AI for Transport and Logistics
Where a forecast error becomes a truck in the wrong place. Eleven verticals, written for the person who has to move the freight tomorrow rather than the person modelling the network.
This is an operations research sector wearing a machine learning label.
Most of the genuinely valuable work in logistics is optimisation rather than prediction: routing, scheduling, loading, slotting, network design, crew rostering. These are constrained optimisation problems with decades of established method behind them, and they work. The reason they are frequently sold as AI is that AI sells better, and the reason they frequently disappoint is that the constraints used in the model are not the constraints that exist on the ground.
The second thing that separates the projects that land from the ones that stall is whether the plan can survive contact with operations. A route that ignores a driver's knowledge of a loading bay, a schedule that assumes a dock turnaround nobody achieves, a slotting plan that puts heavy items on a top shelf — each is technically optimal and operationally rejected. We spend the first weeks of a logistics engagement collecting constraints that exist only in people's heads, because that is what determines whether the optimiser is used.
Where AI actually earns its place in logistics
Ranked by evidence rather than by conference frequency. The maturity column is our own read; the catch column is what the case study leaves out.
| Where | What it does | Maturity | The catch |
|---|---|---|---|
| Route optimisation and scheduling | Sequences stops, assigns vehicles and builds routes against real constraints. | Proven | Decades of established method. The constraint set, not the algorithm, decides whether anyone uses it. |
| Demand and volume forecasting | Predicts shipment volumes by lane, service and location. | Proven | Reliable. In contract logistics you inherit your customers' forecast error, which is worth modelling separately. |
| Warehouse slotting and picking optimisation | Places stock and sequences picks to reduce travel. | Proven | Travel is most of the pick cost, so the gains are real and directly measurable. |
| ETA prediction | Predicts arrival times across road, sea, air and rail. | Strong | Genuinely improved by data. Accuracy is bounded by events nobody can see — port congestion, customs, weather. |
| Document and customs automation | Extracts and validates data from bills of lading, invoices, declarations. | Proven | High volume, structured, immediately measurable. Exception rate decides the saving. |
| Fleet safety and driver behaviour | Detects risky driving from telematics and in-cab video. | Strong | Effective, and it is a workforce intervention before it is a technology one. Design it as such or it will be resisted. |
| Network design and capacity planning | Locates facilities and designs flows against cost and service. | Strong | Operations research, and it works. The hard part is data quality on cost-to-serve. |
| Autonomous long-haul trucking | Driverless operation on defined corridors. | Early | Real and small. Around 440,000 driverless miles across ten commercial routes by mid-2026, against billions of truck miles annually. |
Autonomy is real, and it is smaller than the coverage suggests
Driverless trucks are genuinely operating without a person behind the wheel on public roads, on around ten commercial routes in the US Sun Belt, having accumulated in the region of 440,000 driverless miles by mid-2026 with a stated production target of a thousand trucks a year. That is a real engineering achievement and it is a rounding error against annual US truck mileage. The proportionate response is to keep watching it, to make sure your systems could accept an autonomous carrier as a capacity option, and to spend this year's budget on routing, forecasting, dock scheduling and document automation — which return now, on lanes you actually run.
Road traffic, working time and the paperwork that moves freight
Transport sits under road traffic and transport safety law, working time and tachograph rules, customs and trade regimes, and for some systems the AI Act's critical infrastructure category. This is our reading as at September 2026 and we work alongside your compliance and legal functions rather than in place of them.
Annex III includes road traffic
AI intended for use as a safety component in the management and operation of critical infrastructure, including road traffic, is high-risk with obligations from 2 December 2027. Routing, scheduling, forecasting and ETA systems are operational tools rather than safety components. Systems that intervene in vehicle behaviour, or that manage traffic infrastructure as a safety function, are a different matter and should be classified deliberately with counsel.
The most sensitive system you will deploy
In-cab cameras and driver behaviour scoring touch employment law, data protection and, in the EU, the AI Act's prohibition on emotion inference in the workplace. Fatigue and distraction detection based on observable behaviour is generally distinguishable from emotion inference, and the boundary is worth establishing with counsel before deployment rather than after a works council objects.
Hard constraints, not preferences
Driver hours, rest periods and tachograph rules are legal limits. Any scheduling or routing optimiser must treat them as inviolable constraints rather than costs to trade against — and must handle them correctly across jurisdictions for international operations, where the applicable rules change at a border.
The EU reform is coming, and one change already landed
The EU Customs Data Hub opens for e-commerce consignments in 2028, becomes voluntary for other importers in 2031 and mandatory for all traders in 2034. The €150 duty exemption is abolished and e-commerce platforms become the importer, collecting duties and VAT at purchase. A temporary €3 per-item duty applies from 1 July 2026 until 1 July 2028.
