How to Set the Rules an AI Dispatcher Books Within

An AI dispatcher is only as safe as the rules you give it. Hand it nothing and it will chase cheap freight, run empty across the country, and email brokers you stopped trusting six months ago. The work that makes automation trustworthy happens before any load gets booked: you write the constraints that define what "a good load for this truck" actually means, and the machine books inside them.

This is the part carriers tend to skip because it feels like overhead. It isn't. Rules are the product. A dispatcher's judgment, written down as numbers and lists, is what lets you delegate the booking and still sleep. Below is how to set the ones that matter — RPM floors, deadhead caps, lane preferences, broker blocklists and minimum scores, equipment and hours-of-service constraints, and the thresholds where the system has to stop and ask you.

Start with the RPM floor, and make it clear cost

The first rule is the one that protects you from losing money on every mile: a rate-per-mile floor the AI will not book below. Most carriers set this from gut feel or from what they booked last week, which is how you end up hauling freight that doesn't cover the truck. The honest way to set it is to start from cost. ATRI's 2025 report (on 2024 data) put the average marginal cost of operating a truck at roughly $2.26 per mile — fuel, wages, maintenance, insurance, the truck itself. Whatever your floor is, it has to clear your version of that number with margin on top, or the load is a donation.

The floor is rarely a single number, and that's the point of writing rules instead of trusting a vibe. RPM economics differ by lane, equipment, and season. A reefer floor in produce season is not a dry-van floor on a backhaul lane in February. So you set a base floor, then override it where you have a reason: a higher floor on lanes you know run hot, a lower one on a lane where any revenue beats sitting, a separate floor per trailer type. The discipline is that every number is deliberate. When the AI declines a load because it's eleven cents under floor, you want to know that floor was a decision, not an accident.

There's a second framing worth building into the floor: the rule should compute on all-in revenue over loaded plus deadhead miles, not the broker's quoted line-haul over loaded miles only. A load that pays well on paper but requires a long empty repositioning can clear your line-haul floor and still lose money once you spread the rate across the real miles. Tell the AI to evaluate the trip, not the posting — total revenue divided by total miles, deadhead included. That's the difference between a floor that protects you and one that just looks like it does.

Cap deadhead, and decide what counts as "too far for empty"

Deadhead is the quiet tax on a fleet. Industry deadhead runs somewhere in the 15-30% range depending on how it's measured and what segment you're in, and most of it is invisible at the moment of booking because the empty miles to the next pickup don't show up on the rate confirmation. An AI dispatcher books one load at a time, so without a rule it will happily accept good-looking freight that strands the truck 180 empty miles from the next opportunity.

The constraint that fixes this is a maximum deadhead — a hard cap on the empty miles the system will accept to reach a pickup, expressed either as a flat distance or as a percentage of the loaded miles on the load it's considering. A flat cap (say, no more than 75 deadhead miles to pickup) is simple and works well in dense freight markets. A percentage cap (deadhead no more than 15% of loaded miles) scales better for long-haul, where 100 empty miles to grab a 1,400-mile run is fine but the same 100 miles for a 200-mile run is not. Many carriers set both and take the tighter of the two.

Deadhead rules get more useful when they're tied to position rather than treated in isolation. A high deadhead to reposition into a strong outbound market can be worth it; the same empty miles into a dead market is just loss. The AI can't make that call unless you've told it which regions you want the truck to end up in, which is where deadhead caps and lane preferences start working together. The cap keeps the machine honest on empty miles; the lane rules tell it which empty miles are an investment.

Write the lane rules: preferred, excluded, and the home-base pull

Lane preferences are how you encode where you actually want your trucks to run. At the simplest level this is two lists: lanes or regions you prefer, and lanes you exclude outright. Preferred lanes get a scoring boost so the AI reaches for them first; excluded lanes are hard filters it never books, whether because they're chronically cheap, run through weather or theft corridors you avoid, or just don't fit the network. The exclusion list does as much work as the preference list — saying no to the wrong freight is most of dispatch.

The rule that earns its keep on most fleets is the home-base or relay pull: a constraint that biases the truck toward getting back to a home region by a certain day, or toward a driver's domicile for a reset. This is where automation without rules goes badly wrong, because the cheapest next load is almost never the one that gets the driver home on Friday. You write the constraint — "prioritize loads delivering within 150 miles of Dallas on Thursday or Friday" — and the AI weighs cents-per-mile against getting the truck where it needs to be. The trade-off is yours to set; the machine just executes it consistently, which is more than a tired human juggling six trucks at 4 p.m. usually manages.

A practical lane rule set looks less like a policy document and more like a short table. Here is a workable starting point for a small dry-van fleet running the southern half of the country:

The values are illustrative, not a recommendation — yours come from your own network. The point is that the whole policy fits on a card, and the AI applies it identically to every load it sees, all day, without drifting.

