’t tell you whether a product can execute freight work. You have to look at what’s underneath it.

In Part 1, I explained why brokerages need an execution layer rather than another screen. Now I want to open that layer up. A real freight agent depends on three specific capabilities working together. Take away any one of them, and you don’t have an operating system. You have a polished demo that still leaves your people doing the work.

First Things First: Most of It Is a Chatbot in a Costume

Before we open up the operating system, we need to clear away the label doing most of the damage: agent washing.

The formula is simple. Take a chatbot or a script built to handle one narrow task, call it agentic, and price it like it can run an operation. I saw plenty of that at FreightWaves F3. The demos looked smooth because nothing unexpected happened. Freight rarely gives you that luxury.

Gartner predicts more than 40% of agentic AI projects to be abandoned by the end of 2027, largely because the costs outrun the value, and the controls were never strong enough to begin with. That prediction makes sense when so many products are being sold as something they aren’t.

The dividing line is ownership. A chatbot completes a prompt. An automation runs a step. A real agent owns the outcome, adjusts when conditions change, and keeps working until the job is done.

The Three Layers

Having said that, owning an outcome takes far more than a model that can answer questions or trigger a workflow. The agent needs a way to understand freight, carry out the work, remember what it learns, and show your team exactly what happened along the way.

That operating system rests on three layers. Each solves a different problem, but none works well on its own.

Layer One: TOAS, the Part That Shows Its Work

The first layer is the Transportation Observability Action System, or TOAS. It gives your team a live view of what the agent did, why it made each decision, and what happened next. A rep can follow the work as it unfolds and step in before a missed pickup becomes a rolled load.

Dashboards have displayed logistics data for years. The difference is that TOAS connects visibility directly to execution. FedEx found that 97% of logistics leaders no longer believe visibility alone is enough, and I agree. Knowing a load is in trouble helps. Seeing what the agent has already done about it, what it plans to do next, and where it needs human judgment is far more useful.

TOAS also keeps a running record of every action and decision. That audit trail is what allows a brokerage to hand live freight to an agent without surrendering control.

Layer Two: The LMA, the Agent That Owns the Load

TOAS lets you inspect the work. The Logistics Management Agent, or LMA, is the part doing it.

An LMA owns a freight outcome from beginning to end rather than completing one isolated task. It can source capacity, run outreach, negotiate inside the brokerage’s limits, respond to changes, and keep moving until the load is covered.

That requires more than general intelligence because freight has its own language, operating rules, and consequences. A general purpose model may recognize terms such as TONU, detention, or drop and hook, but recognition isn’t the same as understanding how they affect the next decision.

One bad response during a reefer breakdown can cost real money and destroy a rep’s trust on the first load. Freight fluency has to sit inside the agent from the start. Bolt it on afterward, and you’re still relying on a chatbot that learned enough vocabulary to sound convincing.

Layer Three: The Carrier Context Graph, the Memory Nobody Can Copy

TOAS shows you what the agent is doing. The LMA does the work. Envoy’s Carrier Context Graph makes sure it doesn’t forget everything the moment the load is covered.

Every carrier call, quoted rate, counteroffer, rejection, booking, service failure, and clean delivery gives the agent more context for the next load. It starts learning who answers on certain lanes, where a carrier will usually land on price, who performs when the market tightens, and which small warning signs tend to become big problems.

That history has to live somewhere smarter than a giant prompt. Shoving every past interaction into every request is slow, expensive, and mostly noise. Mem0 found that pulling only the relevant memories used about 7,000 tokens per call, compared with 26,000 when everything was packed into the prompt.

Most AI tools wake up with amnesia. The Carrier Context Graph doesn’t. After a year, your agent knows your carrier network in a way no competitor can buy off the shelf.

The Model Is the Easy Part

After walking through those three layers, you may have noticed what barely came up: the model itself.

That’s intentional. Most AI vendors are renting from the same small group of model providers. Nobody has a secret brain in the basement. The model can reason, but it doesn’t know your carriers, your rate strategy, your customer promises, or when a rep needs to take the wheel. All of that comes from the system built around it.

A real agent needs four pieces working together: a model to think, tools to act, memory to carry the right history forward, and policies that keep it inside your rules. Max pay, approval limits, rate floors, escalation points. That’s where the operation lives.

The tricky part is deciding what the agent should know at each moment and what it should ignore. People call that context engineering. I call it the difference between an impressive demo and software you’d trust with a live load.

The Next Problem Sits in the Corner Office

Once you can tell a real agent from a dressed-up chatbot, the buying decision gets easier. The harder part starts after you put one to work.

A machine that can source, negotiate, escalate, and finish the job changes who is responsible for what inside the brokerage. Someone has to decide how much authority it gets, which calls still belong to a person, what proof earns more trust, and who owns the answer when the software gets it wrong. No vendor can make those calls for you.

That’s where the next advantage will come from. Brokerages that learn to manage autonomous work before the market turns will be able to add volume without rebuilding the floor every time. The rest will buy the same technology and wonder why nothing really changed.

Ellie is already running real freight inside that model and keeps getting sharper on your lanes the longer she runs. Part 3 is about what has to change above her: how freight executives evaluate vendors, govern machine labor, keep people accountable, and become the boss of machines.