Not yet adopted, and worth watching closely
The IMO Net-Zero Framework was agreed in principle in April 2025, its adoption vote was delayed from October 2025, and it remains unadopted — with negotiations continuing through 2026 and projected entry into force around 2028. Anyone selling you a compliance product on the basis that the framework is in force is ahead of the facts.
Telematics is personal data
Vehicle telematics, driver scoring and in-cab video are personal data about workers, with the transparency, minimisation and purpose limitation obligations that follow. In many jurisdictions this also engages works council consultation, which is a project timeline consideration rather than a compliance footnote.
What this means for a build
Two things are architectural. Legal driving time limits belong in the constraint set, encoded as hard and jurisdiction-aware, never as penalties an optimiser can trade against. And any system observing drivers should be designed as a workforce intervention first: what is collected, who sees it, what it is used for and what it is never used for, agreed before deployment. The technology in both cases is straightforward; the failure mode is a system that is legally exposed or operationally rejected. See AI policy development and AI data privacy compliance.
Who we write for
Each page starts from that organisation's own problems, names the regulatory exposure it carries, and routes into the engineering. Depth varies and is stated on each page.
Logistics & 3PL
Quoting, cost-to-serve, routing and customer forecast bias — where margin is won at the tender.
Read more →Freight & trucking
Load matching, pricing, ETAs and driver retention — with autonomy sized honestly.
Read more →Shipping & maritime
Voyage optimisation, fuel and emissions, port call planning and container repositioning.
Read more →Air cargo & freight aviation
Capacity and revenue management, load planning and booking accuracy — where a booking is not a shipment.
Read more →Rail
Rolling stock condition, infrastructure inspection and delay analytics — outside the signalling safety case.
Read more →Last-mile delivery
Routing with real service times, failure prediction and slot management — where the cost is the second attempt.
Read more →Warehousing & fulfilment
Slotting, pick paths, labour forecasting and dock scheduling — where travel is most of the cost.
Read more →Fleet management
Maintenance, fuel analytics, safety programmes and replacement timing — plus honest EV transition modelling.
Read more →Courier & postal
Sortation and network flow, peak forecasting and address quality — at network scale.
Read more →Customs & trade
Classification support, declaration automation and screening — with the EU customs reform timeline mapped.
Read more →Supply chain planning
Demand forecasting, inventory optimisation and lead time variability — where the plan meets the process.
Read more →FAQ
Marked up with FAQPage schema so these answers can surface in search results and inside AI assistant responses.
Do you have logistics experience?
Yes in routing and scheduling optimisation, demand and ETA forecasting, warehouse operations analytics and document automation. Less in maritime operations specifically, and none in vehicle autonomy engineering — we would not claim otherwise. Each of the eleven vertical pages states our depth in that area rather than implying uniform expertise across the sector.
Is this machine learning or operations research?
Mostly operations research, and we say so because it changes what you should buy. Routing, scheduling, slotting, loading and network design are constrained optimisation problems with decades of established method and reliable results. Machine learning contributes the inputs — demand forecasts, ETA predictions, service time estimates — that make the optimisation realistic. A supplier presenting routing as a machine learning breakthrough is describing a solved problem in fashionable language.
Why do our optimisation projects get ignored by operations?
Almost always because the constraint set is incomplete. Drivers, planners and warehouse supervisors hold constraints that exist nowhere in writing: which bay a vehicle cannot reach, which customer will not accept a delivery before ten, which pick face jams if you slot heavy items there. An optimiser missing those produces plans that are technically optimal and practically wrong, and it takes about two rejections before people stop looking at it. We spend the first weeks collecting those constraints, because that is what determines adoption.
Should we be planning for autonomous trucks?
You should be able to accept one as a capacity option and you should not be reorganising around it this year. Driverless trucks are genuinely running without a person aboard on around ten commercial US routes, with roughly 440,000 driverless miles by mid-2026 — real, and a rounding error against annual truck mileage. The proportionate move is systems that can handle an autonomous carrier in the mix, while this year's budget goes to routing, dock scheduling and document automation that return now.
What is the most common way logistics AI projects fail?
They optimise something the business cannot execute. A perfect route that ignores the dock booking system, a slotting plan the warehouse management system cannot implement, a forecast that arrives after the carrier commitment deadline. We ask what the operational process can actually act on, and by when, before scoping anything — and that question has ended engagements before anyone spent money.
Start with the problem, not the technology.
Thirty minutes, no charge, no deck. Tell us what is going wrong in your organisation and we will tell you whether AI is the right instrument — including when it plainly is not.