SEA Last-Mile Efficiency at Scale: Absorbing E-Commerce Growth Without Eroding Margins in 2026

What last-mile efficiency at scale actually means in Southeast Asia

Last-mile efficiency at scale is the ability to hold cost per successful delivery flat, or reduce it, while order volume rises. It is a different discipline from last-mile efficiency in general, because the constraint is not the cost of today’s deliveries but the slope of the cost curve as volume climbs.

Two operators can post identical cost per delivery this quarter. One will absorb a 30% volume increase at the same unit cost. The other will absorb it at 15% higher unit cost because every incremental order requires an incremental rider, and the marginal rider is less productive than the average one. The difference between them is almost never fleet size. It is whether the decisioning layer can extract more throughput from the same capacity as density rises.

The metric has to be cost per successful drop rather than cost per order. Cost per order counts attempts. Cost per successful drop counts outcomes, and it absorbs failed attempts, re-attempt labour, and exception handling. In high-growth Southeast Asian markets those three categories grow faster than volume does, which is precisely why margins erode during growth phases even when the headline delivery cost looks stable.

Why the standard efficiency playbook does not transfer to Southeast Asia

Four structural conditions separate Southeast Asian last-mile operations from the networks most planning systems were designed around.

Two-wheeler fleets change the optimisation problem, not just the vehicle. A motorbike fleet has different capacity constraints, different access rights, different route feasibility, and different weather exposure than a van fleet. Optimising a two-wheeler network is not a van model with smaller payload assumptions. Stop density per hour can be far higher, which raises the value of good sequencing, while carrying capacity per trip is far lower, which raises the frequency of hub returns and makes replenishment timing a first-order routing variable rather than an afterthought.

Mixed and fluid labour models. Most operators at scale run some combination of employed riders, contracted fleets, and gig capacity, with availability that varies by hour and by day rather than by shift roster. Capacity is therefore not a fixed input to the plan. It is a live variable, and a planning system that treats the rider pool as known at 6am is planning against a number that will be wrong by 10am.

Address and kerbside informality. In many high-density Southeast Asian neighbourhoods, the address as captured is not sufficient to reach the door. Access depends on local knowledge, informal landmarks, gated communities, and buildings without formal delivery provision. This means service time per stop varies far more than in planned-grid networks, and a routing engine that assumes uniform service duration will produce plans that fail in the same neighbourhoods every day.

Geography that forces multi-modal handoffs. Archipelagic and river-delta geographies push volume through sea, air, and ground legs before it reaches a rider. Each transfer is a handoff, and handoffs are where cost accumulates invisibly. McKinsey estimates inefficient logistics handovers account for 13% to 19% of logistics costs. A network with more legs has more seams, and fragmented networks pay for every one of them.

The pressure behind all four is infrastructure growth lagging demand growth. The World Economic Forum projects urban delivery demand rising 78% by 2030 globally, with 36% more delivery vehicles in inner cities and emissions up 60% without effective intervention. In megacities where road capacity is already the binding constraint, that trajectory cannot be absorbed by adding vehicles.

Also Read: The End of the “Captive Fleet Only” Era: Orchestrating Hybrid Last-Mile Capacity in 2026

The scale-margin equation: why adding riders is the wrong first response

When volume rises, the instinctive response is to add capacity. It is also the response that guarantees unit cost rises, for a reason worth stating precisely.

Rider productivity is not uniform. The most productive rider-hours are those spent in dense clusters with tight sequencing and short inter-stop distances. When an operator adds riders to absorb volume without re-optimising the network, the new capacity is allocated at the margin, which means it inherits the least dense work. Average deliveries per rider-hour falls. Cost per drop rises. The operation has bought throughput at declining marginal efficiency.

The alternative is to extract the additional throughput from existing capacity first, which rising density actually makes easier rather than harder. More orders in the same service area means shorter distances between stops, more opportunities to cluster multiple drops per journey leg, and better utilisation of each hub return. Growth improves the raw material of route optimisation. Whether the operation captures that improvement depends entirely on whether its planning system re-optimises continuously or reproduces yesterday’s zone structure at higher volume.