Broker blocklists and a minimum broker score

Who you book with is a rule, not an afterthought. There are roughly 27,000 brokers in the market, and they are not equal — some pay on time and reconfirm cleanly, some are slow, and a growing few are fronts for fraud. Double-brokering is rising, and cargo theft has gotten expensive: CargoNet logged around $725 million in reported losses in 2025, with an average of about $273,990 per event. A meaningful share of that traces back to booking with a party you shouldn't have. A blocklist is cheap insurance.

So the AI carries two broker constraints. The first is a hard blocklist — MC numbers it will never book with, full stop, sourced from your own bad experiences and from ops as new fraud patterns surface. The second is a minimum broker score: a threshold below which the system won't book without a human looking first. The score can blend credit/days-to-pay data, your own payment history with that broker, and how long they've been authorized. New authority plus a too-good rate is a classic double-broker signature, and a score floor catches exactly that combination before the load is on your truck instead of after.

Brokers also sit on the other side of your margin, which is worth remembering when you set negotiation rules alongside the score. Broker margins run around 13.5% on average (DAT, 2023), so there's usually room in a posted rate. You can let the AI negotiate within a band you define — open at a target, accept down to your floor, walk below it — but the broker score should gate whether it negotiates at all. A high-trust broker gets the automated counter; an unknown one gets flagged to you regardless of the rate, because the rate is not the risk. The party is.

Equipment, hours, and the constraints that aren't negotiable

Some rules aren't preferences you're optimizing — they're physical and legal facts the AI must respect or it books something the truck literally cannot do. Equipment is the obvious one: trailer type, length, and any specialized capability (reefer with a working unit, food-grade wash, liftgate, team-capable) are hard filters. There's no scoring trade-off here. A load that needs a reefer and a truck that pulls a dry van is not a slightly-worse match; it's not a match, and the system should treat it as invisible.

Hours of service is the constraint where automation has to be most conservative, because the cost of getting it wrong is a violation or an exhausted driver, not just a thin margin. The AI needs the driver's available hours and clock as an input and must not book a load whose pickup, transit, and delivery appointments can't be met legally within the remaining drive and on-duty windows, including realistic detention and break time. This is also where you build in slack: detention alone costs the industry an estimated $1.1-1.3 billion a year, and an AI that schedules to the theoretical minute will book trips that blow up the moment a shipper runs two hours late. Tell it to assume things slip, because they do.

These hard constraints are the ones you should be least tempted to soften for a good rate, and the system should be built so it can't be tempted either. Equipment mismatch, an HOS-impossible schedule, a load requiring an endorsement the driver doesn't hold — these are filters, not factors. Keeping them absolute is part of what makes the rest of the automation safe: when the AI does surface a load, you already know it's a load the truck can legally and physically run, and the only open questions are the ones worth a human's attention.

Approval thresholds: where the machine stops and you decide

The last rule set is the most important for trust, and it's about where automation hands control back. An approval threshold defines the line between what the AI books on its own and what it stages for your sign-off. Set it too loose and you've built the black box nobody wanted; set it too tight and you've automated nothing, because every load waits on a human anyway. The right line is specific to your risk tolerance, and it usually keys off a few clear triggers.

A sane default: let the AI auto-book loads that clear the RPM floor, sit inside deadhead and lane rules, come from a broker above the score threshold, and fit equipment and HOS cleanly — the loads where every rule is comfortably satisfied. Everything else stages for review. That includes any rate within a few cents of the floor, any broker below score or with new authority, any deadhead near the cap, any load that pushes HOS, and anything above a dollar ceiling you set per load. The thresholds are yours to tune, and most carriers start strict — auto-book almost nothing — then loosen specific triggers as they watch the system make calls they agree with.

This is the honest answer to "does AI replace the dispatcher." No. It moves the dispatcher from doing the repetitive booking to setting the rules and handling the exceptions the rules flag. The human stays in control of exactly the decisions that should never be fully automated — price near the edge, unfamiliar counterparties, anything irreversible — while the machine handles the high-volume, clearly-in-bounds work that used to eat the day. That division is the whole point. Good rules don't make the dispatcher obsolete; they make the dispatcher's judgment scale.

The takeaway is simple: automation is only as safe as the constraints behind it, so spend your effort on the rules, not on hoping the AI guesses right. Write the RPM floor against real cost, cap your deadhead, list your lanes and your bad brokers, lock equipment and HOS as hard filters, and draw a clear approval line — then let the system book inside the box you built and bring you the rest. That's how you delegate dispatch without giving up control of it. If you want to see what booking inside carrier-defined rules looks like in practice, that's the model behind AI Hub.