Last-Mile Delivery Software: How It Is Built in 2026

Last-mile delivery software is the system that takes a parcel from the local depot to the door and proves it got there. It plans the day's stops, assigns them to drivers, sequences each route, runs the app in the driver's hand, pushes a tracking link and an ETA to the recipient, records the handover, and deals with the stops that go wrong. What makes it hard is the setting: the last mile is the leg of the chain where the vehicle spends most of its time parked, where every address is a separate small negotiation with a building, and where a missed handover means doing the work twice.

The money at stake is large and the margins are thin. US e-commerce reached $340.2 billion in the second quarter of 2026, 17.1 percent of retail sales, according to the Census Bureau's seasonally adjusted estimate (revised figures are due on 19 November 2026). Amazon's shipping costs were $102.7 billion in 2025, about 14 percent of its net sales by our arithmetic. UPS told shareholders that its US Domestic cost per piece rose 8.1 percent in 2025, faster than revenue per piece, and 16.8 percent year on year in the second quarter of 2026 (8.0 percent adjusted), mainly on driver buyout costs and USPS fees but also because of "lower average daily volume and stops." The software that runs this leg is a market Gartner calls extremely fragmented, and the top of the search results for it is a row of vendor homepages and listicles that cite no measurements at all.

This article is for the people who have to pick or build this software: operations leads, CTOs and product managers. It covers what the measured evidence says about last-mile economics, how each component of a last mile delivery platform is built and where it fails, what the published prices are, and when building your own is a sound decision. Every number links to the page it was read on.

The short version

• The last mile is a density problem. The USPS now delivers 2.4 pieces of mail and parcels per delivery point per day, down from 5.5 in 2007, and UPS lists "lower average daily volume and stops" among the reasons its cost per piece rose. Software that raises items per stop and first-attempt success moves cost more than software that shortens driving distance.
• A parcel driver's day is spent mostly parked. In measured London rounds the van was parked 62 percent of the time; in six Seattle ride-alongs across mixed commercial vehicles it was 80 percent. The money is at the stop: parking, walking, handover and failure handling.
• Routing engines learned humility from 9,184 real Amazon routes: on 1,107 high-quality routes, experienced drivers' routes ran 12.9 percent longer than the travel-time optimum, and the winning model scored by imitating the driver, not by minimising time.
• The driver app is bounded by the phone. Android gives a background app location "only a few times each hour", iOS lets an app watch at most 20 regions, Android 100, and one study measured an iPhone's GPS error at 7 to 13 metres on a university campus, against 4.9 metres under open sky. Offline-first design is a requirement, not a feature.
• Buying is cheap. Published list prices for 5,000 deliveries a month run from $285 to $1,349 (read 1 October 2026). Build when the delivery layer is the product or the constraints are yours, and buy the solver even then.
• Compliance is now product design: the EU Platform Work Directive, due in national law by 2 December 2026, bans collecting a courier's data while they are not working and requires a human to take any deactivation decision.

What last mile delivery software is, and what it is not

Last mile delivery software is the planning and execution layer for the final leg: it sits between the order system that knows what must be delivered and the driver who delivers it. Gartner's Peer Insights category describes it as "specialized, customer-centric transportation management solutions that focus on managing the delivery process by which a consumer receives the products they order online", and lists 42 products in it as of October 2026. The June 2024 Market Guide for the category, the latest edition with a public abstract, made two points: the market "remains extremely fragmented", and most vendors "do not yet offer all the capabilities required to fully support all last-mile fulfillment processes."

Four neighbouring things get called by the same name, and the confusion costs money in procurement:

A retailer that owns its vans needs the first row. A shipper that tenders parcels to carriers needs a transportation management system and a tracking aggregator, not a dispatch tool. A marketplace that pays gig couriers per drop needs the first row plus an assignment engine that is the business itself. Our sibling article covers route optimization software and the vehicle routing problem underneath it; this one covers the platform around the solver.

The economics of the last mile, with numbers that can be checked

Last-mile cost is a labour cost spread over stops, and the two levers are how many items each stop delivers and how often a stop has to be repeated. The widely quoted shares are weaker than they look. The "41 percent of supply chain costs" figure comes from a Capgemini survey of 500 grocery and consumer-products executives in late 2018, so it is what executives said, not what accountants measured. The "50 percent of parcel delivery cost" line is a single unsourced sentence in McKinsey's 2016 last-mile report. We found no primary study behind the "53 percent" in vendor blogs, and the "13 to 75 percent" range in a 2009 University of Antwerp conference paper was "compiled using academic literature and contacts with last mile service providers." None of these should be repeated as data.

The volume base is large and still growing. The Pitney Bowes Parcel Shipping Index, a vendor index whose method is not published, counted 23.1 billion US parcels in 2025, up 3.3 percent, at an average revenue of $9.34 per parcel: a revenue pool of about $216 billion by our arithmetic. Amazon Logistics was the largest carrier by volume at 6.9 billion, ahead of USPS at 6.2 billion, UPS at 4.3 billion and FedEx at 3.9 billion.