This is where most operators lose the margin. Gartner finds that 95% of supply chains must react quickly to change while only 7% can execute decisions in real time. Nearly everyone can see that volume has shifted. Very few can re-decide the network before the shift has already cost them.

The stakes are set by how much of total cost sits in this leg. Capgemini Research Institute puts last-mile delivery at 41% to 53% of total logistics and shipping cost. In a dense e-commerce operation that share sits at the upper end, which means the slope of the last-mile cost curve substantially determines whether growth is profitable.

Lever 1: integrated dispatch allocation

Allocation is the decision that determines which order goes to which capacity, and it is the single highest-leverage decision in a growing network.

The failure mode is allocation by zone and rule. Orders are assigned to a zone, the zone is assigned to a rider pool, and the pool absorbs whatever arrives. This works at stable volume and degrades predictably as volume grows, because zone boundaries drawn for one volume level are wrong at another and nobody redraws them weekly.

Integrated allocation means every order is evaluated against every eligible capacity option at the moment of assignment: employed riders, contracted fleets, gig capacity, and third-party carriers, scored on cost, current position, remaining capacity, skill or handling requirement, and the marginal effect of the assignment on the rest of the plan. The last clause is the one that separates real allocation from dispatch. Assigning an order to the nearest available rider is often the wrong choice if it breaks the sequencing of a denser cluster that rider was about to serve.

Locus handles this through the Capacity Agent, which evaluates available capacity across owned, contracted and gig pools and forecasts demand to right-size the fleet, and the Carrier Agent, which holds every carrier contract and rate structure as the live source of truth and allocates across 1,000+ pre-integrated carriers on cost, SLA and serviceability.

Also Read: Which Dispatch Decisions Should Your AI Make? A Decision-by-Decision Autonomy Map for 2026

Lever 2: dynamic re-routing

Every plan built before the day starts is wrong by mid-morning. In a high-growth e-commerce operation it is wrong sooner, because same-day and next-day orders enter the system after the plan was built.

The metric to hold vendors to is re-decisioning latency: the interval between a disrupting signal and a revised, dispatched sequence. Static overnight planning with manual intervention produces latency measured in hours, which means the network operates against a stale plan for most of the day it governs. Continuous re-optimisation compresses that to the sense-decide cycle.

The economics are well documented. McKinsey estimates AI-driven, multi-constraint routing delivers 10% to 25% cost reductions against a static daily plan. In a two-wheeler network with high stop density, the recoverable share tends toward the upper end of that range, because sequencing errors compound faster when stops per hour are high.

Dynamic re-routing also does something a static plan cannot: it converts intra-day volume into density rather than into overflow. When a batch of same-day orders arrives at 11am, a static network treats them as a separate wave requiring separate capacity. A continuously re-optimising network folds them into journeys already in progress where geography allows.

Lever 3: rider capacity balancing, and the real source of idle time

Idle rider time is the symptom most operators want to fix. The cause is usually upstream of the rider.

Riders are idle for four reasons: they are waiting at a hub for a consignment that is not ready, they have been allocated work in a low-density pocket, they are travelling between clusters that should not have been split, or the plan assigned them fewer stops than their shift can carry because it was built against a forecast rather than actual volume.

Only the first is a warehouse problem. The other three are planning problems, which is why rider-level interventions such as incentive changes and app nudges produce limited improvement. The rider is not choosing to be idle.

The research-grade framing for this comes from workforce planning rather than logistics. McKinsey finds that static planning models can leave as much as 60% of operating hours either understaffed or overstaffed. Fixed plans misallocate by construction, because the conditions they were built against have already moved by the time the shift starts. That is the structural source of idle time, and it is not fixable at the rider level.

