The Next Competitive Moat Isn’t Better AI. It’s Better Operational Context.

AI has given freight brokerages a very sophisticated new way to procrastinate, and I say that as a guy who sells AI to freight brokerages.

I sit in these meetings. Everyone is “evaluating.” Of course they are. There’s a committee, six vendors, 47 rows on the scorecard, and a board presentation that’s been “next month” since February. Ask what will be live by the end of the quarter, and suddenly everybody needs to circle back.

And I get it. Nobody in that room has ever been fired for waiting another six months, and some of you got burned by a vendor who shipped a demo and called it a deployment. But that’s what a committee is for: somewhere to put a decision nobody wants their name on.

What gets me is that the scorecard isn’t measuring anything real. All six vendors rent the same two or three frontier models, and your youngest rep has one open on his phone right now for free.

Those models have read more about freight than anyone alive, and not one has been punched in the mouth by a Monday. I’ve had my teeth rattled plenty, and so has your floor, and that’s where the useful insights live. Like which carrier you’d hand a produce load on a Friday, or the one MC you won’t rebook and couldn’t explain why if I asked.

That’s the operational context, and it’s why we built Ellie at Envoy as an execution layer that can use those live rules, exceptions, and preferences while the work happens.

Operational Context Is the Part You Can’t Buy

I ask the same thing on every call: How does your team decide who gets the load? The first answer is always the policy, some document written for an insurance audit that the floor never opens. The second answer, the one that’s true, is a name.

That person carries the exceptions, and the exceptions are the business: the carrier worth $200 over because he never misses a 6 a.m. appointment, the MC with two bad pickups who still gets freight because his dispatcher called back when the market flipped.

Harvard Business Review named this in February: when everyone runs the same models, the difference is context, and it lives in how work gets done, not documentation. Harvard calls that an advantage. I’d call it an advantage you’re renting from an employee.

She has every reason to keep it that way too. Since writing it down is essentially the same as writing herself down. Ellie catches it as the work happens, so every booked load builds a Carrier Context Graph the brokerage owns.

Custom Models Are an Ego Purchase

Say the word own in a boardroom, and somebody suggests building your own model, trained privately on your data with the keys in your drawer. After 20 years of renting software, it’s the rare pitch that flatters everybody in the room.

Michelle, our head of product, spent two years building one before she joined us: a service-prediction model across $19B in freight. It worked. The floor kept covering loads the way it always had, and people asked it questions like you’d ask an expensive coworker with no authority.

Her verdict, and I’ve never improved on it: “A custom model is an infrastructure decision and an infrastructure investment. It is not an operational strategy.”

The exceptions I just described move every week. Weights don’t. You’d spend six to nine months cleaning data to bake last year’s rates into a model. Gartner has 40%+ of agentic projects dead by 2027, mostly from starved context. You never needed more data. A model that memorizes your history still shows up knowing nothing about this morning.

June Repriced the Market, and Nobody Told Your Model

I spent June on calls with operators rewriting max pay in the middle of the week. Rejections hit 17.64% on June 21, the worst since March 2022. The spot printed an all-time high of $3.78 a mile seven days after that. By August it settled at 14.1%, still triple last year’s 4.75%.

Your max pay table was written when the spot looked nothing like that. If it went untouched through June, your team spent the month bidding to lose, and everyone assumed the reps were slow. That is a policy problem in a technology costume.

So when the next vendor demos, make them show you where today’s numbers come from. If explaining it takes more than a sentence, the system is pricing your freight off the spring while a faster shop takes the truck.

That race is decision latency, and live operational context is the only thing that shortens it.

Wrong, Confidently, in Perfect English

Speed only helps if the agent is right, so let me argue against my own pitch for a second. A fast agent with no rules is a quicker way to hand your freight to a carrier you would never have approved.

VentureBeat surveyed 101 enterprises in June, and 57% had caught an AI being confidently wrong from missing business context. ChatSee reviewed 10,000+ AI failures in July and put hallucinations behind under 10%, while execution failures climbed 62% since Q2 2024.

What makes that dangerous on a carrier desk is that the failure looks like competence. A rate $180 over your ceiling, going to a carrier you pulled off that lane in March, in a sentence written better than anything your team sent that day. Somebody skims it and approves.

Guardrails are the piece most vendors leave to you. Ellie carries yours into the negotiation itself, which is the entire point of the logistics AI stack she runs on.

Montgomery Turned Your Best Rep Into a Liability

I bring this up on calls now and watch the room change. In May, the Supreme Court decided Montgomery v. Caribe 9-0; FAAAA preemption is gone, and negligent selection goes to state court, where how you picked the carrier is evidence.

I’m not a lawyer, and I won’t pretend this ends brokerage. But an attorney will ask your ops director to explain, one load at a time, why that carrier got that freight. “Our senior guys know a bad MC when they see one” is a rough answer to give under oath.

The ruling itself won’t save you, and fraud makes it worse. CargoNet counted $304M in cargo theft losses in Q2, double last year, most of it account compromise and carrier impersonation. Those crews are organized, and they are beating vetting that lives in somebody’s head.

Ellie checks MC, DOT, authority, and safety before she books, and stores why she cleared each one. Ask your process what it does at 4:47 with the TMS down and the carrier ducking calls. That answer is your exposure.

While You Evaluate, Somebody’s Compounding

Exposure is the cheap part of waiting, and I say that as a guy whose pitch benefits from you feeling exposed. A better model arrives for everybody at once, while whatever separates you from the shop across town accrues a load at a time.

So when BCG’s January survey finds nearly 70% of shippers still exploring or piloting and 1% with AI inside a core process, I don’t read an industry behind on technology so much as one that hasn’t started building the asset technology can’t hand it.

And I get the stalling. Most of these projects arrive as an integration, six months of IT before a load moves, and refusing that is right. Ellie runs as a Chrome extension over what your floor already has open, which is less impressive than it sounds and more useful. That’s because operational context starts building the day she covers freight.

In our live accounts, 75% of freight books through her, and every one of those loads makes the next negotiation sharper. That isn’t something you can buy back with better software in 2028.

Bring Me a Load You Rolled Last Week

I’ve said for years the carrier rep role doesn’t survive to 2036 in the shape it’s in, and I’ve taken plenty of heat for it. So here’s a little more. The shops treating 2026 as another evaluation year will spend 2028 buying their way out at worse terms, from vendors with less reason to be generous.

Let’s skip the nurture sequence. Book a demo, bring your live portal and a load you rolled last month, and we’ll cover it in front of you. Or go around the form and email me at robby@tryenvoy.ai with your worst lane and your current time-to-cover. I’ll answer those myself.