AI Hub for Logistics: How Agentic Dispatch Workflows Run

A dispatcher running a small fleet today works across a load board, two or three broker portals, an email inbox, a phone, a spreadsheet, and a TMS, and the day is mostly spent moving the same load details between those screens. Each of those tools does one thing well and knows nothing about the others. The result is predictable: good loads get missed because the team noticed them late, and the same rate gets re-typed four times before it reaches a rate confirmation.

An AI Hub is the response to that fragmentation. Instead of one more standalone tool, it is a place where AI agents run the dispatch loop as a single coordinated workflow: find the load, rank it, draft the reply, surface it to a human, and book it once a person approves. This post explains what that means concretely, how "agentic" software actually behaves, and where the human stays firmly in control. Numeo's AI Hub is the working example throughout.

What an AI Hub actually is

Start with what it is not. An AI Hub is not a smarter load board, and it is not a chatbot bolted onto a TMS. A load board is a catalog you search. A TMS is a system of record you fill in after decisions are made. Both are passive: they wait for a dispatcher to come to them, do something, and leave. An AI Hub is active. It watches the sources you care about, does the preliminary work, and brings you a short list of things worth a decision.

The "hub" part matters because dispatch fragmentation is the core problem, not any single missing feature. Loads arrive on boards and in broker emails. Context lives in market data and your own lane history. Communication happens over email. The commitment gets recorded in a TMS. When those four areas sit in four disconnected tools, the dispatcher is the integration layer, manually carrying state between systems all day. A hub collapses that into one surface so the handoffs happen in software instead of in a person's short-term memory.

The "AI" part is what lets the hub do work rather than just display it. Reading an unstructured broker email and pulling out the lane, equipment, pickup window, and rate is an AI task. Comparing a posted rate against what the lane has been paying is an AI task. Drafting a counteroffer in your voice is an AI task. None of these are commitments — they are the preparation that used to eat the dispatcher's day. The hub does the preparation and leaves the commitment to a human.

What "agentic" means, without the hype

"Agentic" gets used loosely, so here is a plain definition: agentic software takes multiple steps toward a goal on its own, and pauses for human approval before any step that creates a real commitment. That is the whole idea. It is more than a single API call that returns an answer, and it is less than a fully autonomous system that acts without oversight.

Concretely, a non-agentic tool waits for one instruction and returns one result: you type a lane, it returns matching loads. An agentic workflow is given a goal — "keep this truck loaded on lanes that fit our rules" — and then chains several actions to pursue it: it monitors boards, reads new broker offers as they land, normalizes each into structured fields, scores them against your rules, and prepares a draft reply for the best ones. It does all of that without being re-prompted at each step. What it does not do is hit send, accept a load, or commit a truck. Those are where it stops and waits for a person.

The honest framing is that the agent does the legwork and the human makes the calls. That boundary is not a limitation the technology will eventually outgrow; in freight it is the correct design. A booked load is not a click — it commits a truck, a driver's hours, a delivery appointment, and a customer relationship. Software should not own that decision, and a well-built AI Hub does not pretend to. It earns trust by being right about the preparation, not by removing the human from the commitment.

It is worth being precise about communication, too. Numeo's agents negotiate with brokers primarily by email today — drafting and sending structured offers and counteroffers through the inbox where most broker business already happens. That is a deliberate, reviewable channel, not an autonomous phone bank.

How the dispatch loop runs inside the hub

The workflow is a loop with five stages, and the human sits at one specific point in it. Here is the shape of it:

Find is signal collection. The agent watches the load board and the broker portals you use, and it reads new broker emails as they arrive, so a posted load or an updated offer reaches the hub without anyone refreshing a tab. The point is that the dispatcher stops being the polling mechanism. In Numeo, this is where Load Hub feeds the picture by pulling from many freight sources into one place, and Load Radar surfaces matching loads the moment they post.

Rank is where raw freight becomes comparable. Broker postings are inconsistent — one leads with price, another with the delivery appointment, a third buries the commodity. The agent turns each into the same structured fields, then scores it against rules you define: target rate per mile, maximum deadhead, lane preferences, equipment fit, driver hours, and broker history. Market context goes in here, so a rate is judged against what the lane is actually paying rather than in a vacuum. The output is a ranked shortlist, not a longer list.

Draft is preparation for action. For the top loads, the agent writes the broker reply or the counteroffer in your voice and terms, ready to review. Surface is the handoff: the shortlist and the drafts are presented with the reasoning attached, so the dispatcher can see why a load ranked where it did before deciding anything. Book is the only stage the agent does not own. The dispatcher approves, edits, or rejects; on approval, the agent sends the message or records the booking. The loop then repeats for the next truck. The agent compresses four screens of preparation into one reviewable shortlist; the person keeps the decision.

Why this matters for carrier economics

Controlled automation is worth building because trucking margins are thin enough that preparation quality moves real money. ATRI's 2025 report put the marginal cost of operating a truck at roughly $2.26 per mile in 2024, and every empty mile is paid at that cost with no revenue against it. Deadhead commonly runs 15 to 30 percent of miles. A workflow that scores deadhead into every ranking, before a load is taken rather than after, is defending a number that directly determines whether a lane is profitable.

The same logic applies to rate. Brokers operate on a net margin around 13.5 percent (DAT, 2023), which is the spread between what a shipper pays and what the carrier sees. A carrier that responds faster and counters with market context has a better chance of closing that spread in its favor. The agentic loop helps here not by being clever but by being fast and consistent — it drafts the counteroffer the moment the offer lands, with the lane's recent rates already attached, instead of the dispatcher getting to it an hour later.

Scale is the other reason this is a hub problem rather than a headcount problem. FMCSA counted roughly 787,000 active carriers at the end of 2023, and the ATA reports that about 91.5 percent of them run ten trucks or fewer (2025). Those are exactly the operations that cannot hire an analyst to watch every board and a back-office team to chase every rate confirmation. The hub gives a two-person dispatch desk the coverage that used to require a room of people, without handing any of them the authority to commit the company.

Where the hub stops, on purpose

The clearest test of whether an AI Hub is built honestly is where it refuses to act. Automation belongs on work that is repetitive, rules-based, time-sensitive, and easy to review: monitoring saved searches, normalizing load details, flagging incomplete postings, drafting routine broker replies, preparing counteroffers. Get those right and the dispatcher's day changes from data entry to decisions.

Automation should stop where the task needs commercial judgment, relationship judgment, or a commitment the carrier cannot easily reverse. Final price, accepting a load, assigning a driver, and the broker-relationship calls all sit on the human side of the line. The discipline is that every recommendation carries its reasoning, every draft is editable, and every commitment pauses for approval. That audit trail is also what makes the system defensible after a dispute — an operations lead can see exactly what was recommended, what was sent, and who approved it.

This boundary is sound for security reasons as well as operational ones. Broker emails, rate data, and driver records are commercial and personal information, and cargo theft alone reached an estimated $725 million in 2025 (CargoNet). A workflow that keeps a human on every outbound commitment and logs every action is easier to govern than one that acts on its own and explains itself later.

The takeaway is simple. An AI Hub is not autonomy and it is not a faster load board — it is the dispatch loop run as one coordinated, agent-driven workflow with the human kept on the commitments that matter. Done right, the agents handle the finding, ranking, and drafting that used to fill the day, and the dispatcher spends time deciding instead of typing. That is the model Numeo's AI Hub is built around, and it is the standard worth holding any "AI logistics" product to.