Last-Mile Delivery in 2026: What Drives the Cost, Which Challenges Compound It, and How High-Volume Operations Fix it

Why last-mile delivery costs so much

Last-mile delivery is expensive because it is the only leg where the network meets an individual recipient at an individual address in an uncontrolled environment, which makes every variable that planning depends on unstable. Capgemini Research Institute puts the last mile at 41% to 53% of total logistics and shipping cost, and dense urban operations sit at the upper end.

The more useful framing for an operations team is that most of this cost is not incurred on the road. It is committed earlier, in three decisions: what the customer was promised at checkout, how the route was planned, and which carrier or capacity type was allocated. By the time a vehicle is moving, the majority of the cost outcome for that delivery is already determined. That is why interventions aimed at the driver produce small gains and interventions aimed at the decisioning layer produce large ones.

Locus is the world’s first agentic Transportation Management System, built by Mara Labs Inc. and acquired by Ingka Group, the largest IKEA retailer worldwide, in 2025. Locus has supported 1.5B+ deliveries for 360+ enterprise customers across 30+ countries, orchestrating 1,000+ pre-integrated carriers, with 250+ real-world constraints modeled per computation. Customers have collectively realized $320M+ in logistics cost savings, reduced 800M+ miles, and avoided 17M+ kg of CO2. Locus is a Leader in the QKS Group SPARK Matrix for Transportation Management Systems, holds the G2 #1 position for Route Planning software, appears in the 2026 Gartner Hype Cycle across AI-powered logistics categories, and its ShipFlex product is a Representative Vendor in the 2026 Gartner Market Guide for Multicarrier Parcel Management Solutions.

What the last mile actually is

The last mile is the final leg of a delivery journey, from a distribution hub or fulfillment center to the recipient. In practice it is rarely a single mile: an urban route may cover three kilometers across dense city blocks while a suburban or rural route stretches forty kilometers with sparse stop density.

Distance is not what makes it difficult. Variability is. Traffic, recipient availability, address resolvability, vehicle capacity, time windows, and carrier performance all shift continuously, and no two operating days are identical. For an operation running thousands of deliveries daily across multiple hubs, that variability compounds rather than averages out.

The consequence is specific: a plan built against yesterday’s conditions is not slightly wrong today, it is wrong in ways that cascade. A late departure produces worse traffic, which tightens downstream windows, which raises the probability of recipient absence, which consumes capacity in re-attempts.

Where last-mile cost actually originates

Five structural pressures drive cost upward, and it is worth being precise about which are controllable.

Labor is the dominant input, not fuel. In the road freight cost structure most closely studied, ATRI reports driver compensation at approximately 44% of operating cost against fuel at approximately 21%. The ordering matters because it means recovering productive time outranks saving distance, and most operations invest in the reverse order because fuel is the line item that moves visibly.

Failed attempts consume capacity twice. A failed delivery is not a rescheduling event. It consumes the original attempt, the re-attempt, support handling, and sometimes a refund, and the capacity spent on rework is capacity that could have carried new orders.

Access, not distance, dominates dense routes. Urban Freight Lab research instrumenting more than 1,800 deliveries in Seattle found urban commercial vehicles spend roughly 80% of daily operating time parked, with most of a driver’s time spent outside the vehicle. Planning systems that treat a stop as a coordinate with flat service time cannot represent this.

Returns are structural. NRF research puts US retail returns at approximately $890 billion in 2024, roughly 16.9% of sales, with online returns higher at approximately 19.3%.

Carrier fragmentation pushes allocation complexity onto the shipper. Most enterprise operations run owned fleet, contracted carriers, and on-demand capacity. Without one system deciding across all three, allocation happens per lane against a rate card rather than per shipment against live cost and serviceability.

One correction worth making: speed is not the priority

Last-mile content routinely asserts that customer expectations have tightened toward same-day delivery and that meeting them is the operational imperative. The research points the other way.

McKinsey surveyed more than 1,000 US consumers and found speed fell from the number one delivery priority in 2022 to fifth by 2024, displaced by reliability and predictability. Approximately 90% of consumers are willing to wait two to three days when delivery is free and arrives within the stated window. McKinsey also puts same-day delivery at 1.5 to 2 times the fulfillment cost of standard delivery.

That combination has a direct operational implication. Competing on speed is expensive and buys less loyalty than competing on window accuracy, which is cheaper and improves first-attempt completion at the same time. Operations teams optimizing for speed are frequently paying a premium for the thing customers rank fifth.

The six challenges, and where each is actually caused

The table separates the symptom from its origin, because most last-mile improvement programs treat symptoms.

