AI for Last-Mile Delivery
Last mile is the most expensive segment of the chain and its dominant avoidable cost is the second delivery attempt, which is predictable from data most operators already hold.
A failed delivery costs the drop twice and frequently loses the customer's confidence in the retailer who used you. Predicting failure before the van leaves is more valuable than optimising the route it takes.
AI for last-mile delivery covers route optimisation using measured service times, delivery failure prediction and prevention, time slot offering and management, courier allocation and capacity planning, address quality and geocoding, and customer communication.
Service times, which are usually assumed and rarely measured
Route optimisation depends on how long each stop takes, and most systems use a flat assumption or a crude category. Real service times vary enormously by address type, access, parcel size, recipient behaviour and time of day — and they are measurable from GPS traces you already collect.
- Derive service time per address from historical traces. A flat above street level, a gated estate, a business with a goods-in desk and a rural farm are not comparable stops.
- Access and parking dominate in dense urban areas. Time from van stop to door and back is frequently larger than the driving time between stops.
- Courier variation is real and worth modelling. Familiarity with a round substantially affects service time, which matters for allocation as well as routing.
- A route built on wrong service times fails predictably. Early stops run to plan, the error compounds, and the last stops are late or undelivered.
- Measured service times improve the plan more than a better algorithm. This is the highest-return input improvement available.
Address-level service time is the single biggest routing improvement available
Most last-mile operators run good optimisation software on poor service time assumptions, which produces plans that look feasible and are not. Deriving per-address service times from your own GPS traces, segmented by access type and parcel profile, typically improves plan reliability more than any change to the optimiser. It is unglamorous data engineering over data you already collect, and it is where we usually start.
Failed deliveries, the dominant avoidable cost
Failure is predictable from history and address characteristics
Whether someone is home, whether the address is accessible, whether previous attempts failed, whether the slot suits the recipient — these are learnable, and a first-attempt failure probability is actionable before the van is loaded.
The intervention matters more than the prediction
A high-risk delivery can be re-slotted, diverted to a collection point, confirmed by message, or routed to a time when the recipient is likely home. Prediction without an available intervention is a report, and we ask which interventions exist first.
Address quality is a large and boring cause
Incomplete addresses, wrong geocoding, missing access instructions and unit numbers cause failures that look like recipient absence. Cleaning and enriching address data is unglamorous and directly reduces failure.
Measure first-time delivery rate, not attempts completed
A round that completes all attempts including three that failed is not a successful round. First-time delivery rate is the number that reflects both cost and customer experience.
Slots, couriers and the employment question
- Slot offering is a yield problem. Which slots to show, at what price, given remaining capacity and route efficiency, is a revenue management question with a routing constraint attached.
- Tighter slots reduce failure and cost capacity. The trade is real and should be modelled rather than assumed.
- Courier allocation should account for round familiarity. A courier who knows a round is materially faster on it, which is an argument for stability over pure optimisation.
- Courier employment status shapes what you can direct. In many jurisdictions the degree of control exercised over self-employed couriers is legally consequential, and an optimiser that dictates every minute may create exposure.
- Monitoring couriers is personal data. GPS, behaviour scoring and app telemetry carry transparency and purpose limitation obligations, and courier-facing systems attract scrutiny.
How closely you direct couriers is a legal question, not just an operational one
In several jurisdictions the level of control a business exercises over self-employed couriers bears directly on employment status determinations. A system that assigns work, dictates sequence, prescribes timing and monitors compliance minute by minute increases that exposure. This is worth establishing with employment counsel before designing the system rather than after a tribunal, and it is a genuine constraint on how prescriptive the optimiser should be.
How an engagement runs
Service times measured from your own traces before any routing work begins.
Scope and intervention audit
Which failure interventions actually exist, and what the routing process can act on by when.
Service time derivation
Per-address service times from GPS traces, segmented by access type, parcel profile and courier familiarity.
Build
Routing on measured service times, and delivery failure prediction linked to available interventions.
Trial
Against control rounds, measured on first-time delivery rate and drops per hour together.
Operation
Service times updated as rounds change, address quality maintained, failure causes tracked.
What you receive
More first-time deliveries and routes that finish when the plan says they will.
Measured service times
Per address from your own traces, by access type, parcel profile and courier familiarity.
Routing on real inputs
Plans that hold to the end of the round rather than degrading after stop twenty.
Delivery failure prediction
Before the van is loaded, linked to an intervention that actually exists.
Address quality improvement
Geocoding, access instructions and unit-level data cleaned and enriched.
Slot offering optimisation
Yield against capacity and route efficiency, with the failure trade modelled.
Courier allocation
Round familiarity weighted, within whatever direction constraints your employment position requires.
Is this the right starting point?
Worth being direct. There are situations in last-mile delivery where custom AI work is the wrong spend, and those are listed rather than buried.
Worth doing if
- Routes are built on assumed service times and consistently over-run.
- Failed deliveries are a material cost and prediction is not attempted.
- Address data quality causes failures that are recorded as recipient absence.
- Slot offering is fixed rather than optimised against capacity.
- GPS traces are collected and never used to derive service times.
Do something else if
- GPS traces are not retained, so service times cannot be measured.
- No failure interventions exist, so prediction changes nothing.
- The employment position rules out the direction the optimiser would require.
- Round volumes are too small for per-address patterns to be reliable.
Frequently asked questions
Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.
Why do our routes always over-run?
Almost always service time assumptions. Most operators run capable optimisation software on flat or crudely categorised service times, which produces plans that look feasible and are not — early stops run to plan, error compounds, and the last stops are late or undelivered. Deriving per-address service times from your own GPS traces typically improves plan reliability more than any change to the optimiser.
Can we predict failed deliveries?
Yes, from delivery history, address characteristics, previous attempt outcomes and slot fit — and the prediction is only useful if an intervention exists. Re-slotting, diverting to a collection point, confirming by message or routing to a time the recipient is likely home are all real options. We ask which of them you can actually do before building the prediction, because otherwise it is a report.
What is the most under-rated cause of failure?
Address data quality. Incomplete addresses, wrong geocoding, missing unit numbers and absent access instructions produce failures that get recorded as recipient absence, which means the real cause is invisible in the reporting. Cleaning and enriching address data is unglamorous, cheap relative to its effect, and it reduces failure directly.
Should slots be tighter or wider?
It is a genuine trade and worth modelling rather than assuming. Tighter slots reduce failure because recipients can plan around them, and they cost capacity because they constrain routing. The right answer depends on your failure cost, your route density and your customers, and it differs by area — which is exactly the kind of question a model answers well once service times are measured properly.
How prescriptive should the optimiser be with couriers?
That is partly a legal question. In several jurisdictions the degree of control exercised over self-employed couriers bears on employment status, and a system that assigns work, dictates sequence and monitors minute-by-minute compliance increases that exposure. Establish the position with employment counsel before design — it is a real constraint on how prescriptive the system should be, not a compliance footnote.
Related verticals
Organisations in last-mile delivery usually share data, buyers or regulators with these. All fourteen are listed on the Transport & Logistics page.
AI for Courier and Postal Operators
Sortation and network flow, peak forecasting and address quality — at network scale.
Read more →AI for Freight and Trucking
Load matching, pricing, ETAs and driver retention — with autonomy sized honestly.
Read more →AI for Fleet Management
Maintenance, fuel analytics, safety programmes and replacement timing — plus honest EV transition modelling.
Read more →Tell us what the problem looks like.
Thirty minutes, no charge, no deck. We will tell you whether this is an AI problem, a data problem, or a process problem — and we will say when the honest answer is to buy something rather than build it.