The Dispatch Cost of Over-Automation: Where Full Autonomy Erodes Peak-Season Margin

AI dispatch assigns orders to vehicles and drivers against live constraints, and logistics automation is usually sold on a single axis: the more decisions run unattended, the better the operation. That axis is wrong at the top end. A fully autonomous dispatch engine does not fail by being slower than a human. It fails by being confidently wrong faster than anyone can catch it, and the cost of that failure scales with the same volume that makes automation attractive in the first place. At 200 decisions an hour a fifteen-minute detection delay puts 50 decisions on a bad premise. At the 800 an hour a peak surge produces, the same delay puts 200 on it.

The right frame is not how much autonomy you can achieve. It is which decision classes earn it, computed from how reversible each decision is and how many more get made before anyone notices the first one was wrong.

Why full autonomy is the wrong target for AI dispatch

The failure mode here is old and well documented, which is worth knowing before treating it as an artifact of recent AI. In the canonical human factors treatment, Parasuraman and Riley classified human interaction with automation as use, misuse, disuse and abuse, defining misuse as over-reliance that produces failures of monitoring and decision bias. That paper is from 1997. The mechanism it describes, where an operator supervising a system that is usually right stops catching the cases where it is not, is the exact risk a dispatch operation takes on when it moves a decision class to full autonomy and keeps the same monitoring posture.

What has changed is the decision rate. A dispatcher reviewing plans one at a time was a bottleneck and also a filter. Remove the bottleneck and you remove the filter, and the volume that flows through is the volume that determines the damage.

Peak makes this materially worse for two independent reasons, and it is the combination that erodes margin.

The first is distribution shift. Models are fitted on history, and peak is precisely the period least represented in it. The standard engineering account of production machine learning names changes in the external world among the risk factors that make deployed models expensive to maintain, and a demand surge is that risk arriving on a schedule. US parcel volume illustrates the scale: carriers handled an estimated 2.3 billion packages across the 2025 holiday season, up 5% year on year, and the intra-season concentration is far sharper than that annual figure suggests.

The second is that the cost function changes shape. Peak surcharges are not a flat uplift. FedEx ties residential delivery surcharges to shipping volume measured against a June baseline, and going 200% over that baseline can triple or quadruple per-package fees, reaching $7.50 to $8.75 per package at the height of the season. An allocation engine calibrated on linear per-unit costs will cheerfully push volume across a threshold that multiplies the rate, because from inside the model nothing about that movement looks expensive. This is the mechanism behind the quiet part of margin erosion: no single decision is obviously wrong, and the invoice arrives in January.

Ground cost is rising underneath all of it. ATRI’s 2026 report puts the industry-average cost of operating a truck at $2.336 per mile in 2025, a record for the series and 3.4% above the prior year, with $1.854 of that excluding fuel.

How to set autonomy levels for peak

1. Sort decision classes by reversibility, not importance

The instinct is to gate the decisions that look big. The correct sort is by whether a decision can be undone in the next planning cycle at no cost. Re-sequencing two stops on a route that has not departed is fully reversible. Dispatching a vehicle 40 miles, committing to a delivery window with a customer, tendering a load to a carrier at a surcharge tier, or releasing a driver for the day are not. Reversibility, not size, determines whether an error is a correction or a loss.

2. Compute blast radius for each class

Blast radius is the number of decisions made on the same wrong premise before somebody notices. It is decision rate multiplied by detection time, and it is where volume enters the calculation.

The right-hand column is the peak-season column, and it is the argument. Nothing about the model changed between the two columns. Only volume did, and the exposure quadrupled.

3. Set the level from blast radius against review capacity

Automation throughput scales with volume. Review capacity does not, because it is headcount, and headcount is set weeks before peak.

A configuration that is comfortable at baseline diverges at peak, and it diverges for structural reasons rather than because anything broke. The operation that leaves autonomy at its baseline setting through surge is not holding a position, it is accumulating an unreviewed queue at 34 items an hour.

4. Stop gating autonomy on model confidence alone

The common design is to auto-execute above a confidence threshold and escalate below it. That design has an unfortunate property at peak, because confidence and accuracy decouple exactly when conditions leave the training distribution.

The figures are illustrative, but the direction is not in doubt. A confidence gate does not tighten under stress, it stays where it is while the accuracy behind it falls, which means the system auto-executes the same proportion of decisions at a materially higher error rate. Confidence is a useful input and a poor sole gate. Pair it with a condition test, so that autonomy steps down when volume, mix or geography moves outside the range the model was fitted on.

