Route Optimization When Capacity is Opt-In: Why the Route Nobody Takes is Never Random

06:12 and four routes have no driver

A US last-mile operation runs a mixed fleet. Overnight the optimizer builds 84 routes from the day’s orders. Fifty-three go to employed drivers as assignments. Thirty-one are released as offers to a flex pool.

By 06:12, twenty-seven have been claimed. Four have not.

Those four are not four random routes out of thirty-one. They are the two longest, the one covering a low-density outer zone, and the one with a 90-minute window in a district where parking is difficult. Anyone who has worked a dispatch desk could have named them the night before without looking at the acceptance data.

That predictability is the whole point. The pattern is not noise in the acceptance process. It is a structural consequence of running an optimizer that assigns against a workforce that chooses, and it is worth understanding as a design problem rather than as a staffing shortfall to be escalated each morning.

Also Read: Driver Onboarding and Scheduling Software Guide 2026

Assignment and offer are different problems

A vehicle routing engine solves an assignment problem. Given a set of work, a set of resources, and a set of constraints, produce the best matching of work to resources. That formulation is correct when the resources are yours to direct: employed drivers on shift, owned vehicles, contracted capacity with an obligation to perform.

Treating owned fleet, 3PL capacity, and gig riders as one allocatable pool is the right first move, and it is what modern platforms do. It removes the parallel-systems problem and lets one decision consider all available capacity.

The second move is the one that gets skipped. One of those pools answers back.

When a route is offered rather than assigned, the plan’s output is a proposal. Every route can be individually feasible, correctly sequenced, and fully compliant, and the plan can still fail, because feasibility is a property of the route and acceptance is a property of the person looking at it. The optimizer has no representation of the second.

So the failure mode is new and specific: a plan that is valid and unstaffed. Nothing in the solve is wrong. The plan simply made a claim about resources it does not control.

Adverse selection is the mechanism

Here is why the unclaimed routes are predictable, and it is a consequence of two parties each doing their job well.

The optimizer concentrates efficiency. That is what optimization means. Given a day’s orders, it will cluster the dense work into tight, short, well-sequenced routes, and it will push the awkward geography, the isolated stops, and the difficult windows into whatever routes are left. It does not distribute difficulty evenly, because distributing difficulty evenly is less efficient. A good solve produces a set of routes with high variance in attractiveness.

Drivers evaluate offers on effective earnings per hour. A flex driver comparing two routes will take the one that pays similarly for less time, less distance, and less hassle. This is entirely rational and requires no coordination.

Put those together and the result follows mechanically. The most attractive routes are claimed first. What remains is not the average route, it is the tail of the distribution the optimizer created.

The uncomfortable corollary is that a better optimizer makes this worse. A weaker engine spreads difficulty more evenly and produces a flatter distribution of route attractiveness, which is less efficient in total and easier to staff. A stronger engine concentrates the good work, which raises plan efficiency and sharpens the gap between the routes that go instantly and the routes that go nowhere.

That is not an argument against optimization. It is an argument that plan efficiency and staffing certainty are competing objectives, and that most operations optimize the first while assuming the second.

Also Read: The Three-Workforce Fleet Reality: Owned, 3PL, and Gig Drivers

Why the orphan costs more than it looks

An unclaimed route gets covered one of three ways, and none of the costs appears in the routing report.

Premium spot capacity. A third-party operator takes it at short notice, at short-notice pricing. Immediate, expensive, and the cost lands in carrier spend rather than in planning.

Owned-driver overtime. An employed driver absorbs it after their own route, or a supervisor runs it. The cost lands in labor, and it also consumes the goodwill of the people you least want to burn.

Deferral. The work moves to tomorrow. The cost lands in service, in customer contact, and often in a failed promise that generates its own recovery cost.

There is a compounding problem underneath all three. The orphaned routes are disproportionately the remote, the low-density, and the awkward, which means they are also disproportionately the deliveries with the highest failure risk and the longest recovery time. So the work that ends up covered late, by whoever is available, is the work that could least afford it.

This is why orphan rate deserves to be a routing metric rather than a dispatch anecdote. The routing function created the distribution. Currently it is charged for none of the consequence.

Four levers that change acceptance

Route composition. Deliberately blend attractive and difficult work rather than allowing the solve to purify routes. A route with eight dense drops and two awkward ones will be taken; a route with ten awkward ones will not. This is the most powerful lever and the least used, because it costs measurable plan efficiency to buy unmeasured staffing certainty, and only one of those two numbers is on the report.

Offer sequencing. Whether offers go out simultaneously or in waves, and who sees what first. Tenure-based priority rewards retention and accelerates the separation problem. Releasing difficult routes first, with the attractive ones held back, changes the calculus entirely. Both are policy choices, and most operations have never made one deliberately.

Price. A premium on the tail. The cheapest instrument to implement and the easiest to overuse, because once drivers learn that waiting produces a premium, waiting becomes the strategy. If price is used, it needs a governed ceiling and it needs to be attached to route characteristics rather than to time elapsed.

Information. What the driver actually sees when deciding: estimated earnings, realistic duration including dwell, access notes, parking difficulty. Under-disclosure raises initial acceptance and produces the worst outcome in this whole system, which is acceptance followed by abandonment. A declined route is a problem at 05:30 when you can still act. An abandoned route is a problem at 11:00 when you cannot.