The per-piece numbers the carriers disclose tell the cost story better than any share. UPS's US Domestic revenue per piece was $13.21 in 2025, up 7.1 percent, while cost per piece rose 8.1 percent (UPS does not publish the absolute cost figure). FedEx's US ground yield was $12.49 per package in the fiscal year to May 2026, and the company contracts "approximately 5,300 independent small businesses" for certain linehaul and pickup-and-delivery operations. The USPS, by our arithmetic on its FY2025 10-K, earned about $5.55 per Ground Advantage piece on 2.93 billion pieces. The three figures are not like for like: UPS's $13.21 includes air services (its Ground average was $11.60), and Ground Advantage is a slower economy product, so read them as price points rather than three prices for one physical act.

Why last mile delivery cost is a density problem

Density decides the economics because the driver's time is the cost and the number of items handed over per stop is the output. The clearest long-run series is the USPS's, which counts letters and parcels together and ties the fall to "lower mail volume" and more addresses: "a drop in the average number of pieces delivered per delivery point per day from 5.5 pieces in 2007 to 2.4 pieces in 2025", while "the number of delivery points continues to grow with an increase of 1.8 million in 2025, which drives up delivery costs." Amazon's 10-K says the same thing from the shipper's side: a poorly run fulfilment network "increases our net shipping cost by increasing the distance products are shipped and reducing the number of units per shipment or delivery."

UPS gave the field a natural experiment in 2025. It brought its SurePost product in-house as Ground Saver in January 2025 and found that "pickup and delivery costs were higher than in 2024". By December it had signed a new agreement with the USPS "to assist with final-mile delivery for a portion of our Ground Saver" volume. In the second quarter of 2026, having cut its Amazon volume by more than half from 2024 levels, it reported cost per piece up 16.8 percent (8.0 percent adjusted), driven mainly by separation costs from its Driver Choice buyout programme and higher fees paid to the USPS, with "lower average daily volume and stops" also on UPS's list. The density equation shows through the filing, but it is not the main reason for that quarter's jump.

Labour is the biggest line. The BLS counts 983,300 light truck drivers at a median $21.57 an hour (May 2025, excluding the self-employed); its Occupational Outlook Handbook, which counts about 1.1 million jobs including the self-employed, puts 37 percent of them with couriers and messengers and projects 6 percent growth to 2035. Wages are 68.6 percent of total compensation for transportation and material moving occupations in private industry (BLS, June 2026), so the mean $23.45 wage loads to about $34 an hour. Amazon says its funding will help Delivery Service Partners raise driver pay to a national average of nearly $23 an hour.

A worked example, assumptions stated: a 7.3-hour round at $34 an hour costs about $250 in driver time. Spread over the 72 customers and 118 items of a measured London round (below), that is $3.47 per customer and $2.11 per item before the van, fuel, insurance, depot and the dispatcher. Spread over 36 customers, it is $6.93. Nothing a routing engine does to driving distance changes that arithmetic as much as one more item per stop does.

Failed deliveries and time windows

A failed first attempt is the most expensive event in the last mile because it repeats the whole stop. The only government series we found is Japan's: the Ministry of Land, Infrastructure, Transport and Tourism sampled 3.3 million parcels from six carriers in April 2026 and measured a redelivery rate of about 7.6 percent, down from 8.4 percent a year earlier, with urban areas at 8.5 percent and rural at 6.0 percent. Among the three largest carriers, the share of parcels received without a face-to-face handover (delivery boxes, leave-at-door and other options) was about 31.0 percent, and the government's target is about half by fiscal 2030. The much-repeated "a failed delivery costs $17.20" comes from a vendor survey of 304 retail executives in December 2020, so treat it as an opinion with a sample.

In the UK, Ofcom's 2025 online survey of 4,058 people aged 16 and over found that 68 percent had experienced a delivery issue with any operator in the previous six months, led by delays, parcels left in inappropriate places and drivers not knocking loudly enough, and that "better tracking information" was the most requested improvement. Across the EU, Eurostat reports "delivery slower than indicated" as the most common online shopping problem in 2025, hit by about 20 percent of recent buyers.

Time windows are the other cost multiplier. The peer-reviewed estimate is from Boyer, Prud'homme and Chung (2009): "offering a 3 hour delivery window is 30 to 45% more expensive than offering unattended (9 hour delivery window) delivery." A promise engine that offers narrow windows by default is pricing the operation out of its own margin; the window is a product decision to be made with the cost in view.

Returns ride the same vans. In a summer 2025 survey of 358 e-commerce professionals at US merchants with more than $500 million in revenue (plus 2,006 consumers), the National Retail Federation and Happy Returns, a UPS company, found that retailers expect 15.8 percent of their 2025 sales to be returned, 19.3 percent of online sales, and that 9 percent of returns are fraudulent. A platform that treats a pickup as a first-class stop type, with its own POD rules, is doing something the first generation of delivery tools did not.