5. Price selective review before assuming you cannot afford it

The standard objection to human-in-the-loop dispatch is throughput. The arithmetic does not support it, provided review is selective rather than blanket. At 40 seconds per review:

Reviewing everything at peak requires nearly nine full-time dispatchers and is correctly rejected. Reviewing the 8% of decisions that are irreversible and consequential requires 0.71 of one. That gap is the whole case for autonomy levels per decision class rather than a single global setting: the expensive option was never selective review, it was the absence of a way to select.

6. Rehearse the escalation path at surge volume

An escalation path that works at 200 decisions an hour has not been tested. Run the surge scenario before the season with the volume dialed up, and measure the thing that actually fails first, which is rarely the engine. It is the queue: whether escalations reach a named person, whether that person has the authority and the context to override, and whether the alert rate stays low enough that they keep reading it. An escalation channel producing 40 items an hour will be ignored by the second day, which is Parasuraman and Riley’s disuse failure arriving on schedule.

What each autonomy level should be used for

The pattern worth taking from the table is that peak should move autonomy down for some classes and leave it untouched for others. A blanket reduction gives up the throughput that made logistics automation worth buying, and a blanket hold gives up the margin.

Five criteria for evaluating autonomy in an AI dispatch platform

Autonomy configurable per decision class, not per user or per site. A global setting cannot express the reversibility distinction, which means it forces the operation to choose the wrong level for most classes.

A condition test independent of model confidence. Ask what causes the system to step autonomy down on its own. If the only mechanism is a confidence score, autonomy will not respond to distribution shift.

Detection time as a reported number. Blast radius is unknowable without it. Ask how long a systematically wrong decision typically runs before the platform surfaces it, and how that is measured.

Escalation volume controls. Ask what the platform does when escalations exceed review capacity, because a queue with no ceiling becomes a queue nobody reads.

Cost-model configurability for surcharge tiers. If the optimizer cannot represent a threshold that multiplies the rate, it cannot avoid crossing one. Ask specifically how non-linear carrier pricing is expressed.

What this looks like in enterprise deployments

A leading North American retailer running multimodal logistics automation across several hundred stores replaced six legacy systems, reduced manual dispatch by more than 80% and resolves exceptions in under two hours, alongside 99%-plus on-time store delivery and 95%-plus route compliance. The two figures belong together. More than 80% of dispatch running unattended is a high autonomy posture, and a sub-two-hour exception resolution is what makes it safe, because the blast radius on a store delivery is bounded by a window measured in hours rather than minutes. High autonomy was affordable there because detection was fast relative to the consequence.

A Fortune 50 parcel operation running centralized dispatch across a 120-country network and 51 sites lifted weekly execution adherence from 75% to 92% and surfaced more than $14 million of unused capacity, including $565,000 at a single site. Adherence is the relevant metric for this argument. It measures whether execution matched the plan, which is the signal that tells you an autonomous decision went wrong while there is still time to act, and moving it 17 points is what shortens detection time across the network.

Four mistakes operations make on dispatch autonomy

Treating autonomy as a maturity score. L3 everywhere is presented as the destination and is not. The destination is the right level per decision class, and for irreversible classes that is permanently L1 regardless of how good the model becomes.

Leaving autonomy settings untouched through peak. The setting did not change, so it feels like continuity. What changed is decision rate, which is the multiplier on blast radius, so holding the setting is an increase in exposure rather than a neutral act.

Assuming human-in-the-loop costs throughput. Blanket review does, at nearly nine FTE for an 800-per-hour operation. Selective review of the 8% that is irreversible costs 0.71 FTE, and conflating the two is what leads teams to reject the affordable option.

Measuring automation rate as a success metric. The share of decisions running unattended tells you nothing about whether the right ones are. A dashboard reporting 94% automation and no blast radius figure is reporting the input while omitting the risk.

How Locus approaches dispatch autonomy and governance

Locus, the world’s first Decision-Intelligent, Agentic TMS, runs more than 12 million automated decisions a day, and the governance layer exists because that volume is only safe when the authority behind each decision class is explicit. Autonomy Levels run per agent and per domain: L1 requires human approval before the system acts, L2 acts within defined guardrails, and L3 operates autonomously within high-confidence thresholds. That per-domain granularity is what makes the reversibility sort above implementable, rather than forcing one setting across an operation where a stop re-sequence and a carrier tender carry completely different consequences.

