Last-Mile Delivery is a Time Problem, Not a Distance Problem, and the Industry is Optimizing the Wrong Thing

The inconvenient math

Last-mile delivery absorbs a disproportionate share of logistics spend. Capgemini Research Institute puts it at 41 to 53 percent of total logistics and shipping cost. Almost all optimization investment aimed at that cost targets one variable: road distance.

Research on how the day actually divides suggests that is the wrong target. Urban Freight Lab research at the University of Washington, based on more than 1,800 real deliveries and published in Transportation, found urban commercial vehicles spend around 80 percent of daily operating time parked, with most of a delivery driver’s time spent outside the vehicle walking the final stretch to the customer. For cargo-cycle drivers the split was roughly 60 percent parking and walking against 40 percent driving.

Sit with the arithmetic that implies. If driving is a fifth of the shift, a ten percent improvement in road distance moves two percent of the day. A ten percent improvement in what happens at stops moves eight percent. The industry has spent fifteen years and considerable capital getting very good at the smaller number.

This is the structural explanation for a pattern many operations leaders recognize: the routing platform delivered the efficiency gain it promised, and cost per delivery did not fall proportionally. Nothing was broken. The gain was real and it applied to a minority of the cost base.

The thesis is straightforward. Last-mile delivery efficiency is a stop-time problem wearing a routing problem’s clothes, and the operations that figure this out first will hold an advantage that is difficult to copy, because it depends on accumulated stop-level data rather than on a software license.

Also Read: What Last-Mile Delivery Efficiency Actually Means: Six Dimensions Fleet and Operations Leaders Should Measure

What dwell time is made of

Dwell time is the elapsed period between a driver arriving in a delivery zone and departing it. Treating it as a single number is the first mistake, because it is four different things with four different causes.

Parking and approach. Finding a legal, usable space and walking to the delivery point. This is where the Urban Freight Lab finding concentrates, and it varies more by street geometry and time of day than by anything the driver controls.

Building navigation. Intercom, security desk, elevator, stairwell, finding the correct unit. Multi-family residential and office towers are where this dominates, and it is the component least visible to any system that models a stop as a coordinate.

Recipient interaction. Whether someone is present, whether a signature or ID check is required, whether there is a conversation, whether the driver returns to the vehicle for a second item.

Exception handling. A failed attempt consumes the full parking, approach, and navigation cost and produces no completed delivery, then generates a second stop that repeats all of it. This is the most expensive category of dwell by a distance, and it is the one most often counted only as a failed-delivery statistic rather than as time.

Now the modeling problem. Most routing engines represent dwell as a constant, either one value across all stops or one per broad category such as residential and commercial. A constant is a reasonable engineering choice when you have no better data, and it produces a plan that is structurally wrong before the vehicle leaves the depot, because the actual value varies by building, by floor, by time of day, and by whether the concierge desk is staffed.

Facility-side dwell shows the same pattern in freight. ATRI found drivers were detained at 39.3 percent of all stops in 2023, losing between 117 and 209 hours per year depending on sector, at a cost of 3.6 billion dollars in direct expenses and 11.5 billion dollars in lost productivity. Different segment, same lesson: time at the stop is where the money goes and it is systematically underestimated in planning.

The four stop archetypes

A route of 25 stops is not 25 equivalent events. It is a portfolio of stop types with different cost profiles, and planning it without knowing the composition is like pricing a menu without knowing ingredient costs.

Two things this table deliberately does not contain: minute ranges and cost multipliers.

That is not an omission. Credible dwell time benchmarks by stop type do not exist. Figures in circulation for cost per stop, stops per route, and first-attempt failure rates trace to software vendors and aggregator pages rather than to research firms, government sources, or peer-reviewed work. Publishing invented ranges here would contradict the argument in the same paragraph as making it, because the argument is that almost nobody has this data.

Your own numbers are the only ones worth planning against, and the good news is that they are recoverable from timestamps you may already be collecting.

Also Read: Why Address Validation Isn’t Enough for Big-and-Bulky: The Building Intelligence Architecture US Furniture Delivery Actually Needs

Drivers are not the variable, the system is

The instinct when dwell time varies is to look at driver performance. That instinct produces marginal gains and considerable resentment, because most of the variance is systemic.

Sequencing by building type, not just by geography

When high-dwell stops are scattered through a route, the driver absorbs building navigation overhead repeatedly with no offsetting benefit. Clustering them into route segments lets a single parking event and a single building entry serve several deliveries, which converts fixed overhead into shared overhead.

Pure geographic optimization does not do this, because geographically adjacent stops can have completely different access profiles. Two addresses fifty meters apart, one a doorstep and one a twelfth-floor unit behind a security desk, are not comparable stops.

