AI Dispatch Rollout: A Real Pilot-to-Production Plan

Most AI dispatch rollouts fail for a boring reason: the carrier turns it on across the whole board, nobody trusts it, and it gets quietly switched off within a month. The technology is rarely the problem. The rollout is. What works instead is narrow and measured — one lane, one dispatcher, every recommendation reviewed, and expansion that follows evidence rather than enthusiasm. This is the version of that plan a carrier can actually follow, with the parts where a human stays in control called out honestly.

Baseline before you change anything

You cannot tell whether AI dispatch helped if you never wrote down where you started. This is the step everyone skips, and it is the one that decides whether the pilot ends in a real number or a shrug. Spend the first few days measuring your current operation by hand, on the lane and dispatcher you plan to pilot — not the whole fleet. You want a small, honest baseline, not a fleet-wide average that hides the truth.

Pull these from your TMS, your load-board history, and a week of watching one dispatcher work. The point is to capture how the job runs today, friction included.

Two of these deserve context so you read them right. Deadhead in the 15–30% range (the spread varies by fleet and lane) is the single biggest controllable cost after the truck itself, which ATRI's 2025 report put around $2.26 per mile for 2024. And broker margins sitting near 13.5% (DAT, 2023) are a reminder that the rate you negotiate by email is real money — a few points of RPM on a lane compounds fast. Write these numbers down somewhere you will look at them again in four weeks. That document is the whole point of the pilot.

Scope the pilot so it can't hurt you

The instinct is to pilot broadly to "really test it." Resist that. A good pilot is deliberately small: one lane or a tight lane group, one equipment type, and one or two dispatchers who are good at their job and willing to be honest about what they see. Narrow scope is not timidity — it is what makes the results legible. When something looks wrong, you want to know exactly which lane, which broker, and which rule produced it, not sift through the whole board.

Pick a lane you run often enough to get a real sample inside a few weeks, with brokers you already know. Familiar territory means the dispatcher can immediately tell a good AI recommendation from a bad one, because they already know what the right answer looks like. Avoid piloting on your most chaotic, exception-heavy freight; you are testing the workflow, not stress-testing it. Save the hard lanes for after the system has earned some trust.

Before you switch anything on, write down the rules the AI operates inside — and keep them tight. Minimum acceptable RPM for the lane, maximum deadhead you will tolerate, brokers to exclude, your driver's hours and home-time constraints, and the tone you want broker emails written in. These are not settings you set once and forget; they are the carrier's operating judgment, made explicit so the software can respect it. The narrower and clearer the rules, the easier it is to see when the AI is following them and when your rules themselves were wrong. Most pilots discover their own rules need fixing before the AI does.

Run it review-only, with a human on every commit

For the whole first stretch, the AI proposes and the dispatcher disposes. It can watch the boards, rank loads against your rules, draft the broker email, flag a good lane, and surface the market context — but it commits to nothing. No load booked, no rate sent, no driver assigned without a human clicking approve. This is not a temporary inconvenience you tolerate until the AI "graduates." For the decisions that move money and relationships, a human staying in the loop is the design, not a training-wheels phase.

This matters most in negotiation. Numeo negotiates with brokers primarily by email today, and a drafted email is exactly the right unit of human control: the dispatcher reads the counter the AI wrote, sees the reasoning and the market context behind the number, edits the tone if it is off, and sends it — or doesn't. The broker relationship is the carrier's asset, and price is the carrier's call. The AI's job is to get a well-reasoned draft in front of the dispatcher in seconds instead of minutes, not to quietly commit the carrier to a number. That is the line that keeps a rollout safe, and it is the line you should refuse to blur early.

Review-only is also how you find the system's blind spots cheaply. When the dispatcher edits a recommendation, that edit is data — it tells you a rule was wrong, a data source was missing, or context the AI couldn't see mattered. Keep a running list of why dispatchers override. Early on they will edit nearly everything, and that is fine; you are not looking for zero edits, you are looking for the edit rate to fall as the rules and data tighten. A dispatcher who trusts the draft enough to send it with a one-word tweak is the signal you are actually waiting for.

Read the results honestly

At the end of the pilot, put the new numbers next to the baseline you wrote down and resist the urge to grade on a curve. The comparison only means something because you captured the "before" by hand. Look at the same metrics: realized RPM against market, deadhead on the lane, time to first action, broker-message turnaround, and the exception rate. The first four should move in your favor. The last one — exceptions, the manual rescues — must not get worse. A pilot that books faster but creates more messes downstream is not a win; it has just moved the work.

The most informative metric is the one that is easy to overlook: how often the dispatcher accepted the AI's recommendation with little or no edit, and whether that acceptance climbed over the weeks. Rising acceptance on a stable or improving exception rate is the real green light. Flat acceptance means the rules or the data still don't reflect how you actually operate, and that is a fixable problem — but you fix it before you expand, not after. Be suspicious of a single headline number in isolation. Faster is good only if the rate held and the mess didn't grow.

Separate the wins by type, because they expand differently. If the alerts were accurate and the dispatcher caught loads they'd have missed, that part is ready to widen. If the email drafts saved real time and went out with minor edits, that part is ready to widen. If the load ranking was weak, that part is not ready — improving the rules behind it comes first. Most pilots are not uniformly good or bad; they are good at two things and mediocre at a third. The honest read tells you which is which.

Expand only what earned it

Production is not a switch you flip; it is the pilot widening one proven piece at a time. Take the parts that worked and extend them to an adjacent lane, the next equipment type, or another dispatcher — and let each new slice clear the same bar the first one did before you widen again. The thing you are scaling is consistency: a system that prices and negotiates roughly the way your best dispatcher would, applied evenly instead of depending on who is at the desk that day.

Autonomy should be earned per decision category, not granted wholesale. A specific lane, a specific broker, a specific rate band where the AI has been reliably right for weeks can move to lighter review — the dispatcher spot-checks rather than approves every one. Everything outside that proven box stays review-only until it earns its way out. This keeps the risk bounded at every step: the worst case on any new category is a few caught edits, not a board full of bad commitments. Try to flip everything to autonomous booking at once and you will get the same quiet sabotage that kills most rollouts, no matter how good the model is.

A few honest expectations as you scale. The economics are real — a dispatcher costs roughly $46,860 a year (BLS, 2023), and the leverage of AI dispatch is consistency and reach per person, not replacing that person. Keep the human on price, broker relationships, driver assignment, and anything touching compliance or service commitments, because those are exactly the decisions where being wrong is expensive and hard to reverse. And keep your baseline document alive; re-measure each new lane so expansion stays evidence-driven rather than a leap of faith.

The whole plan reduces to one discipline: measure first, scope small, keep a human on every commitment, and let each piece earn its expansion. Done that way, the downside is capped at every step and the upside shows up in the numbers you already track. If you want to see what a carrier-native version of this looks like, the AI Hub is the layer that ranks, drafts, and negotiates on top of the systems you already run — with your dispatchers holding the approval the whole way.