The fourth lever is where most operations have the largest immediate gain, because it costs nothing and improves the quality of the acceptance rather than the quantity.

Also Read: Driver Routing and Scheduling: What Enterprises Get Wrong

Three ways operations handle opt-in capacity

The row that matters is the second-to-last one, because it is the trade nobody writes down. Acceptance-aware planning produces a marginally worse plan on the metric the routing team reports and a better outcome on the metric the business pays. Until orphan recovery cost is attributed back to the plan, the incentive runs the wrong way.

Also Read: Route Optimization Software: Dynamic Re-Routing 2026

What to measure

Orphan rate, and orphan rate by route characteristic. Not just how many routes went unclaimed, but what they had in common: length, stop density, window tightness, zone. The second half is the finding. The first half is a headcount problem.

Acceptance rate by offer ordinal. Whether the first offer released is claimed at a different rate than the tenth. This shows how steep your attractiveness distribution is.

Time to full staffing. Minutes from offer release to last route claimed. A long tail here is the same signal as an orphan, arriving slightly earlier.

Abandonment rate after acceptance. Routes claimed and then dropped, and how far into the day. This measures your information quality, and it is the most expensive failure in the system.

Orphan recovery cost per route, by recovery mode. Premium spot, overtime, or deferral, each costed. This is the number that makes the composition trade arguable rather than theoretical.

The equity tension worth naming

Blending routes to improve acceptance has a second effect that deserves stating rather than glossing.

If difficult work is spread across routes rather than concentrated, earnings differences between drivers narrow. That is good for retention and for fairness, and it is also a reduction in the upside available to the fastest drivers, who will notice.

If instead the tail is priced, you are paying more for worse work, which is defensible and arguably correct, and it establishes that difficult routes carry a premium, which changes behavior in ways that persist beyond the day.

Neither is free and neither is obviously right. What is clearly wrong is the default: concentrate the difficulty, offer it at standard rates, and treat the resulting shortfall as a supply problem.

Also Read: Best Driver Management Software for Delivery Fleets 2026

How Locus fits

Locus, the world’s first Decision-Intelligent, Agentic TMS, allocates work across owned fleet, contracted 3PL capacity, and gig riders in one decision rather than in parallel systems, which is the prerequisite for treating acceptance as anything other than a downstream surprise. Within its DiSCO framework, the Digital Supply Chain Officer, the Dispatch agent plans and re-sequences against a model of more than 250 real-world constraints, and the Capacity agent holds the full roster with live state on availability, remaining hours, and qualifications across owned, contracted, and on-demand pools.

Two properties are directly relevant to the levers above. Because route composition is expressed through constraints rather than emerging only from a cost objective, requirements such as a minimum assignment-quality floor or a cap on consecutive difficult routes can be stated as constraints alongside cost, which is what makes deliberate blending an option rather than a manual override. And because the Capacity agent holds allocation as state across a period rather than recomputing per route, cumulative distribution of good and difficult work is visible to the system, which is the prerequisite for both the composition lever and the equity tension described above.

Worth being straight about the frontier. Modeling a driver’s probability of accepting a specific offer, and shaping the plan against that prediction, is the harder half of this problem and it is where the category as a whole is still early. What a platform can do today is unify the pool, express composition as a constraint, hold cumulative allocation as state, and give the driver enough information at the point of decision to prevent abandonment. Those four are available now and most operations use none of them.

Locus has processed more than 1.5 billion deliveries for 360-plus enterprise customers across 30-plus countries at 99.99% uptime. It is 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. Further analyst recognition is published in full.

Two deployments show mixed-pool allocation at scale.

A Fortune 50 parcel and logistics provider centralized dispatch across 51 sites in a 120-country network against a driver pool of 4,500 split between captive and third-party capacity. That split is the condition this piece is about: two pools with different obligations to perform, allocated from one plan. Weekly execution rate rose from 75% to 92% at 99.99% uptime, and more than $14 million in contracted capacity that local practice had never used was surfaced, including $565,000 at a single site. Unused contracted capacity is the mirror image of the orphan problem, and both come from allocating without seeing the whole pool.

A global FMCG operation across ten countries runs more than 1,000 distributors and a rider pool exceeding 5,000, reaching over 1.8 million retail outlets, with more than 12,000 trips saved per month at 3X ROI. At that rider count, route attractiveness is not a soft consideration. It determines whether a plan built at night is staffed in the morning.

Request a Locus route acceptance assessment to measure your orphan rate by route characteristic, cost your current orphan recovery by mode, and model what deliberate route blending would cost in plan efficiency against what it would save in recovery.

Name the four routes before you release the offers

One exercise, tonight, at no cost.

Before offers go out, have the dispatch lead write down which routes they expect to go unclaimed. Then compare the list against what actually happens by 06:30.

If the prediction is good, and it usually is, you have established two things. Orphaning is a property of your plan rather than of your labor market. And the information needed to prevent it existed before the plan was released, which means the plan could have been different.

Then cost last month’s orphan recoveries by mode and put that number next to the plan efficiency you would give up by blending. That comparison is the whole decision, and almost nobody has both halves of it on the same page.