Cross-Border Route Optimization in North America: Why the Border is a Queue and Not a Road Segment

Cross-border route optimization in North America plans movements that pass through a port of entry between the United States, Mexico and Canada. Nearly every routing engine represents that crossing the same way it represents a bridge or a motorway segment: an edge in the road graph with a travel time attached, perhaps adjusted for time of day. A port of entry is not an edge. It is a queue with a finite number of open lanes, published operating hours, an eligibility-gated express lane, and a wait that rises nonlinearly with how many trucks are trying to cross. Modeling it as an edge is not a small simplification. It produces plans whose error is largest precisely when the network is busiest, and it removes the one property that makes fleet-level planning work everywhere else: that delays average out across vehicles.

Why the crossing deserves its own model

The volumes make the queue real rather than theoretical. The Bureau of Transportation Statistics border crossing data program publishes inbound crossings at port level for the US-Canada and US-Mexico land borders, and recorded 687,577 trucks entering the United States from Mexico in October 2025, up 1.4% on the same month a year earlier. That is a single month on one border, arriving through a limited number of commercial ports with a limited number of booths.

The wait is also measured and published, which is unusual and useful. US Customs and Border Protection operates a live border wait times service reporting current and average commercial vehicle delays per port along with the number of lanes open, and BTS maintains the wait time data links for historical series. Published commercial averages at major crossings commonly sit in the tens of minutes, with a distribution that runs much longer. Any planner can obtain this data, and very few routing configurations consume it.

The cost of the delayed hour is well documented. ATRI’s operational cost analysis put the industry-average cost to operate a truck at $2.336 per mile in 2025, up 3.4%, with truckload and refrigerated operating margins below 1.0%. On margins that thin, a systematically optimistic border assumption is not an inconvenience. It is the difference between a lane that clears its cost and one that does not.

The policy backdrop adds a reason to build this properly now rather than later. At the joint review on 1 July 2026 the United States declined to confirm its intention to extend USMCA for a further sixteen years, triggering annual reviews, with tighter automotive rules of origin expected to feature in continuing discussions. Cross-border flows are not becoming simpler, and documentation and inspection intensity is the variable most likely to move.

Also Read: Truck Route Optimization: Enterprise Guide 2026

How to model a port of entry properly

1. Understand why border delay does not average out

This is the property that matters most and it is almost never stated. On an open road network, congestion delays hit different vehicles at different places and times, so they are quasi-independent and the fleet-level average is stable. The standard deviation of the fleet mean falls as the square root of the number of vehicles.

A border queue does the opposite. Every truck in the queue at a given hour experiences the same conditions, so the delay is common-mode and the fleet-level variance does not shrink at all.

Vehicles crossing together	Independent delay, fleet-mean spread	Common-mode delay	Relative disruption
4	sigma divided by 2.0	sigma	2.0x
10	sigma divided by 3.2	sigma	3.2x
20	sigma divided by 4.5	sigma	4.5x
50	sigma divided by 7.1	sigma	7.1x

A border delay of the same magnitude as a traffic delay is therefore roughly the square root of n times more disruptive to a fleet plan. At twenty trucks crossing in the same window, 4.5 times. Scale does not protect you here, which is the opposite of the intuition that larger networks absorb variability better.

2. Size the buffer against the tail, not the mean

Queue waits are right-skewed. Treating them with a standard exponential approximation gives a clear and uncomfortable result.

Target reliability	Buffer required, as multiple of mean wait	Applied to a 22-minute published average
50%	0.69x	15 minutes
80%	1.61x	35 minutes
90%	2.30x	51 minutes
95%	3.00x	66 minutes
99%	4.61x	101 minutes

Planning on the mean wait itself succeeds only 63% of the time. This is a queueing approximation rather than a measured distribution, so calibrate it against your own port-level history, but the shape is the point: reliability at a border is bought in multiples of the average, not in additions to it.

3. Treat the operating-hour cutoff as a cliff

Many commercial crossings have limited hours. If the queue has not cleared by closing, a delay measured in minutes becomes one measured in hours, which makes the delay distribution bimodal rather than merely skewed.

Slack before cutoff	Probability of missing it	Expected delay
15 min	50.6%	364 min
30 min	25.6%	184 min
60 min	6.5%	47 min
90 min	1.7%	12 min
120 min	0.4%	3 min

On a 22-minute mean wait against an overnight penalty, arriving thirty minutes before cutoff carries about 184 minutes of expected delay while arriving two hours before carries about three. The value of early arrival at a border is wildly nonlinear, and no distance-minimizing objective will ever discover that, because arriving early costs miles and saves nothing the optimizer is measuring.

4. Expect the error to be worst on the days that matter

Queue wait rises with utilization rather than in proportion to it. Using the standard relationship, relative wait moves as utilization over one minus utilization.

