I work in a logistics SaaS company. I have closely studied the operations of various courier-express-parcel companies and e-commerce/quick-commerce companies, particularly in India and the Middle East. This is my attempt to document the day-to-day operations of a typical last-mile hub, the challenges they face and the white-spaces in the industry - based on my observations on ground.
What in the world is last-mile logistics
Most logistics networks follow a hub-and-spoke topology. This means its entire serviceable area is covered by a set of hubs - each hub is responsible for picking up, consolidating and delivering shipments in their designated areas. These hubs connect to each other or via a parent hub that consolidates from and redistributes to its child hubs. In scaled up operations, there could be intermediate layers - for example, Fedex network has a hierarchy of local stations/depots, regional hubs and superhubs.
But why hub-and-spoke? Consolidation of load is key to achieving economies of scale in logistics. Hub-and-spoke ensures that shipments originating from and terminating in disparate points are consolidated into a few high volume corridors where larger vessels/trucks/aircrafts can be filled to reduce unit cost. As against a point-to-point model, which can only work with high predictable demand on bilateral corridors.[note in appendix]
Accordingly, each shipment passes through a network of hubs and its journey can be broken into three components: first-mile i.e. picking up from origin to delivering it to the designated hub of the area; middle-mile i.e. moving the shipment from the origin hub to destination hub; and last-mile i.e. picking up from the destination hub and delivering it to the destination.
Since the service of delivering from A to B is in itself commoditized, a courier company’s differentiation and right-to-win is determined by broadly three levers: cost per shipment, time per shipment (or % of shipments delivered within the promised time) and customer experience. A courier company will win against its competition or earn more surplus, only if it has lower costs, faster deliveries and happier customers (promise adherence, reliable tracking, delivery scheduling for example). With varying degrees of preference, the same metrics also measure the goodness of the logistics ops of e-commerce/quick-commerce companies.
A day at a last-mile hub
The last mile journey begins with receiving shipments from the middle mile (i.e. coming in from another hub) or first mile (i.e. if origin and destination hubs are the same). These shipments would come in throughout the day and are unloaded from the truck/van into the staging area of a hub.
The first step is sorting. The ultimate objective of the hub is to divide all shipments coming to the hub into a set of buckets, such that each bucket can be delivered by a single driver in a single shift. Why do we need sorting at all? At scale, thousands of consignments coming in everyday at the hub cannot be stored randomly as it will become tough to find any particular consignment for its eventual dispatch. So, sorting is needed to make the downstream task of search (a.k.a. picking) efficient.
To enable that, each consignment address is mapped to a zone. A sorter’s job is to scan the shipment at receipt and place it in the designated pallete/shelf of the zone. In operations with a large number of zones or no clear mapping, sorting may be done in two-passes. For example, some hubs in the UAE serving multiple emirates, first take all shipments of an emirate in the respective aisle and then shelve it based on the particular zone in the emirate.
Zones are essentially polygons drawn out on a map that further sub-divide the polygon served by the hub. For simplicity, a courier company could use a standard set of polygons like pincodes. Or for optimality, they may create their own set of zone polygons. The polygons are created to ensure broadly equitable distribution of loads across zones by studying the historical demand patterns. Zone creation should also take into account specific topological features of the area like a zone wouldn’t cross a highway or a river as it would be difficult and time-consuming for a driver to cross them.
But how does one map the shipment to its zone? E-commerce and quick commerce companies receiving orders through an app are better off in this regard, as they capture a customer location pin when booking an order and so can automatically assign the zone based on the polygon in which the coordinates lie. But for most courier or B2B logistics companies, they receive only textual addresses - in which case, the mapping of a consignment to its zone is a cause of major inefficiency. Typically, geocoding APIs provided by Google, AWS and their likes are used to get the latlong/coordinates of a textual address. But these APIs are expensive and often inaccurate, particularly in South East Asia and the Middle East. The other option is to rely on the knowledge and experience of the sorter to read the address and figure out the zone - this slows down sorting and is error-prone. A third alternative is to form some kind of mapping of areas in the address to zones and ask your customers to choose one of the pre-defined areas while booking the shipment.
Irrespective of the system used, an incomplete and unstandardized address input is a pain for the operator. In such cases, either the customer needs to be called up to get extra details - which means more time and cost per shipment. Or else, the shipment is sorted based on the best-guess zone. Mapping a shipment to the wrong zone would result in the wrong driver loading it in his van in the morning, eventually failing to deliver it and then bringing it back to the hub by evening, only to be delivered the next day.
Some countries’ national posts have made standard conventions to address this problem. Like the Saudi National Addresses mandates a 8 digit code for each building and Qatar’s Anwani system assigns 3 numbers to each location in the country (a zone, street and building number). But in most other places, this is a key whitespace to solve for, thanks to the inherent complexity and heterogeneity of addresses.
