How AI-Driven Dispatch and Allocation Improve Grocery Delivery Productivity

Why grocery is not a general routing problem

Grocery delivery looks like last-mile delivery with a tighter clock. Planned as though it were, productivity stalls at a level well below what the same fleet could achieve.

Five constraints separate it, and they interact rather than adding up.

Temperature classes. A single order can span ambient, chilled, and frozen. That constrains which vehicle or which compartment can carry it, how long it can sit, and which other orders it can travel with. Compatibility is a routing constraint rather than a loading preference.

Substitutions. The order the customer placed and the order the store picked are frequently different. Substituted items change weight, volume, and occasionally temperature class, and they do so after the route was planned.

Narrow committed windows. One and two-hour windows are standard rather than premium, which removes most of the recovery time a general model relies on.

Recipient presence. Chilled and frozen goods usually cannot be left, which makes availability a hard delivery condition rather than a preference. A general model treats a failed attempt as a reattempt; a grocery model has to treat it as a loss.

Perishability. The product runs a clock that no dispatch decision suspends, so time in transit is a quality cost as well as a service cost.

Each constraint alone is manageable. Together they define a feasible set narrow enough that plans built without modelling them are frequently invalid rather than merely suboptimal.

Also Read: Grocery Delivery Logistics: How AI Dispatch Handles Fresh, Frozen, and Ambient in One Route

Measure productivity as deliveries per driver hour

Most grocery operations track stops per route or orders per shift. Neither isolates productivity from the shape of the work.

Deliveries per driver hour does, and it is the metric AI-driven dispatch and allocation actually moves. The reason is that grocery service time is long and highly variable: crates to carry, a customer to meet, sometimes a doorstep handover with substitutions to explain. Route length says little about output when the stop itself consumes most of the time.

The research on urban delivery makes the point in general terms. Urban Freight Lab work at the University of Washington, based on more than 1,800 real deliveries, found urban commercial vehicles spend around 80 percent of daily operating time parked, with most of a driver’s time spent outside the vehicle. In grocery the ratio is at the harder end of that range, because the handover itself is slower than a parcel drop.

Two corollaries follow for anyone building a productivity programme.

Distance optimisation addresses the minority of the shift. A ten percent reduction in drive time moves a small share of the day when driving is a fifth of it.

Service time accuracy is the highest-leverage input. A plan assuming four minutes per delivery and encountering nine will miss from the middle of the route onward regardless of how well the stops were sequenced.

Where the productivity gains actually come from

Three mechanisms, in order of typical size.

Continuous re-planning as the picture changes

A grocery plan built at 06:00 is invalid by 09:00 for reasons specific to the category: orders arrive after cutoff, picks complete late, substitutions change load composition, and items go out of stock entirely.

An operation re-planning once absorbs all of that manually, which means a dispatcher patching a plan they cannot see the whole of. An operation re-planning continuously converts each change into a revised allocation across the affected routes.

The gap this closes is documented. Gartner found that while 95 percent of supply chains must react quickly to change, only 7 percent can execute decisions in real time.

Multi-trip routing within a shift

The largest single lever in most grocery operations, and the one most often left unused.

A driver completing a route and returning to store has hours of shift remaining. Whether those hours produce further deliveries depends on whether the system plans multiple trips as one continuous assignment, accounting for reload time, remaining hours, and the windows still available.

Planned as separate dispatch events, the second trip either does not happen or happens inefficiently. Planned as one assignment, the same driver and vehicle produce substantially more output against the same fixed cost.

Siam Makro, the largest B2B online-to-offline retailer in Asia, shows the mechanism at scale. Across 160+ stores and 10,900+ active riders, continuous wave-based dispatch planning in 30-minute increments with multi-trip routing raised orders per rider per day from 10 to 15 up to 18 to 20, while dispatch time per store fell from two hours of human planning to under 30 minutes and logistics cost fell 16.7 percent.

Service time learned per address

Grocery service time varies more by location than by order. A ground-floor house with a driveway and a fourth-floor apartment with a security door are different jobs with the same basket.

Systems using a flat constant misallocate time systematically in both directions, and the compounding effect is what produces routes that run late from stop six onward. Learning service time per address, from observed delivery history, is the change that makes narrow windows holdable rather than aspirational.

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

Also Read: Five Complex Problems AI-Driven Dispatch and Allocation Immediately Resolves

The substitution problem

This deserves separate treatment because it is genuinely specific to grocery and almost nothing models it.