Lane utilization	Relative expected wait
50%	1.0
70%	2.3
80%	4.0
90%	9.0
95%	19.0

Moving from 50% to 90% utilization multiplies expected wait around nine times. A buffer calibrated on an average day is therefore least adequate on the days volume peaks, which is exactly when the plan is load-bearing. Lane closures compound this, because CBP publishes lanes open alongside wait and a reduction in booths raises utilization on the remainder immediately.

Also Read: Best Truck Routing Software in 2026

5. Model express-lane eligibility as a product, not a status

Trusted-trader and express programs are the main lever for reducing border wait, and eligibility is compound: the carrier must be enrolled, the driver must be enrolled, and the shipment and its documentation must qualify.

Carrier qualified	Driver qualified	Shipment qualified	Loads eligible
95%	90%	85%	72.7%
90%	80%	70%	50.4%
80%	70%	60%	33.6%

At 90%, 80% and 70% the express lane is available to only half the loads. Because eligibility multiplies rather than averages, a program reported as broadly in place can cover a minority of actual movements, and the assignment of eligible driver and tractor to eligible load becomes a routing decision rather than a compliance record.

6. Choose the port as a decision, not a default

Where multiple crossings serve a lane, port selection is a live trade-off between distance, published wait, operating hours, lane availability and eligibility. A slightly longer drive to a quieter port is frequently the cheaper plan once the tail and the cutoff are priced, and that comparison cannot be made by an engine that holds one fixed travel time per crossing.

7. Separate the crossing from the customs decision

The queue is an operational constraint and admissibility is a documentary one. They interact, since incomplete documentation lengthens processing and can trigger secondary inspection, but they are different systems with different owners. Routing should consume the crossing’s timing behavior while the trade management platform owns admissibility, and the interface between them is where most cross-border programs are underscoped.

Where a border differs from every other network element

Property	Road segment	Depot or dock	Port of entry
Delay correlation across vehicles	Low, averages out	Moderate, within site	High, common-mode
Availability	Continuous	Shift hours you control	Published hours you do not control
Capacity	Fixed by geometry	Adjustable by staffing	Set by lanes open, changes without notice
Access differentiation	None	None	Eligibility-gated express lanes
Failure mode	Slower	Queueing	Missed cutoff, overnight
Data availability	Traffic feeds	Internal	Published live per port

The last row is the useful one. Border timing is one of the few genuinely public operational datasets in freight, and the gap between its availability and its use in planning is the opportunity.

Also Read: Route Optimization for Enterprise Logistics: 2026 Guide

Five criteria for evaluating cross-border routing capability

1. Does it consume live and historical port wait data? Ask specifically whether published per-port wait and lanes-open data feed the plan, or whether the crossing carries a static allowance.

2. Can it hold a port-specific reliability buffer rather than a flat one? Buffers should scale with each port’s own distribution. One global border allowance is wrong at every port simultaneously.

3. Does it model operating hours as a hard cutoff with an overnight consequence? The plan should treat a missed cutoff as a discontinuity, not as additional minutes, and should value early arrival accordingly.

4. Can it assign eligible driver and tractor to eligible load? Express-lane benefit is only realized when all three qualifications line up on the same movement. Ask how that pairing is enforced at assignment.

5. Can it re-select the port mid-plan? If wait at the intended crossing spikes after dispatch, the platform should be able to evaluate a diversion against the remaining sequence rather than committing to the original choice.

Also Read: Top 10 Logistics Management Software Companies in Mexico

What this looks like in multi-country deployments

A Fortune 50 parcel and freight enterprise operating across a 120-country network centralized dispatch across 51 sites and a 4,500-strong pool split between captive and third-party capacity, lifting weekly execution rate from 75% to 92% and surfacing more than $14 million in unused capacity. The mechanism was removing the boundary between sites so one allocation decision could see all the capacity. A border is the same class of problem expressed geographically: a shared, contended resource that no single site’s plan can see the whole of.

A leading North American retailer running multi-hundred stores across ocean, rail and road replaced six legacy systems with a single orchestration layer, reaching over 99% on-time store delivery, route compliance above 95%, exceptions resolved in under two hours and break-even in year one against savings above $1 million. Multimodal North American networks are where crossings sit, and exception resolution time is the metric that matters when a crossing behaves unexpectedly, because the useful response is a re-plan rather than a report.

Neither deployment is presented as a border-crossing implementation. What they show is the mechanism: contended shared resources modeled inside the plan rather than assumed around it.

Four mistakes in cross-border route planning

Carrying a static border allowance. A single fixed number per crossing is wrong in both directions, generous on quiet days and badly short on busy ones, and it never triggers a diversion.

Planning on the average published wait. It fails roughly 37% of the time. Reliability at a queue is bought in multiples of the mean.

Treating the cutoff as a soft target. Missing it is not a longer delay, it is a different category of delay. Distance-minimizing objectives will always shave the slack that protects against it.

Assuming express-lane enrolment equals express-lane access. Eligibility is compound across carrier, driver and shipment, so a program described as in place can apply to half the loads.