It would be fair to ask why do we need zones at all when I can use a computer program to optimally decide driver routes for each day/shift and then the shipments can be sorted directly by the driver routes of the next day? I, however, found that zoning is almost ubiquitous in the industry for three main reasons: one, with a large number of drivers, the optimal routing problem is computationally complex - zoning reduces the problem size; two, zone to driver mapping is key to driver productivity. A driver is at their highest productivity when they are assigned deliveries in a fixed area everyday, as they have an acquired knowledge of the area’s addresses; three, typically dispatch to drivers is happening every morning and until a few hours before that, the hub manager does not have the exact list of consignments to be delivered that day. Thereby, allocation of shipments to routes/drivers before the shipment reaches the hub is practically tough.
After the consignments have been sorted per their zone, the second step is planning. A few hours before the drivers are to come in, the hub manager needs to decide which, if not all, consignments will be dispatched today and which driver will be allocated which shipments. One might think that there is no need for planning if each zone is mapped to a driver so when they come to the hub every morning, they pick up the already-sorted shipments of their allocated zone.
However, the load for each zone varies by each day. If the drivers follow a one-to-one zone mapping, they’ll either be overburdened or underutilized. Additionally, if all shipments cannot be dispatched today due to capacity and time constraints, then shipments need to be prioritized. Priorities can be based on service type (a premium delivery service promising fixed datetime over a standard shipment that can wait for a day), the type of goods (perishable goods will lose their value if they wait) or customer preference. Additionally, some couriers may allow their customers to schedule delivery dates and time. Replenishment of dark stores and mall-based stores often need an appointment-based delivery to avoid disruption of their ops. Therefore, while zone-to-driver mapping helps prepare a first draft of the day’s plan, it is then adjusted to factor in constraints and to optimize the allocation.
Constraints ensure that the plan is feasible i.e. the vehicle can realistically carry all the shipments allocated and the driver can realistically attempt all deliveries allocated within their shift time. This sounds simpler than it is in practice. Why? Firstly, a typical courier company uses a mix of bikes, cars, small vans and mini-trucks and delivers everything from an envelope, a garment, a refrigerator to a carton. The heterogeneity of fleet and goods means a feasible plan should factor in the dimensions and weights of each shipment and the dimensions and weight capacity of each vehicle. Additionally, certain goods like pharmaceuticals or refrigerated food require temperature-controlled vehicles; certain service types like bank/identity documents or electronics with open-box deliveries/coupled installation service can only be allocated to the specially trained drivers. I have also seen permit-related constraints like each vehicle may not have the permit to deliver in all emirates in the UAE. Secondly, time taken to deliver is not a direct function of the number of shipments. A driver going to a far away zone should be given fewer shipments than another driver going to a nearby zone (assuming the same shift time). The estimate of the time on the road should factor in the traffic and the exact route they’ll take as mini trucks may not be able to take the same route a car or bike can. Moreover, in B2B logistics where there may be a number of shipments to be delivered at a single address, shipments are grouped into stops and the number of stops (instead of number of shipments) is used as a proxy to estimate delivery time.
While constraints are necessary for good planning, they are not sufficient - it also needs to be optimized on either of the three objective functions I mentioned above: the lowest cost per delivery, least time per delivery or least customer promises breached. Or a complex mix of all three. Optimality of a route plan is often measured in terms of driver productivity, vehicle utilization and % overlap of routes. Some courier companies run a mix of self-owned fleet, vendor-leased fleet and gig-worker fleet. Accordingly, the cost input for each driver and vehicle may vary - it is computed based on fuel consumed and driver salary for self-owned fleet, on per-day used for leased fleet and on per-delivery or per-hour for gig workers.
After all adjustments, there are always exceptions and one-off cases: return shipments, cancelled orders, orders that were dispatched yesterday but couldn’t be successfully delivered, discrepancy in the shipment dimensions, a sick driver who cannot turn up today or a vehicle that broke down. Lastly, efficient couriers would share the same fleet for first-mile and last-mile i.e. the driver will do pickups as well as deliveries on their route. If so, the optimization also needs to factor in the pickup stops and time.
Planning logic or algorithm is rarely documented - traditional courier companies rely on heuristics-based rules or intuition of veteran planners who’ve done it for years on Excel. Modern companies (like Delhivery) and particularly, e-commerce/quick-commerce companies (like Zomato, Noon) have analyst teams and software to optimize routes.
Therefore, optimal planning is one of the highest leverage problems in the industry. Great planning can take a driver’s productivity from 20-25 deliveries/day to 70-80 deliveries/day, given enough density of orders. While the computational problem of route optimization is largely solved, the harder problem is coordinating imperfect data, physical operations, human judgement and constant exceptions.
