AI Freight Dispatch Software: How It Finds and Ranks Loads

Most explanations of AI freight dispatch software stop at "it finds you better loads." That is a marketing claim, not a description of how the thing works. If you run a fleet, you want to know what the software is actually doing between the moment a load gets posted and the moment your truck is rolling under a signed rate confirmation.

There are three real stages, and they are not equally automated. The software finds loads by searching many sources at once and turning messy postings into clean data. It ranks them by doing math you would do yourself if you had time, against rules you set. And it drafts negotiation messages toward a target price, but a human approves before anything goes to a broker. This post walks through each stage and is honest about which parts run on their own and which parts stay in your hands.

Stage one: finding loads across fragmented sources

Finding loads sounds simple until you count the places a load can live. A small carrier might watch two or three load boards, a handful of broker portals and TMS connections, plus an email inbox where brokers send direct offers. With roughly 27,000 active brokers in the market and around 787,000 carriers competing for freight, the postings are scattered, duplicated, and inconsistent. No dispatcher can refresh every source every minute. That is the gap the software fills first.

What the find stage does is poll those sources continuously instead of on a human refresh cycle. It holds your saved searches, your lanes, and your equipment types, and it checks each connected board and portal on a tight loop. When a matching load appears, it pulls the posting in. This is the part that is genuinely automated, and it is automated because it is repetitive and time-sensitive: the difference between seeing a good load when it posts and seeing it twenty minutes later is often the difference between booking it and watching it disappear.

The harder, less obvious work in this stage is normalization. A load board posting and a broker email describe the same load in completely different shapes. One leads with rate, another buries it. One gives a pickup window, another gives a single appointment time. Commodity, weight, equipment, and accessorial terms show up in different fields, in different words, or in a free-text blob a human wrote in a hurry. The software reads those inconsistent postings and maps them into the same structured fields every time: origin, destination, miles, equipment, pickup window, delivery window, commodity, weight, posted rate, broker. Once every load is the same shape, you can compare them, sort them, and rank them. Without normalization there is nothing to rank, because you cannot score a pile of postings that do not share a vocabulary.

It is worth being clear about what the find stage does not do. It does not decide anything. It does not commit you to a load or hide loads it dislikes. It is plumbing and translation: gather widely, then make everything legible. The judgment comes next.

Stage two: ranking by all-in rate against your rules

A list of clean loads is still just a list. The rank stage is where the software earns its keep, because it scores each load the way an experienced dispatcher would if the dispatcher had a calculator running on every posting at once.

The number that matters is the all-in rate, not the posted rate. Posted rate is what the broker advertises for the loaded miles. All-in rate is what you actually net once you account for the deadhead miles your truck runs empty to reach the pickup. Deadhead commonly runs anywhere from 15 to 30 percent of total miles, and ignoring it is the most common way a load that looks good turns out to lose money. The software pulls the deadhead from your truck's current or planned position to the origin, adds those empty miles into the denominator, and computes a real revenue-per-mile figure you can trust.

Here is the math, worked on two loads that look similar on a board:

On the board, Load A pays more in total and Load B has the higher posted RPM. But once deadhead is folded in, Load B is clearly stronger at $2.64 all-in versus $2.11. A dispatcher juggling forty postings will not run that arithmetic on each one. The software runs it on all of them in the same pass, which is the whole point: it surfaces the real ranking, not the advertised one.

All-in RPM is only the first filter. The rank stage then scores each load against the rules you define, the same constraints a good dispatcher carries in their head. A useful floor to anchor against is cost: ATRI's 2025 report put the marginal cost of operating a truck at roughly $2.26 per mile for 2024, so a load whose all-in RPM lands below your cost is one you usually want flagged or dropped, not surfaced. Other rules stack on top:

What the ranking weighs

• RPM floor. Your minimum acceptable all-in rate per mile, set against your real operating cost, not a generic benchmark.
• Lane fit. Whether the destination moves the truck toward your next committed load or your home base, or strands it somewhere with no outbound freight.
• Equipment match. Dry van, reefer, flatbed, and the specific requirements the load demands.
• Broker quality. Payment history and a reliability score, so a slightly cheaper load from a slow-paying broker can rank below a cleaner one.
• Timing. Whether the pickup and delivery windows fit the driver's available hours without forcing a violation.