Balancing capacity properly requires three things: demand forecasting granular enough to right-size the pool by area and time band, continuous re-allocation so that capacity follows volume during the day rather than after it, and hub readiness coordinated with departure timing so riders are not waiting on consignments. In Locus, the Capacity Agent handles forecasting and pool sizing, the Dispatch Agent re-sequences continuously, and the Hub Agent runs outbound readiness and handoff as one chain of custody so that dock release and rider departure are one coordinated decision rather than two independent ones.

Also Read: The Three-Workforce Fleet Reality: How Owned, 3PL, and Gig Drivers Actually Operate at Most Enterprises

Deliveries per hour: the metric everyone quotes and nobody can benchmark

Deliveries per hour, or stops per rider-hour, is the right internal productivity metric for a Southeast Asian e-commerce operation. It captures sequencing quality, cluster density, service time and idle time in one number, and it moves when the decisioning layer improves.

It is also a metric for which no credible external benchmark exists. Every published deliveries-per-hour or stops-per-route figure traces to software vendors or aggregator content rather than to a research firm, a consultancy, a government statistics body or a peer-reviewed source. The same is true for first-attempt failure rates, absolute cost per drop, fleet utilisation by vertical, and idle-time reduction percentages attributed to specific levers.

That is worth knowing before a business case is built on one, because a figure that cannot be sourced will not survive a finance review. The defensible approach is straightforward.

• Measure deliveries per rider-hour segmented by density tier, not as a network average. A network average across a dense city core and a peri-urban fringe describes neither.
• Segment by vehicle type, because two-wheeler and van economics are not comparable.
• Track the trend against your own prior period rather than against an industry figure.
• Pair it with first-attempt completion, because deliveries per hour can be improved by attempting more stops rather than completing more, which raises cost per successful drop while the productivity metric appears to improve.

That last point is the one that catches operators. Optimising the wrong half of the ratio is a common and expensive mistake.

Lever 4: promise discipline at the point of order

Cost per successful drop is partly determined before the order reaches the network, by what the customer was told at checkout.

A delivery date generated from a static lead-time table rather than from live network capacity produces one of two outcomes. Over-promising produces failures, re-attempts and support contacts. Under-promising costs conversion. Both are margin.

Capacity-aware promising means the commitment shown at checkout is computed from what the network can actually execute, including current capacity, serviceability by area, and the cut-off realities of hub operations. It is the intervention that most directly protects cost per successful drop, because a promise the network can hold is a promise the recipient is more likely to be present for.

Also Read: Stop Routing Bad Promises: Why Last-Mile Efficiency Actually Starts at the E-Commerce Checkout

Lever 5: exception recovery cost

At scale, exceptions become a cost centre rather than an operational irritation.

The full cost of an exception includes re-attempt labour and distance, support handling, refunds or credits, and the capacity consumed by rework that could have carried new orders. That last component is the one growing operations under-count, because it is an opportunity cost rather than an invoice.

Two disciplines matter. Detect exceptions during execution rather than at end of day, so the remaining plan can be re-decided while recovery is still cheap. And make recovery a system decision rather than a rider decision, so that a failed attempt triggers an evaluated next action, whether that is re-sequencing for a later attempt the same day, redirecting to a pickup point, or rescheduling with the recipient.

Why point tools underperform at scale

Each of the five levers can be bought separately, and many operators do exactly that: a routing tool, a rider app, a tracking platform, a warehouse system.

The levers are coupled. Hub readiness determines departure timing, which determines route viability. Allocation determines cluster density, which determines the value of re-routing. Promise accuracy determines exception volume, which determines how much capacity is consumed by rework. Optimising each in isolation leaves the coupling unmanaged, and the coupling is where the scale economics live.

Deloitte finds that enterprises which orchestrate AI agents well could increase the value they capture by 15% to 30%. Read alongside the McKinsey handover figure of 13% to 19% of logistics cost, the two describe the same value pool from opposite directions: it sits in the coordination rather than in the individual functions.