Once the shift’s plan is finalized, the next step is dispatch. The planner will furnish the runsheet of each driver that has details of the shipments they are supposed to deliver in the shift. Based on the runsheet, a picker will procure the shipment from the specific shelf/palette and drop it in a cage/bucket designated for the particular driver. This is often done late at night or early in the morning, so that as soon as the driver arrives at the hub, they can quickly load the consignments of their designated bucket into their vehicle, after scanning/verifying them against the runsheet. Here again, optimally sequencing the deliveries of a single route and accordingly loading the vehicle can improve the productivity significantly. While software solutions exist for sequencing and guided loading, experienced drivers often have a better sense of the optimal sequence in their area and do not refer to any mandated sequences.
After dispatch, you know the story - the driver goes to location, calls each customer just before reaching the destination, marks the delivery outcome in their app or on paper and captures a picture as a proof of delivery. During this period, hub supervisors monitor and observe the drivers and intervene if needed. Exceptions may either be driver-raised like a vehicle has broken down or a road is closed. Or they may be customer-raised, like a customer called to inform them that they are not available to take the delivery. An attentive supervisor may also identify exceptions, like a driver deviated from the route or is whiling away time at a stop. Technology companies have found this monitoring and actioning as a key space to deploy AI agents that pro-actively identify deviations and take action autonomously.
After they have completed their shift, they will return back to the hub for the last step in the hub’s day - debriefing. This includes returning back the failed or unattempted deliveries to the hub, reconciling any cash-on-delivery and at some places, a supervisor review of the proofs-of-deliveries and reasons for failure.
This marks the end of a day or shift, only to start all over again. A great hub runs this process like clock-work, day-after-day, with perfectly orchestrated staff and utmost punctuality so that productivity of each resource is maximised. The more precisely you time and optimize each step - sorting, planning, dispatching, monitoring and debriefing - the higher efficiency and lower costs you’ll achieve. A quick-commerce hub (a.k.a. dark store) runs the same process but on steroids - instead of a single or shift-wise dispatch, they keep allocating and dispatching orders as and when they come in.
Technology-based applications have a massive opportunity to revolutionize these traditionally manual operations processes. As computation has become cheaper and faster, just-in-time optimization is replacing heuristics-based processes. Imagine this: addresses are cleaned and grounded to their coordinates as soon as a shipment is booked; if anything is incomplete, an AI agent can Whatsapp or call the customer to seek details and update the lat/long accordingly. Before the shipment arrives in the last-mile hub, the software has already mapped the zones for each shipment based on its lat/long and so a robot can scan each label and place it in the right shelf. The route optimizer would run an optimization before every shift and allocate shipments to each driver. The robot can again pick the shipment from the shelf and place it in the right basket for the respective driver. Once dispatched, an AI agent can track route deviations and idle vehicles on-the-go, based on the GPS pings of the drivers phone or IOT devices on the vehicle and take corrective actions autonomously like nudging drivers, calling customers to inform of delays.
As we speak, tech-first logistics companies are already putting these solutions in place. However, most scaled-up traditional courier companies suffer from the upstream problem of data sanctity. A complex web of different softwares - order management system, warehouse management system, transport management system, emails and phone calls from customers, an accounting system or ERP for finances and billing - results in siloed data with unstandard schemas. This makes integrations and optimization challenging in day-to-day ops. Also, ops veterans who have run these processes for years, do not trust tech solutions they don’t understand. And for good reason - they are wary of a black box optimizer messing up their route plan for the day or a flaky integration layer disrupting their ops. For smaller courier companies, no cheap, out-of-the-box solution fits given the bespoke nature of constraints and investing in these rather expensive custom solutions has limited ROI.
That said, operational efficiency is only one of the multiple levers that determine the efficiency of the courier company. The topology of the network, density of demand, cost structures of the fleet, location of the hubs and zoning of their territories are all important factors but are beyond the scope of this essay.
Some companies have made tactical deviations from a strict hub-and-spoke topology. For example, the Ministry of Foreign Affairs (MoFA) in the UAE runs a document attestation service with certain courier partners. Herein, a courier picks up original documents from the customer's home, transports them to MoFA for a physical official stamp, and delivers them back to the customer. One of these courier companies I worked with, runs a milk-run topology wherein drivers are assigned specific areas where they keep doing pickups and deliveries. A mid-size truck goes around the city, meeting each driver on the road to take all the pickups they have done so far and handover the deliveries. After the truck has done this exchange with each zone’s driver, it will terminate its trip at MoFA or a hub. Delhivery follows a mesh-of-hubs model wherein each shipment is dynamically routed to the next hub based on their algorithm, instead of a fixed hierarchy of hubs; Amazon in 2023, announced a policy of regionalization to enable same-day and next-day deliveries. They have created a set of microcosms of hub-and-spokes in each region.