The output is an ordered shortlist with the reasoning attached, not a single take-it-or-leave-it pick. You see why Load B ranked above Load A, and you can override the order when you know something the rules do not. The ranking is automated; the decision about which ranked load to pursue stays yours.

Stage three: negotiating and booking, under human approval

The posted rate is rarely the final rate, and brokers price with that in mind. Industry margin data from DAT put the average broker gross margin around 13.5 percent, which tells you there is usually room between the first posted number and what the broker will actually pay. The negotiate stage is where the software helps you close that gap, and it is also the stage where human control matters most.

Mechanically, the software drafts the outreach. For a ranked load you want, it composes a message to the broker, often an email, that opens at a target rate. That target is not a guess: it is derived from your RPM floor, the all-in math from the rank stage, and the room the posted rate implies. When the broker counters, the software reads the reply, updates the picture, and drafts your counter toward the same target, holding the line you set rather than drifting toward whatever the broker offers. It handles the repetitive back-and-forth drafting that eats a dispatcher's day, the same way a strong assistant would.

What it does not do is send anything on its own, and that is deliberate. Every offer and counter sits in front of a human before it goes out. You read the draft, adjust the number or the tone, and approve it, or you reject it and write your own. The reason is not timidity; it is that a rate commitment is a real commitment. It affects revenue, the relationship with that broker, and your driver's next several days. Automating the drafting saves time without much downside. Automating the send would hand away the one decision that defines whether the load is profitable, and would risk your standing with a broker you may run a hundred more loads with. The honest version of this software keeps the human as the approver, not the bottleneck. You are not writing every message from scratch; you are reviewing and authorizing.

Booking works the same way. Once a rate is agreed, the software can prepare the rate confirmation, populate the load into your records, and queue the next steps. But the click that accepts the load, like the click that sends the counter, is a human one. The pattern across all three stages is consistent: automate the gathering and the math and the drafting, keep the commitments under a person's hand.

Why the human stays in the loop

It is tempting to ask why any of stage three needs a human at all, if the software can already find, score, and draft. The answer is that the cost of a wrong commitment in freight is asymmetric and hard to reverse. The errors are not abstract. Detention alone costs the industry an estimated $1.1 to $1.3 billion a year, and cargo theft tracked by CargoNet reached around $725 million in 2025, much of it through fraud and misdirected loads that a hurried, unreviewed booking makes easier to fall for. A dispatcher's judgment about a sketchy broker, an appointment that will not hold, or a lane that strands the truck is exactly the kind of context the rules do not fully capture.

There is also a simpler economic point. A dispatcher's median pay runs around $46,860 a year, and the leverage of the software is to make that person handle far more loads at higher quality, not to remove them. When the find and rank stages do the searching and the math, the dispatcher spends their time on the decisions that actually need a human: which broker to trust, when to walk from a negotiation, which load protects next week's schedule. That is the real shape of AI freight dispatch software working well. It is not an autopilot. It is a system that does the tireless, repetitive, error-prone work continuously and accurately, and then hands a clean, ranked, half-drafted decision to a person who is still in charge.

The takeaway is to judge any dispatch tool by where it draws that line. If it claims to book loads without you, be skeptical, because the stages that should stay human are the ones it is automating away. If it finds and ranks relentlessly and drafts negotiations toward your target while leaving the commitment to you, it is built the way the work actually runs. That is the model behind Numeo AI Hub: the AI dispatcher works under the dispatcher's control, finding, ranking, and negotiating, with the human approving every load that matters.