Last-mile capability has moved through three generations: monitoring systems that report rider location, analytics systems that report performance after the fact, and orchestration systems that sense conditions, decide, execute and learn continuously. Locus operates in the third tier through its SDEL architecture, Sense-Decide-Execute-Learn, across the DiSCO agent suite of Capacity, Carrier, Dispatch, Hub, Customer, Settlement and Orchestrator agents, with Mycroft AI Co-Pilot as the natural-language interface.

Six governance mechanisms, Explainability, Traceability, Evaluation, Autonomy Levels, Execution Sandbox and Human-in-the-Loop, keep automated allocation and dispatch decisions auditable, which matters when software is deciding how thousands of rider-hours are spent each day.

Also Read: How AI Dispatch Reduces Cost-Per-Stop: A Benchmarking Guide for Last-Mile Operations

Deployment evidence from Southeast Asian operations

Multi-market distribution at scale: a global food and beverage leader. This operation runs one of the largest F&B distribution networks across Southeast Asia and MENA, serving 150,000+ retail outlets. In Thailand alone it spans 100+ distribution centres, 33+ cities and 5,000+ vehicles dispatched monthly. Before Locus, routes and dispatch were built manually on informal logic that ignored real operational constraints, riders and vehicles were tracked manually with no alerts when an SLA slipped, transporter management was handled market by market with no consistent way to compare rates, and proof of delivery was verified by hand.

On Locus, the Dispatch Agent plans and sequences every route against 250+ live constraints modelled as the customer’s own business rules and re-routes in real time. The Capacity Agent forecasts demand and right-sizes the fleet. The Carrier Agent scores every transporter on cost and service with competitive trip bidding. The Hub Agent runs hub and multi-leg movements as one chain of custody with AI-verified proof of delivery at the drop. Results across six markets: 97%+ SLA adherence, 18M+ orders planned per year, 22% reduction in procurement costs, 15% improvement in rider time efficiency and approximately 90% of proof-of-delivery reviews automated. Detail in the global FMCG logistics automation case study.

Two figures matter for the argument of this piece. The 15% rider time efficiency improvement is the idle-time recovery the fifth lever describes, achieved through planning and forecasting rather than rider-level intervention. And 18M+ orders planned annually at 97%+ SLA is the scale evidence: SLA held while volume was carried, which is the definition of absorbing growth without eroding service.

Multi-carrier e-commerce parcel: a leading ASEAN apparel retailer. This retailer runs a large store network alongside a global e-commerce business, with last-mile running almost entirely through carriers, each with its own systems, rates and service areas. Carrier onboarding took over three months per carrier as a full engineering project, which capped how fast the network could add capacity. Without a date computed across the carrier mix, the storefront showed only a rough lead time, driving hundreds of thousands of delivery and returns complaints in a single half-year. And every carrier reported delivery events in its own status codes, so no common view of the network existed.

On Locus, allocation runs on serviceability and the retailer’s own hard rules, then selects across cost, speed or performance on the mix the retailer sets. A network-aware delivery date is computed across the carrier mix so the storefront shows a date the operation can hold. Every carrier’s status is harmonised into one standard set and synced back to the retailer’s OMS and WMS. Results: carrier onboarding from three months to three days, a 40%+ drop in WISMO and returns queries, 99%+ delivery SLA and sub-500ms carrier label generation. Detail in the multi-carrier parcel management case study.

The onboarding figure is the one to read as a scale lever. Three months to three days changes what capacity strategy is available during a growth phase, because a network that can add a carrier in days can respond to a volume surge with capacity rather than with overtime.

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 recognised Locus for seven consecutive years. The full set is at Locus analyst recognition.

Five questions for a scaling Southeast Asian operation

• If volume in your densest city rose 30% next quarter, would cost per successful drop go up, stay flat, or fall? Can you show why?
• What is your measured latency from a disruption signal to a revised, dispatched sequence?
• When a rider is idle, can you attribute it to hub readiness, allocation density, or plan sizing?
• Is the delivery date shown at checkout computed from live network capacity, or from a lead-time table?
• How long does it take to onboard a new carrier or fleet partner, and is that a technology constraint or a commercial one?

Learn more, visit locus.sh