What a driver's day looks like, measured

The best evidence about where last-mile time goes comes from researchers who rode along, and it points away from the road. Allen and colleagues accompanied 25 parcel rounds for two carriers in central London in 2016 (the author preprint is open access). A round lasted 7.3 hours, delivered 118 items to 72 customers from 37 parking stops, 95 percent of them on-street, at 4.1 minutes per customer once parked. Within the delivery area the vehicle drove 11.9 kilometres at an average 7 km/h and the driver walked 7.9 kilometres. "For 62% of this time, the vehicle was parked while deliveries or collections were being made." The finding that matters most for software buyers is a sentence most summaries skip: "neither carrier used any form of routing and scheduling software to manage the last-mile transaction", because the address on the manifest was not where the van would park.

The University of Washington's Urban Freight Lab reached the same shape from the other coast. Across 2,900 parcel trips in downtown Seattle, drivers spent an estimated 2.3 minutes per trip cruising for parking, 28 percent of trip time. In six ride-alongs with four carriers, urban commercial vehicles "spent 80% of their daily operating time parked". A 2025 study of more than 1,800 deliveries found a median dwell time of 17.5 minutes per stop for one parcel carrier, 3.4 deliveries per stop, and that "it is still entirely up to the driver to choose where and how far to park the vehicle from the delivery customers." UPS's own figure, from its January 2020 routing announcement, is that many of its drivers average 135 stops a day.

Three design consequences follow. Service time is the variable to model, learned per address from the app's own timestamps, because a manifest cannot know that one building has a loading bay and the next has a concierge who takes ten minutes. The parking location and the delivery address are different entities; a platform that stores only the address will produce ETAs and geofences that fire at the wrong place. And walking is a routing problem of its own: a cluster of six addresses served from one parked position is one stop to the vehicle and six to the driver, and the app should show it that way.

Inside a last mile delivery platform: the seven components

A last mile delivery platform is seven cooperating services, and the useful way to describe each one is by the way it fails. The feature lists on vendor pages are nouns; the failures are where the engineering time goes.

Order intake and geocoding is where most "routing" problems are born. The platform receives orders from the order management system, a store, a marketplace feed or a CSV, and must turn a free-text address into a point on the map. Google charges $17 per 1,000 Address Validation Pro requests after 5,000 free a month, and plain geocoding $5 per 1,000 after 10,000 free, which is cheap compared with one failed attempt. The failure mode is the plausible wrong point: a geocoder that drops a pin on the street centroid or the wrong building sends the driver to the right postcode and the wrong door, and the ETA engine will defend that pin all day.

Dispatch and assignment turns a pool of stops into a set of routes with drivers, vehicles, start times and depots. The hard constraints are capacity, driver skills, vehicle type (a refrigerated van, a cargo bike), shift length and the stops that must stay together. The failure mode is the plan nobody drives: a dispatch board that produces routes drivers reorder on the first corner has lost the field, and the London study above is what that looks like at carrier scale.

Routing and sequencing orders the stops within a route and re-sequences them in-day when a stop runs long or an order is added. The engine can be a commercial API, an open-source solver such as OR-Tools (Apache 2.0) or VROOM (BSD), or a vendor module. We cover the solver choices in route optimization software; the next section covers what last-mile data taught the solvers.

The driver app is the field terminal: manifest, navigation hand-off, stop status, proof of delivery, exception reporting and messages from dispatch. Its failure mode is connectivity. A van in an underground car park or a rural lane has no signal, and an app that cannot complete a stop offline, queue the evidence and sync later will be worked around with phone calls and paper.

Tracking and ETA publishes the driver's progress to the customer: a link, a map, a window that narrows as the route proceeds. The failure mode is the honest-looking lie: an ETA computed from the planned route rather than the driver's actual position and service-time history, which is confidently wrong by the third stop.

Proof of delivery closes the stop with evidence: photo, signature, barcode scan, PIN. The failure mode is evidence that does not survive a dispute, because the photo has no GPS stamp, the geofence fired 200 metres away, or the signature is a squiggle with no identity behind it.

Exceptions and returns are the stops that did not go to plan: not at home, access refused, damaged parcel, wrong address, a pickup that was not ready. The failure mode is the exception that lives in a WhatsApp group between the driver and the dispatcher instead of in the system, so the redelivery never gets planned and the customer never gets told.

Analytics is the eighth component that lives inside the other seven: service time per address, first-attempt rate per route, ETA error by hour, cost per stop. Without it the platform cannot learn, and the plan on day 200 is the plan from day one.