Time-of-day assignment by stop type

Dwell varies through the day for reasons that have nothing to do with traffic. Commercial receiving windows are the obvious case, where arriving outside the window converts a predictable stop into an open-ended queue. Residential buildings vary too: whether a concierge is on duty, whether residents are home, whether the loading bay is occupied.

Most routing systems model time-of-day effects on travel, because traffic data is readily available, and model no time-of-day effect on dwell at all.

Information asymmetry at the stop

A driver who knows before arriving that a building requires a specific entrance, that the door code changed, that the customer asked for the parcel to go to the side gate, or that the service elevator is out completes that stop faster than a driver who discovers each of those on arrival.

This is the cheapest available intervention and the one most often left undone, because it requires capturing what drivers learn and returning it to the next driver at that address. Most operations lose this knowledge every time a driver leaves.

The conclusion for operations leaders: dwell reduction is a data and systems problem. Approached as performance management it yields a few percent and damages trust. Approached as sequencing, timing, and information it compounds across every route.

Also Read: Last-Mile Delivery Efficiency in Dense Urban Areas: Why Standard Operational Playbooks Fail

What instrumenting the stop reveals

Most operations hold route-level completion time and failed-delivery rate. Stop-level dwell is the most underused diagnostic available, and four findings emerge almost immediately from having it.

Which stop types systematically overrun the model, which lets the routing engine be recalibrated with observed distributions instead of constants. This alone usually improves ETA accuracy more than any change to the routing algorithm.

Which addresses are persistently expensive, as opposed to which zones. Buildings behave consistently, so a per-address dwell history is more actionable than a zone average, and it is what makes clustering and window assignment possible.

The true time cost of a failed attempt, which is larger than the redelivery fee suggests because it includes the original approach and navigation time that produced nothing. Counting failures as a rate rather than as consumed hours understates them.

Where targeted investment pays. Electronic proof of delivery, customer pre-notification, building access arrangements, and parking strategy all reduce specific dwell components, and stop-level data tells you which component your operation actually suffers from rather than which one a vendor sells against.

The reframe worth carrying into your next planning cycle: a routing platform that captures stop-level timestamps is not only a navigation tool. It is a cost accounting system for the largest and least understood part of your delivery cost.

One economic note for network design: the US Postal Regulatory Commission has found average cost per delivery in rural areas runs approximately twice that of urban areas. Urban stops are cheaper per delivery and, as the research above shows, dominated by time rather than distance, which means the cost lever differs by geography and a single optimization strategy across a mixed network will underperform in both.

Where this shows up in deployment

Locus, the world’s first Decision-Intelligent, Agentic TMS, models stop-level constraints and feeds executed outcomes back into planning, which is the loop that turns dwell observations into better plans. Within DiSCO, the Dispatch agent plans and re-sequences against 250+ real-world constraints, and the Learn stage of the Sense, Decide, Execute, Learn cycle is where observed service times replace assumed ones.

Two deployments show stop-time efficiency rather than distance efficiency. Siam Makro, the largest B2B online-to-offline retailer in Asia, raised orders completed per rider per day from 10 to 15 up to 18 to 20 across 160+ stores and 10,900+ active riders, with multi-trip routing and 30-minute incremental dispatch waves. That is throughput per shift, which is the metric dwell time governs. A global food and beverage leader operating across six markets recorded a 15 percent improvement in rider time efficiency specifically, attributed to constraint-aware routing and real-time re-sequencing freeing rider hours across daily runs.

Locus 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.

Also Read: Deliveries Per Hour: The Rider Productivity Metric That Reveals Hidden Last-Mile Waste in 2026

Five questions about your operation’s dwell profile

1. Do you capture arrival and departure timestamps at stop level, or only route start and route completion?
2. What dwell value does your routing engine currently assume, and is it one constant, one per category, or learned per location?
3. What is the spread of actual dwell within a single stop category, and how does that spread compare to the constant you are planning against?
4. Which individual addresses appear repeatedly in your longest-dwell decile, and what do they have in common?
5. How many hours per week does your operation lose to failed attempts when approach and navigation time is counted, rather than counting failures as a rate?

Question one determines whether the rest are answerable. Question five usually produces the number that changes the conversation.

Also Read: The First-Attempt Delivery Rate: A Key Metric That Decides Last-Mile Profitability in 2026

The closing reframe

Mapping and routing algorithms are commoditized. Any operation can buy a good one, and the difference between a good one and an excellent one is a few percent of a fifth of the day.

Stop-level intelligence is not commoditized, because it cannot be bought as a feature. It accumulates: which buildings are slow, which addresses need a note, which facilities queue at 14:00, which zones lose ten minutes to parking. An operation with three years of that data plans better routes than an operation with the same software and none of it.

Distance optimization is table stakes. Stop-time intelligence is where the remaining advantage is, and it compounds for whoever starts measuring first.