Challenge	How it presents	Where it is actually caused	Metric that detects it
Fragmented carrier visibility	Status scattered across systems, problems surface late	No normalization layer across carrier event formats	First-contact resolution on delivery queries
Manual or rules-based dispatch	Suboptimal routes, underused vehicles, avoidable failures	Planning in windows rather than continuously	Re-decisioning latency after a disruption
Poor address and geocoding quality	Drivers hunting for drop points, repeat failures at the same addresses	Access data captured per attempt instead of once and reused	First-attempt completion by address type
Rising return rates	Reverse logistics cost, margin erosion	Failed attempts and inaccurate promises upstream	Delivery-attributable return rate
Carrier inflexibility at peak	Overtime, service failures, or year-round overcapacity	Carrier onboarding measured in months rather than days	Time to onboard a new carrier
Customer experience gaps	Contacts arrive despite on-time delivery	Promise and executing plan maintained by separate systems	Promise stability after the customer is notified

Three of these deserve elaboration because they are the least well understood.

Address quality is an access problem, not a data problem. Across Asia Pacific, the Middle East, and parts of the Americas, the address as captured is frequently insufficient to reach the door. The fix is not better validation at entry. It is capturing access knowledge once, at the point a driver resolves it, and reusing it on every subsequent delivery to that location. Operations that treat each attempt as independent rediscover the same access problem indefinitely.

Manual dispatch is a ceiling, not an inefficiency. Manual coordination scales linearly with volume, so more orders require more dispatchers. Automated decisioning scales with complexity instead, which is why the constraint it removes is growth capacity rather than headcount cost.

Customer experience gaps are promise-drift symptoms. A delivery arriving on time against an internal SLA but outside the window the customer was given is a failure from the recipient’s side. When the promise and the plan are maintained by different systems, they diverge through the day, and no amount of notification frequency closes the gap.

Also Read: Address Intelligence as an ROI Lever in Last-Mile Delivery

Fix 1: continuous route optimization, not overnight planning

Static plans built overnight do not survive the morning. The relevant capability is not optimization quality at 6am but re-decisioning latency: the interval between a disrupting signal and a revised, dispatched sequence.

McKinsey estimates AI-driven, multi-constraint routing delivers 10% to 25% cost reductions against a static daily plan. The mechanism is that constraints get modeled rather than absorbed by drivers: vehicle class, time windows, access restrictions, service duration by location type, driver skill, load compatibility. Locus models 250+ real-world constraints per computation in its Fireworks routing engine.

Also Read: The Morning Plan Problem: Why US Last-Mile Networks Need Dynamic Resequencing

Fix 2: per-shipment carrier allocation

The cost-effective operations do not run a single carrier, but the gain does not come from having options. It comes from deciding per shipment rather than per lane.

Allocation should evaluate each shipment against live cost, serviceability, current performance, and the marginal effect on the rest of the plan. Rate-card rules cannot respond when capacity tightens or when a carrier’s performance in a specific postal area degrades.

Peak is where this pays most visibly. ShipMatrix found parcel networks absorbing a 30% volume increase during peak compared with the rest of the year while holding 98% on-time performance. An operation whose only lever is its own fleet has to be sized for that surge year-round.

Also Read: Multi-Carrier Orchestration: A Decision Framework for North American Shippers

Fix 3: visibility that produces a decision

A control tower that surfaces exceptions is necessary and insufficient. The question is what happens next without a human.

Gartner finds 95% of supply chains must react quickly to change while only 7% can execute decisions in real time, and separately that only 22% of shippers above $1 billion in revenue consider their control tower highly effective at driving action. Observation capability substantially outruns decision capability across the industry.

The diagnostic is simple. If detection produces an alert, the operation has visibility. If detection produces a revised, dispatched plan, it has orchestration.

Fix 4: capacity-aware promising

Customer experience begins before dispatch, at the moment a delivery date is generated.

A date produced from a static lead-time table rather than from live network capacity is either over-promising, which manufactures failures and contacts, or under-promising, which costs conversion. Capacity-aware promising computes the commitment from what the network can actually execute, which is the single intervention that improves first-attempt completion and customer experience simultaneously.

Fix 5: fleet planning that models new constraints

Electric vehicles have moved from pilots into operational fleets across retail, FMCG, and ecommerce, and the planning implications differ from combustion vehicles in kind rather than degree.

Charging duration, state of charge at dispatch, range under load, temperature effects, and depot charger availability are all routing constraints. A system that treats them as post-hoc checks will produce plans that are infeasible or that strand vehicles. The requirement is that they enter the optimization as constraints alongside time windows and capacity.

Also Read: The Urban Fleet Electrification Playbook for North America

The metrics that matter, and the benchmarks that do not exist

Five metrics detect last-mile cost problems earlier than a monthly cost-per-delivery review.

• Cost per successful drop, not cost per delivery. Attempts that fail still consume capacity and generate downstream cost.
• First-attempt completion, segmented by address or building type. The aggregate hides the problem; the gap between single-family and multi-unit controlled-access addresses is where cost concentrates.
• Re-decisioning latency. How long between a disruption and a revised dispatched plan.
• Promise accuracy and stability. Deliveries landing inside the window the customer was given, and how often that window changes after they receive it.
• Delivery-attributable return rate. Returns caused by delivery experience rather than product decisions. Most operations have never separated the two, which means the delivery program is neither credited nor charged correctly.