Routing in last mile delivery software: what 9,184 Amazon routes taught the field

The most important routing result of the decade did not come from a faster solver. In 2021 Amazon and the MIT Center for Transportation and Logistics released a data set of 9,184 routes driven by Amazon drivers in 2018 in Seattle, Los Angeles, Austin, Chicago and Boston, 6,112 for training and 3,072 for evaluation (the training routes have 32 to 237 stops, a mean of 148), and asked researchers to predict the sequence an experienced driver would take. Amazon's framing was blunt: drivers "frequently deviate from those computed routes."

The deviation is now measured. Cook, Held and Helsgaun, who won the $100,000 first prize, computed the travel-time-optimal tour for 1,107 routes Amazon had graded "High" quality and fully delivered, and found the drivers' routes were on average 12.9 percent longer than the optimum. Their winning model scored 0.0248 on the challenge's hidden evaluation routes; on the 1,107 high-quality, fully delivered routes the authors used as a test set it scored 0.0198, against 0.0703 for the pure travel-time-optimal tour (zero is a perfect match, a random shuffle scores about 0.9). Forcing stops to be visited one zone at a time cost only 3.5 percent in extra travel time but cut the score to 0.0487, and the second-placed team found that getting the zone order right with the stops inside each zone shuffled already scored about 0.02. Experienced drivers optimise something the objective function did not contain; in Amazon's own words, they know "which roads are hard to navigate, when traffic is bad, when and where they can easily find parking, which stops can be conveniently served together".

Two other details from the data set matter for anyone building a platform. Time windows barely bound in Amazon's real routes: on the 1,107 test routes, ignoring them altogether left at most 695 violations across 163,177 stops, under half a percent, and the drivers' own tours left 647, so the constraint that dominates textbook VRP mattered little in practice. And the data set's service-time field "may include time required to park and hand-off the package at the drop-off location", which is to say that the biggest cost component is stored as one undifferentiated number even at Amazon.

The carriers' own routing programmes point the same way. UPS described ORION to investors in November 2014 as a system expected to cut seven to eight miles from each driver's day and save $300 million to $400 million on full implementation in 2017 (an annual figure in the Interfaces case study), at a build cost the Interfaces case study put above $295 million. The measured result that reached a filing was in the 2016 10-K: delivery stops up 4.4 percent while "average daily package miles driven only increased 0.2%". In January 2020 UPS said ORION had been saving "about 100 million miles and 10 million gallons of fuel per year" and announced Dynamic ORION, in-day re-optimisation; that release has since been removed from UPS's site, and the FY2025 10-K does not mention ORION. FedEx instead bought its routing vendor, RouteSmart, in February 2025, whose service "serves as the backbone for the internal FedEx Route Optimization (FRO) tool." Deutsche Post went furthest toward the driver: its INFORMS case study describes routes built from couriers' historical behaviour across about 13,000 delivery districts, which is what let it publish delivery time windows for 1.6 million parcels a day.

The design lesson for a last mile delivery system is that the solver should be fed the driver's own history (parking positions, zone sequences, service times) as constraints and costs, and judged on whether drivers follow its plan, not on the kilometres it claims to save.

The driver app: offline first, battery, and the geofence ceiling

The driver app is built against limits the operating systems publish, and a product team that reads them first saves months. Android states that when an app is in the background "it can receive location updates only a few times each hour" (Android developer documentation); continuous tracking during a shift therefore has to run as a foreground service, and since Android 14 every foreground service must declare a type such as location. iOS "suspends the execution of most apps shortly after they move to the background" (Apple developer documentation), and only Always authorization will relaunch a terminated app for region or significant-change events.

Geofenced arrival detection, the feature every vendor page lists, has a hard ceiling. Core Location "prevents any single app from monitoring more than 20 conditions of any type simultaneously", and Android allows 100 geofences per app, with alert latency "usually less than 2 minutes" and a recommended minimum radius of 100 to 150 metres. A route of 148 stops cannot be registered as 148 geofences; the app has to rotate the next few stops in and out, or do its own distance checks against a foreground location stream. And the position itself is noisy: GPS.gov gives 4.9 metres under open sky and warns that accuracy worsens near buildings, and a PLOS ONE study measured an average horizontal error of 7 to 13 metres on an iPhone 6 on a university campus. A 100-metre geofence in a dense block will fire for the stop across the street.

Offline operation is a requirement the OS limits impose rather than a premium feature. The pattern is an event log on the device (arrived, delivered, photo captured, exception raised), each event idempotent with its own identifier and timestamp, synced in order when connectivity returns, with the server treating replays as no-ops. Grab's engineering team describes the server side of that pattern in its message centre design: pending events stored and redelivered when the client reconnects. Battery is the other constraint; a shift runs seven hours or more (the London rounds averaged 7.3), the app runs navigation, camera and location all day, and a driver whose phone dies at stop 90 is a driver who stops reporting.