Five further mechanisms complete the framework. Explainability and Traceability record the trigger, context, reasoning, action and outcome for each decision, which is what turns detection time into a measurable quantity instead of an anecdote. Evaluation makes model performance observable, which is the input a condition test needs. The Execution Sandbox allows a policy or autonomy change to be tested before it reaches live plans, which is how a peak configuration gets rehearsed rather than discovered. Human Review provides the escalation pathway itself. Because the route planning system re-optimizes in roughly two minutes against more than 250 real-world operating constraints, a detected error can often be answered by re-planning rather than by unwinding, which is what keeps a lower autonomy level from costing throughput.

Two boundaries belong here. Locus does not decide your autonomy policy, and it should not. Which decision classes are irreversible in your operation depends on your contracts, your customer promises and your surcharge exposure, and mapping that is a named implementation task rather than a default. Second, no governance layer removes the need for review capacity. If escalations exceed the people available to read them, the correct response is to narrow what escalates rather than to raise the threshold and call it autonomy.

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 does full autonomy cost money in AI dispatch at peak? It can, and the mechanism is arithmetic rather than opinion. Blast radius is decision rate times detection time, so a fifteen-minute delay that puts 50 decisions on a wrong premise at 200 an hour puts 200 on it at the 800 an hour a surge produces, with nothing about the model having changed. Peak also decouples confidence from accuracy, so a confidence-gated threshold auto-executes the same share of decisions at a higher error rate, while errors accrue at 40 an hour against a fixed review capacity of six. And the peak cost function is non-linear, since surcharges tied to a volume baseline can triple or quadruple per-package fees, which an optimizer calibrated on flat costs will cross without registering it. The fix is not less logistics automation. It is autonomy set per decision class by reversibility, since reviewing the irreversible 8% costs 0.71 FTE while reviewing everything costs 8.89. Locus supports this through Autonomy Levels running per agent and per domain, Explainability and Traceability making detection time measurable, Evaluation feeding a condition test, and the Execution Sandbox letting a peak configuration be rehearsed before the season. Request a Locus dispatch assessment to set your own autonomy levels before peak.

Frequently Asked Questions

Is full autonomy the goal for AI dispatch? No. The goal is the correct autonomy level per decision class, determined by reversibility. A stop re-sequence on an undeparted route can run unattended indefinitely, while a carrier tender above a surcharge tier or a customer window commitment should stay gated regardless of how accurate the model becomes, because an error there is a loss rather than a correction.

What is blast radius in dispatch automation? The number of decisions made on the same wrong premise before anyone notices, calculated as decision rate times detection time. At 200 decisions an hour a fifteen-minute detection delay yields 50 affected decisions; at the 800 an hour of a peak surge, the same delay yields 200.

Why does peak season make over-automation worse? Two independent reasons. Peak is out-of-distribution, so error rates rise while model confidence does not, and decision volume rises, which multiplies blast radius. Meanwhile review capacity is fixed headcount, so a configuration that absorbs 4 errors an hour at baseline faces 40 an hour at peak against the same 6-per-hour capacity.

Does human-in-the-loop dispatch slow down throughput? Only if review is blanket. At 40 seconds per review, checking all 800 hourly decisions needs 8.89 FTE and is rightly rejected. Checking the 8% that are irreversible and consequential needs 0.71 FTE. The affordable option gets discarded when the two are conflated.

Should autonomy be gated on model confidence? Not on confidence alone. Confidence and accuracy decouple when conditions leave the training distribution, so a fixed threshold auto-executes the same proportion of decisions at a higher error rate exactly when conditions are hardest. Pair the confidence gate with a condition test on volume, mix and geography.

How do peak surcharges change the automation risk? They make the cost function non-linear. Where carriers tie surcharges to volume against a baseline, exceeding it by 200% can triple or quadruple per-package fees. An optimizer calibrated on flat per-unit costs crosses that threshold without registering an expense, which is why the margin loss is quiet and appears on the invoice rather than in the plan.

What should we change before peak season? Sort decision classes by reversibility, compute blast radius at surge volume rather than baseline, step autonomy down on the irreversible classes, add a condition test independent of confidence, confirm the escalation queue has a ceiling and a named owner, then rehearse the whole configuration at peak volume in a sandbox rather than discovering it live.