Also Read: Route Optimization: The Complete 2026 Guide

How Locus handles contended, time-bounded network resources

Locus, the world’s first Decision-Intelligent, Agentic TMS, plans against constraints rather than around them, and a port of entry is a constraint of a specific type: shared, time-bounded and eligibility-gated. The route planning system sequences against more than 250 real-world operating constraints, including time windows, driver hours, vehicle and driver qualification, access restrictions and site-level availability, which is the machinery a crossing needs, because the crossing is a window plus a qualification plus a queue rather than a distance.

Because plans are produced in roughly two minutes and re-optimized continuously as execution events arrive, a wait that spikes after dispatch can be answered with a re-plan or a diversion rather than with an arrival that misses the cutoff. That responsiveness is the practical defense against a queue, since the useful decision is almost always made after departure.

The governance layer decides what happens without a dispatcher. Autonomy Levels run per agent and per domain, so a modest re-sequence can be absorbed autonomously while a port diversion, which carries commercial and customs consequences, routes to a person. Explainability and Traceability record the trigger, context, reasoning, action and outcome, which is what allows a missed cutoff to be reconstructed rather than argued about.

One boundary is worth stating. Locus does not perform customs admissibility determination, classify goods or file entries, and no routing platform should be assumed to. Cross-border deployments need the trade management or brokerage system that does, alongside a live feed of port wait and lane data, and both should be named requirements during implementation rather than assumed. What Locus contributes is the planning and execution layer: the crossing as a modeled constraint, allocation across owned fleet, contracted transporters and a network of more than 1,000 carriers when the qualified asset is not in-house, and a decision record for what was chosen and why.

Locus supports more than 360 enterprise customers across 30-plus countries, with over 1.5 billion deliveries optimized, more than $320 million in documented client logistics savings and 99.99% uptime. It has been recognized by Gartner for seven consecutive years, featured in the 2026 Hype Cycle for Supply Chain Execution and Logistics Technologies, named a Leader in TMS by QKS Group (SPARK Matrix), and ranked #1 in Route Planning on G2’s 2026 Best Software Awards.

In October 2025, Ingka Investments, the investment arm of Ingka Group, the world’s largest IKEA retailer, acquired Locus. Locus continues to operate independently.

So how should cross-border route optimization treat a North American port of entry? As a queue with published hours, published waits and eligibility-gated lanes, not as a road segment with a travel time. The consequences are specific and measurable: border delay is common-mode so it does not average out, making it roughly 4.5 times more disruptive than an equivalent road delay across twenty trucks; planning on the average published wait fails about 37% of the time and 95% reliability costs three times the mean; a missed operating-hour cutoff converts minutes into an overnight, so arriving thirty minutes early carries 184 minutes of expected delay against three minutes for a two-hour margin; and express-lane eligibility multiplies across carrier, driver and shipment, so half the loads can be excluded. 

Locus models the crossing as a window, a qualification and a contended resource inside a 250-plus constraint set, re-optimizes in roughly two minutes so a diversion is still available after dispatch, and routes the port-change decision to a person under per-domain autonomy levels while absorbing the routine re-sequences. Pair it with your trade management platform and a live port-wait feed. Request a Locus route planning assessment to price your own border buffers.

FrAQs

Why do cross-border route plans miss more often than domestic ones? Because a border crossing is a queue, not a road segment, and most routing engines model it as the latter. A queue has a right-skewed wait, published operating hours, variable lanes open and eligibility-gated express access. A static travel time captures none of that, and the error grows as volume rises.

How much buffer should a border crossing get? More than the average wait. Under a standard queueing approximation, planning on the mean succeeds about 63% of the time, 90% reliability needs roughly 2.3 times the mean and 95% needs about 3 times. Against a 22-minute published average that is a 66-minute buffer for 95% reliability. Calibrate against your own port history rather than a single global allowance.

Why does a border delay hurt more than an equivalent traffic delay? Because it is common-mode. Traffic delays hit different vehicles at different times and average out across a fleet, so the fleet-mean spread shrinks with the square root of vehicle count. Every truck in a border queue is delayed together, so the variance does not shrink. At twenty trucks crossing in the same window the border delay is around 4.5 times more disruptive.

How valuable is arriving early at a crossing? Far more valuable than distance-based objectives suggest, and nonlinearly so. On a 22-minute mean wait with an overnight penalty for missing the cutoff, arriving thirty minutes before carries roughly 184 minutes of expected delay while arriving two hours before carries about three. The protection is against the cliff, not against the average.

Does enrolment in a trusted-trader program fix border wait? Only for loads where every qualification lines up. Eligibility is a product of carrier, driver and shipment qualification, so at 90%, 80% and 70% respectively only about half of loads qualify. Realizing the benefit means assigning an eligible driver and tractor to an eligible load, which is an assignment decision rather than a compliance status.

Should port selection be fixed per lane? No, where alternatives exist. Port choice is a live trade-off between distance, published wait, lanes open, operating hours and eligibility, and a slightly longer drive to a quieter crossing is often cheaper once the tail and the cutoff are priced. That comparison requires the crossing to carry live timing data rather than one fixed value.