Store policy shapes the product too. Google Play restricts background location to apps that need it for core functionality, requires a declaration in the developer console, and lists "Delivery/service tracking ... for users (not drivers)" among features that could use foreground access instead, so the customer tracking app and the driver app have different permission designs. Apple's review guideline 3.1.3(e) requires purchases of physical goods and services consumed outside the app to use methods other than in-app purchase, which is why delivery fees carry no App Store commission. A driver app built this way is the shape of work we describe in mobile app development: a field tool first, where the hard problems are the sync log, the battery budget and the permission flow rather than the screens.

Last mile delivery tracking and ETAs: how the promise is kept

Last mile delivery tracking is a prediction problem dressed as a map. The customer link shows the van; the number that matters is the ETA, and the honest way to score an ETA is the one DoorDash uses as its north-star metric, "on-time percentage, or the percentage of orders that had an accurate ETA with a +/- margin of error" (DoorDash Engineering, 2021). The companies that publish real numbers are few. Uber's DeeprETA post-processing model reduced delivery ETA mean absolute error by 7.46 percent against the routing engine's estimate in offline tests, and the 95th percentile error by 11.05 percent, with a median serving latency of 3.25 milliseconds. DoorDash reported a 20 percent relative improvement in ETA accuracy from a mixture-of-experts model in 2024 without defining the metric, and has moved to probabilistic forecasts scored with CRPS. Amazon's SPEEDY framework for sub-same-day delivery predicted narrower promise slots for more than 60 percent of eligible orders in the one locale where it was tested.

For a parcel platform the ETA model has three inputs the planning system must supply: the driver's actual position and sequence (not the plan), the learned service time for the next stops, and a traffic-aware travel time between them. Google's Routes API charges $10 per 1,000 traffic-aware requests (Pro tier) after 5,000 free a month; Mapbox's Directions API is $2 per 1,000 above 100,000 free. The expensive part is the discipline of recomputing the ETA on every status event and pushing the change to the customer only when it crosses a threshold, so the recipient is not pinged twelve times for a four-minute drift.

The notifications themselves are cheap next to labour. Twilio lists US SMS at $0.0083 per segment plus carrier fees of $0.0035 to $0.005 (pricing current as of August 2026); a WhatsApp utility template in North America costs $0.0034 from Meta plus Twilio's $0.005 per message (Meta's rate card effective 1 July 2026; Meta has announced utility pricing changes from 1 October 2026); push through Firebase Cloud Messaging is listed as no-cost.

What recipients want is the thing the ETA is a proxy for: control. In Ofcom's 2025 survey better tracking information (18 percent) came ahead of more time to answer the door (16 percent) and delivery instructions being followed (15 percent). A tracking page that lets the customer change the drop point, add an instruction or move the slot is also the cheapest failed-delivery prevention tool in the stack.

Proof of delivery software: photo, signature, PIN, and what a court will accept

Proof of delivery software records the handover in a form that survives a dispute, and the carriers have converged on the photo. Amazon's help page says plainly that "if your package is delivered by Amazon, we may take a Photo on Delivery", and for some high-value orders it sends a six-digit one-time password that the recipient reads to the driver in person. In June 2022 FedEx announced free Picture Proof of Delivery for residential deliveries released without a signature, with general release in the US and Canada planned for that year's holiday season. UPS's My Choice page says "hundreds of millions of residential deliveries now include a delivery photo."

The legal weight of each artefact is worth knowing before the product promises "proof". In the US, the ESIGN Act provides that a signature or record "may not be denied legal effect, validity, or enforceability solely because it is in electronic form." In the EU, eIDAS Article 25 says an electronic signature "shall not be denied legal effect and admissibility as evidence in legal proceedings solely on the grounds that it is in an electronic form", while only a qualified electronic signature "shall have the equivalent legal effect of a handwritten signature." A finger-drawn squiggle on a driver's phone is a simple electronic signature: admissible, not presumed equal to ink. No statute gives the photo special status either. What makes a POD record hold up is corroboration: a photo, a GPS fix with its accuracy radius, a timestamp from the device and one from the server, the scan of the parcel label, and the driver identity, all bound to the same stop event and stored immutably.

Design the capture for the dispute, not for the dashboard. Store the raw photo with its EXIF, not a compressed thumbnail. Record the geofence outcome and the distance between the fix and the geocoded address, so a 220-metre gap is visible rather than hidden. Keep the retention policy explicit: the French regulator CNIL allows vehicle geolocation data to be kept one year when it serves as proof of the work done and that proof cannot be provided another way, against a two-month default, which is a concrete number to put in the data model.

Integration with the OMS, WMS and TMS: who owns which record

Integration is where last-mile projects lose their schedules, and the cure is to decide, per entity, which system is the source of truth before a single webhook is written. The order belongs to the OMS. Inventory and the physical parcel belong to the warehouse management system, which also knows the staging sequence: a WMS that picks and stages by delivery window and loads the van in reverse stop order is doing half of the dispatcher's job. Carrier selection and the linehaul leg belong to the TMS. The route, the stop, the ETA and the proof of delivery belong to the last-mile platform, and status flows back the other way. The umbrella view, how TMS, WMS, OMS and visibility fit one platform, is in our logistics management software guide.