The sequence is straightforward and its consequences are not. An order is placed. A route is planned against it. The store picks it, substituting some items and shorting others. The load now differs from what the plan assumed, in weight, in volume, and occasionally in temperature class if a chilled item was substituted for an ambient one or vice versa.

Three consequences follow.

Capacity assumptions break. A vehicle planned to capacity on ordered volume may not fit the picked volume, which is discovered at loading under time pressure.

Compatibility assumptions break. If a substitution introduces a temperature class the compartment allocation did not anticipate, the co-load is no longer valid.

Service time changes. A delivery with substitutions to explain takes longer than one without, and in some models the customer has a right to refuse at the door.

The design requirement is that the plan is built or revised against picked quantities rather than ordered quantities, which means the dispatch layer needs the pick confirmation as an event rather than as a status. That is a WMS or store-systems integration question, and it is the single most valuable integration in a grocery dispatch stack.

Window accuracy is a productivity lever, not a service one

In most last-mile categories, window accuracy is a customer experience concern. In grocery it is also a productivity concern, because recipient presence is a delivery condition.

A missed window in grocery frequently means a failed delivery, which means the full approach and handover cost incurred with no completed delivery, plus a reattempt that incurs it again, plus product that may not survive the round trip. Failed attempts therefore consume productivity directly rather than only costing service.

That reframes the promise decision. A window the network can hold produces more deliveries per driver hour than a narrower window it misses, which is the opposite of the intuitive commercial instinct. And customer preference supports the trade: McKinsey found that speed fell from consumers’ number one delivery priority in 2022 to fifth by 2024, displaced by reliability and predictability.

Capacity-aware promising, where the window offered at checkout is computed against real available capacity rather than a static table, is therefore a productivity intervention as much as a customer one.

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

What to measure

Six measures, and the first two are the ones most operations lack.

Deliveries per driver hour, segmented by route type. The headline productivity metric.

Trips per driver shift. This reveals whether multi-trip capacity is being used, and it is usually the largest available gain.

Service time per address, with distribution rather than average. The tail is what breaks windows.

Window adherence by window width. A high adherence rate on four-hour windows and a poor one on one-hour windows tells you where the promise is miscalibrated.

First-attempt rate, with structured failure reasons. Recipient unavailable and access problems require different responses.

Substitution rate and its effect on service time, which almost nobody measures and which explains a meaningful share of window misses.

Where Locus fits

Locus, the world’s first Decision-Intelligent, Agentic TMS, models the grocery constraint set inside the planning engine rather than applying it as a post-plan check.

Within DiSCO, the Dispatch agent plans and re-sequences against 250+ real-world constraints, which include temperature class, compartment allocation, vehicle eligibility, and delivery windows; the Capacity agent matches demand to available driver and vehicle capacity across a shift, which is what makes multi-trip routing a plan rather than a sequence of dispatches; the Hub agent manages store-side readiness and the dispatch handover where pick confirmation arrives; and the Customer agent manages the promise when a plan changes. The Learn stage of the Sense, Decide, Execute, Learn cycle is where observed service times replace assumed ones.

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. In October 2025, Ingka Investments, the investment arm of Ingka Group, the world’s largest IKEA retailer, acquired Locus. Locus continues to operate independently.

Beyond the Siam Makro deployment above, a leading Canadian grocery brand shows the contracted-carrier version of the same problem. Delivering fresh, perishable food across more than 30 cities through 3PL carriers, warehouse associates had been creating shipments manually in each carrier portal and validating addresses by hand, with the team’s stated position being that every hour of manual data entry was freshness lost in transit. Moving order creation and carrier selection into autonomous allocation produced 33 percent faster deliveries, 15 percent lower fulfilment costs, and 25 percent less time on manual shipping tasks, with order frequency rising 10 percent.

Also Read: Agentic TMS for North America’s Cold Chain Logistics: What Food and Grocery Shippers Should Know

The two numbers to pull first

Trips per driver shift, and service time distribution by address.

The first tells you whether you are using the shift you are already paying for. In most grocery operations a meaningful share of drivers complete one trip and return with hours remaining, and closing that is the largest available productivity gain without adding a vehicle or a driver.

The second tells you why your windows miss. If the service time distribution is wide and your plan uses a constant, the misses are a modelling artefact rather than an execution failure, and no amount of driver management fixes them.