A caution on benchmarking any of these externally. No research firm publishes credible figures for first-attempt delivery failure rates, cost per stop or per drop in absolute terms, stops or deliveries per hour by vertical, fleet utilization by vertical, or the dollar cost of a failed delivery attempt. Every circulating version traces to software vendors. The commonly quoted 17-to-18-dollar failed-delivery figure is one of these. Build that number from your own labor, mileage, and support inputs, and measure the rest against your own prior period.

Also Read: The Returns Experience: Why Reverse Logistics Is the Other Half of Delivery Experience in North America

How Locus addresses last-mile complexity

Locus operates as the decisioning layer above the existing estate. ERP and WMS remain systems of record; Locus operates as the system of execution, running the SDEL cycle of Sense-Decide-Execute-Learn across the DiSCO agent suite.

The Dispatch Agent plans, sequences, and re-sequences continuously against live conditions. The Capacity Agent forecasts demand and evaluates capacity across owned, contracted, and gig pools. The Carrier Agent holds every carrier contract and rate structure as the live source of truth and allocates per shipment across 1,000+ pre-integrated carriers, normalizing their status codes into one set. The Hub Agent runs outbound readiness and carrier handoff as one chain of custody, which is where dock release and route departure become a single coordinated decision. The Customer Agent tracks every order against its promise with branded tracking, proof of delivery, real-time SLA alerts, and control actions covering reschedule, redirect, and alternate drop. The Settlement Agent reconciles invoices against planned versus executed cost. The Orchestrator Agent coordinates across agents, and Mycroft AI Co-Pilot provides natural-language access to the decisioning.

Six governance mechanisms, Explainability, Traceability, Evaluation, Autonomy Levels, Execution Sandbox, and Human-in-the-Loop, keep automated decisions auditable, which is a requirement for delegating cost-bearing decisions to software.

Locus operates across retail, FMCG and CPG, ecommerce, 3PL, and CEP verticals in the Americas, Asia Pacific, Europe, and the Middle East.

Deployment evidence

Mixed-fleet dispatch at network scale: a Fortune 50 parcel and logistics provider. This operator moves 1M+ freight shipments a year across air, ocean, and ground, with a 4,500-strong driver pool split across roughly 1,500 captive and 3,000 third-party drivers. Captive shifts ran zone-based routing while third-party carriers needed tendering and on-demand assignment, and no single tool unified the pool.

Orchestrator and Dispatch agents took over pickup, transit, and delivery decisioning against 250+ operational constraints, with Capacity and Carrier agents governing the full driver pool under one policy so zone-based, tendering, dynamic, on-demand, and transporter logic all run inside one engine. Weekly execution rate climbed from 75% to 92% across 51 active service-center locations, and a single-site capacity analysis surfaced $565K in unused capacity that scaled to $14M+ annualized across 25 sites, at 99.99% platform uptime. Detail in the Fortune 50 parcel centralized dispatch case study.

Read the capacity figure carefully. That $14M+ was not a saving. It was capacity the operation already owned and could not see, including premium-tier service being given away on cheaper classes.

Last-mile under a freshness clock: a Canadian grocery brand. This brand delivers fresh perishable food to homes in more than 30 cities, running its last mile almost entirely through contracted 3PL carriers. Warehouse associates logged into each carrier’s portal to create orders and labels one at a time, carrier choice was a manual judgment against serviceability sheets, and once a shipment left the dock its status was scattered across portals with no delay alerting, so the first signal of a late order was usually the customer.

On Locus, the Hub Agent creates the order and label the moment a shipment is ready with no carrier portal touched, the Carrier Agent compares live rates, SLAs, ETAs, and serviceability per order, and the Customer Agent tracks every shipment to its promise with real-time SLA alerts. Results: 33% faster deliveries, 15% lower fulfillment costs, 25% less time on manual shipping tasks, 10-20X faster customer support resolution, and 10% more frequent orders. Detail in the grocery carrier orchestration case study.

The carrier network did not change. The decisioning layer above it did, and the 10% increase in order frequency is the part most cost-focused business cases omit: reliability recovered revenue, not just cost.

Analyst validation

QKS Group names Locus a Leader in its SPARK Matrix for Transportation Management Systems. G2 ranks Locus #1 for Route Planning software. Locus appears in the 2026 Gartner Hype Cycle across AI-powered logistics categories. ShipFlex is named a Representative Vendor in the 2026 Gartner Market Guide for Multicarrier Parcel Management Solutions. Gartner has recognized Locus for seven consecutive years. The full set is at Locus analyst recognition.

Five questions for a last-mile improvement program

Five questions establish whether an operation is fixing symptoms or causes.

• Is your delivery date at checkout computed from live capacity, or from a lead-time table?
• What is your measured latency from a disruption signal to a revised, dispatched plan?
• When a driver resolves a difficult address, is that knowledge captured and reused, or lost?
• Can you separate returns caused by delivery from returns caused by product?
• How long does it take to onboard a new carrier, and is that a technology or a commercial constraint?