The identifiers are standardised even when the APIs are not. GS1's Serial Shipping Container Code, application identifier (00), identifies a logistic unit end to end (GS1), and a scan of it at loading, at the door and on return is the backbone of chain of custody. For EDI-speaking partners the X12 214 transaction carries shipment status "in terms of dates, times, locations, route", and the 204 load tender covers truckload moves, not parcels. For the carrier leg, the carriers' tracking APIs set the pace: FedEx's track API allows 100,000 calls a day and 1,400 transactions per 10 seconds per project, with OAuth tokens that expire every 60 minutes; UPS's Track API returns signature and POD details and rolls data off after 120 days; the USPS is retiring its Tracking 3.0 legacy API on 31 July 2027, so anything built on it has a deadline.

Event-driven is the practical architecture: the platform publishes stop events (assigned, en route, arrived, delivered, failed) and the other systems subscribe, with idempotent handlers because the driver app will replay events after an outage. Our knowledge base covers the pattern in event-driven architecture and the wider system integration work. The classic failure is a synchronous call from the driver app to the OMS at the moment of delivery: when the OMS is slow the driver waits at the door, and when it is down the delivery is "not done" in one system and done in another.

What last mile delivery software costs to buy

Published prices exist only at the small and mid-market end, and even there the unit of pricing differs enough that a like-for-like comparison needs arithmetic. The table lists what the vendors' own pricing pages said on 1 October 2026.

Normalised to one workload, 5,000 deliveries a month with ten drivers on monthly billing where offered, the list prices span about 4.7 times: Spoke Starter at $285, Detrack Advanced $390, OptimoRoute Pro $490, Tookan Standard $499, Shipday $509, Track-POD Advanced $590, Routific $630 and Onfleet Scale $1,349. That is our arithmetic on the vendors' pages; the feature sets are not equivalent, and Onfleet's bundle includes more of the platform than the per-driver tools do. At 50,000 deliveries a month Spoke Starter works out to about $2,085 a month, Routific's tiers to about $2,430 and Spoke Expert to about $3,660; Onfleet's Enterprise tier starts at $3,099 for 10,000 or more tasks with an unpublished overage.

The quiet signal in these pages is the pricing unit. Routific says it moved from per-vehicle to per-order pricing in mid-2024; Onfleet counts a task only when it reaches "Completed"; Spoke prices the stop; Google's Fleet Engine for scheduled tasks bills per successfully completed delivery on a contract. Most of these vendors price on the operator's unit of output, which also means the bill scales with volume with no ceiling for a growing operation.

The enterprise tier is a different market. Descartes, a listed logistics-software company that has been buying final-mile vendors, reported revenue of $729.0 million for the year to January 2026, 93 percent of it services, and its annual information form records the February 2023 purchase of final-mile carrier software GroundCloud for about $136.8 million plus an earn-out of up to $80 million; it does not split out last-mile revenue. Locus was acquired by IKEA's Ingka Group in October 2025 for an undisclosed sum, the buyer citing online sales at 28 percent of IKEA retail sales in FY24 against 11 percent in FY19. When a retailer buys its last-mile vendor outright, that is the build-versus-buy question answered at the scale of a balance sheet.

Build vs buy: the arithmetic for last mile delivery software

Build versus buy for last mile delivery software is not a cost decision at any volume where published prices exist, and the arithmetic shows why. The inputs for a build are public. The BLS puts the mean wage of a US software developer at $148,100 a year ($71.20 an hour, May 2025) and a QA analyst at $111,490; with wages at 69.2 percent of total compensation for professional occupations (BLS, June 2026), a developer costs about $214,000 a year including benefits. A team of four developers and one QA engineer for nine months comes to about $763,000 at those rates, before product management, design, cloud, devices or recruiting. At the modelled Eastern European band in our European rates benchmark, €40 to €64 an hour, the same five people for nine months cost roughly €290,000 to €460,000 at 160 billable hours a month, about $330,000 to $520,000 at the ECB rate of 30 September 2026.

The run cost of a built platform is mostly people too, but the third-party services are cheap and worth listing because they are the ones a build cannot avoid:

One optimisation per shipment is the floor; a platform that re-optimises each route three times a day roughly triples the first line ($420 at 5,000 deliveries, $3,670 at 50,000). Open-source routing engines carry their own tax in the licence terms: OR-Tools, jsprit and GraphHopper are Apache 2.0, VROOM and OSRM are BSD, Valhalla is MIT, but openrouteservice is GPL-3.0, which matters if the solver is linked into a product you distribute. The wider budget of a logistics build, integration surface included, is in our supply chain software cost guide.

A worked example at 5,000 and 50,000 deliveries a month

At 5,000 deliveries a month the annual SaaS bill is $3,400 to $16,200 at list price. The build is about $330,000 to $760,000 in year one plus a smaller team to run it. There is no volume in that range at which the licence saving pays for the build.

At 50,000 deliveries a month the SaaS bill is roughly $25,000 to $44,000 a year on Spoke's and Routific's published tiers, about $60,000 on a per-order plan such as Shipday's, and more with an enterprise vendor on a quote. The build cost is the same as above, and the run cost of a two-person team plus about $57,000 a year of the APIs above is roughly $230,000 to $485,000 a year depending on where the team sits. On licence savings alone the build never pays back, because running it costs more each year than the licences it replaces. The decision is made by something else.

When building last mile delivery software is the right call

The reasons to build are structural: they make the commercial tools unsuitable rather than merely expensive. The delivery layer is the product: a marketplace that pays couriers per drop (DoorDash reported 3.2 billion orders and more than 9 million people who dashed in 2025; Instacart 338.8 million orders and about 600,000 shoppers) cannot rent its assignment engine, because batching, pricing and acceptance logic are the margin. The constraints are yours: two-person crews for heavy goods, temperature ranges per compartment, rounds that start at a customer site and never return to a depot, regulated goods with identity checks at the door, each a hard constraint a vendor either models or does not. The tracking experience must be yours, because white-label pages on a vendor domain are a brand problem for a retailer and a contractual problem for a logistics company selling to other brands. Or you already own the dispatch core, and bolting a vendor's platform on as a generic webhook target costs more in integration than building the delivery layer to match it.

The middle path is the one most teams should take: buy the solver (Google's Route Optimization API, a vendor's routing module, or OR-Tools if the team can run it) and build the driver app, the customer tracking surface and the integrations, which are the parts that carry your constraints and your brand. Our Pikkup case study is the adjacent shape: a geolocation marketplace matching drivers with nearby service centres, with in-app payments, taken from idea to a launched product in five months. The delivery-software analogue is a driver app that owns its event log and POD flow and rents the map and the optimiser underneath.

Compliance: driver classification, location data and the rules of the road

Compliance in last-mile delivery has moved from the legal department into the product backlog, because the rules now describe features. The EU Platform Work Directive 2024/2831, which member states must transpose by 2 December 2026 (Article 29), creates a rebuttable presumption of employment "where facts indicating direction and control ... are found", with the burden on the platform to prove otherwise. Its Chapter III reads like a specification for a dispatch app. Platforms may not, by means of automated monitoring or decision-making systems, collect "any personal data of a person performing platform work while that person is not offering or performing platform work" (Article 7). Automated monitoring systems count as high-risk under GDPR Article 35 and need an impact assessment, the effect of automated decisions must be evaluated at least every two years, and a review request must be answered within two weeks. Any decision to "restrict, suspend or terminate" an account "shall be taken by a human being" (Article 10).

The common-law jurisdictions got there through litigation, and product features decided the cases. The UK Supreme Court found Uber's drivers to be workers in Uber BV v Aslam (2021) partly because Uber monitored "the driver's rate of acceptance (and cancellation) of trip requests" and penalised it by logging drivers off; it found Deliveroo's riders were not in an employment relationship for the purposes of collective bargaining under Article 11 ECHR in IWGB v CAC (2023), because their contract gave "a broad and virtually unfettered right to appoint a substitute". Whether an app polices acceptance rates or allows substitution is a classification decision. In California, the Supreme Court upheld Proposition 22 on 25 July 2024; the statute guarantees app-based drivers 120 percent of the applicable minimum wage for engaged time plus a per-mile amount, so the platform must compute engaged time and engaged miles to the cent. New York City's app delivery minimum rose to $22.13 an hour from April 2026, before tips. At the US federal level the 2024 Department of Labor contractor rule is still on the books but not applied by the Wage and Hour Division since May 2025, with a rescission proposed in February 2026 and a final rule scheduled for October 2026.

Location data has its own rulebook. The French regulator CNIL's guidance on employee vehicle geolocation forbids using it to check speed limits or to track outside working hours, even against theft, requires that drivers can switch it off off-duty, caps retention at two months by default (one year for route optimisation, or as proof of the work done when no other proof is possible; five years for working-time records), and says the driver's name is not to be given to the customer unless it is of particular and indispensable interest. In California, precise geolocation, defined as locating a person within a circle of 1,850 feet radius, is sensitive personal information under the CCPA, which reaches the customer tracking page as well as the driver app. The GDPR's data minimisation and storage limitation principles (Article 5) are the reason a tracking link should expire and a location history should not be kept "just in case."

The vehicle rules are lighter than fleet operators fear, with one 2026 change. In the US, a vehicle is a commercial motor vehicle under 49 CFR 390.5 at a gross vehicle weight rating of 10,001 pounds or more in interstate commerce, so a light van is outside the hours-of-service and ELD rules, and a heavier one on local routes usually fits the 150 air-mile short-haul exception; our fleet management software guide covers the ELD and HOS logic for the fleets that do qualify. In the EU, from 1 July 2026 the driving-time rules and the smart tachograph apply to vans over 2.5 tonnes in international carriage for hire or reward or cabotage (Regulation 2020/1054; the European Labour Authority confirms domestic work is not covered). City access rules are routing constraints: London's ULEZ charges £12.50 a day for non-compliant vans across all boroughs, and Greater Paris restricts Crit'Air 3, 4 and 5 vans from 8am to 8pm on weekdays (with informative checks only until the end of 2026), a regime the Conseil constitutionnel kept alive in May 2026 by striking down the national repeal.

Consumer law sets the outer promise: the EU Consumer Rights Directive and the UK Consumer Rights Act default to delivery within 30 days and keep the risk with the trader until the consumer has physical possession, which is why a parcel "left in a safe place" is a liability question. Safety is the number nobody puts on a vendor page: the BLS counted 950 deaths among driver/sales workers and truck drivers in 2024, 708 of them in transportation incidents, at a rate of 25.7 per 100,000 full-time-equivalent workers against 3.3 for all workers; heavy-truck drivers dominate that figure, and light truck drivers, the delivery-van category, accounted for 65. Route plans that reward the fewest minutes reward the fastest driving, and the telematics module should know that.

What changed in last mile delivery software in 2026

Several things that vendor pages describe as trends became dated facts in 2025 and 2026.

UPS, the largest US parcel carrier by revenue in the Pitney Bowes index, gave up a slice of the last mile. UPS insourced SurePost in January 2025, found its pickup-and-delivery cost rising, and by its February 2026 10-K had signed the USPS to carry part of that final mile again; in the second quarter of 2026 its cost per piece was up 16.8 percent, mostly on buyout costs and USPS fees but also on "lower average daily volume and stops".

Amazon's electric fleet passed a round number: Rivian's second-quarter 2026 release says Amazon has "over 40,000 custom-built Rivian Electric Delivery Vans active in its fleet", and Amazon's 2025 sustainability report says its 52,700 electric vans delivered 2.4 billion packages in 2025. Electric vans change routing through range and charging windows as constraints, and through city access: TfL cut electric vans' congestion-charge exemption to a 50 percent discount from 2 January 2026.

The vendor map consolidated from both ends. FedEx bought RouteSmart in February 2025 and had Network 2.0, its consolidation of pickup-and-delivery routes, live at about 410 locations in the US and Canada by May 2026; IKEA's Ingka Group bought Locus in October 2025; Gartner published a December 2025 edition of its last-mile Market Guide (cited by vendors; the abstract is not public).

Out-of-home delivery kept diverging by country. The Netherlands nearly doubled its parcel lockers to 9,309 in 2025 and still delivered 82.2 percent of parcels to the home; InPost ended 2025 with 61,196 lockers and a 49.0 percent adjusted EBITDA margin in Poland; Ofcom put out-of-home at an indicative 12 percent of UK parcel volume in 2024-25; Japan reached about 31.0 percent non-face-to-face among its three largest carriers. A platform sold across borders needs the pickup point as a first-class stop type with its own capacity, hours and POD rules.

Autonomous delivery produced its first useful utilisation numbers. Serve Robotics reported a fleet of more than 2,000 sidewalk robots at the end of 2025 and 792 daily active robots in the second quarter of 2026, under 40 percent of the year-end fleet by our arithmetic; Starship said in April 2026 it had passed 10 million deliveries; Wing says it has logged more than 1 million residential deliveries since 2012. All company-reported. For a platform a robot or drone is a vehicle type with a payload, a range and a hand-off protocol, and a dispatch engine that cannot model it will be configured around.

Two hard dates fall before the end of 2026: the US Department of Labor's contractor rule is scheduled for a final decision in October 2026, and the Platform Work Directive is due in national law on 2 December 2026. Both change what a dispatch app may do to a courier.

Where last mile delivery software goes from here

The measured facts of the last mile have not changed since the London ride-alongs of 2016: the van is parked most of the day, the driver walks kilometres, the address is not where the vehicle stops, and the stop that fails is the one that costs the most. What has changed is that the data to act on those facts now exists inside the driver app, and that the best routing research of the decade concluded that the driver knew something the solver did not. Last mile delivery software that wins in 2026 learns service times per address, treats parking positions and pickup points as entities, scores its ETAs on an on-time percentage it publishes to itself, and captures proof of delivery that would survive a dispute. It buys the solver and the map, because those are commodities, and it owns the event log, the field app and the integrations, because those carry its constraints.

Buy when your operation is standard, and read the pricing unit before you sign. Build when the delivery layer is the product, when your constraints are not in anyone's configuration screen, or when the tracking experience has to be yours. Either way, measure the stop before the road, and put the compliance rules of December 2026 in the backlog now rather than in the incident report later.