AI in logistics: the use cases that survive contact with the yard

Logistics has properties that make it a genuinely good fit for AI: enormous volumes of repeated decisions, continuous telemetry, and outcomes that are measured within hours rather than quarters. You know by the end of the day whether the route was good.

It is also unforgiving. The users are drivers with gloves on and dispatchers under time pressure, the connectivity is intermittent, and anything that adds a step to a shift gets abandoned within a week.

What the systems that stick have in common

1. They run on data the operation already emits — GPS pings, scans, dock events, EDI messages — rather than requiring anyone to enter something new.
2. They degrade gracefully. Signal drops constantly; a system that stops working in a dead zone is a system drivers stop trusting.
3. They output an action, not a score. "Reroute vehicle 12 via the A-road, arrival 14:40" is usable. "Delay risk: 0.71" is a research finding.

A dispatcher does not need a probability. They need to know which truck to move, and where.

Routing and dispatch

The classic case, and still the most reliable. Planning against live traffic and live vehicle positions rather than a snapshot from an hour ago is a straightforward improvement with an immediately measurable effect on ETAs and fuel.

The engineering difficulty is rarely the optimisation. It is the constraint set — driver hours, vehicle capabilities, customer time windows, site access restrictions, union rules — most of which lives in people's heads rather than in a system. Extracting those constraints is the project. DSV Fleet Management is our build here: real-time tracking and dispatch with full visibility over every vehicle, driver, and route.

The same pattern extends past freight. FieldRoute applies it to field service — auto-dispatching the best-matched technician, optimising routes, and tracking first-time-fix against service levels — and Charge Pulse applies live-availability routing to EV charging.

Predictive ETAs, with a confidence figure

Carrier-provided ETAs are frequently optimistic and rarely updated. A model trained on your own historical lane performance, weather, and dwell times will beat them, often substantially.

The important design choice is exposing confidence alongside the estimate. A precise ETA that is silently unreliable is worse than a range, because planners will build a schedule on it. ShipSight tracks shipments across ocean, air, and ground and attaches a confidence score to each predicted ETA, flagging the at-risk ones — which is the part that actually changes behaviour.

Inside the four walls

Warehouse work is dense with repeated micro-decisions, which makes small percentage gains add up quickly.

• Slotting: placing SKUs by velocity and affinity so the fastest movers are nearest the pick face. Re-slotting continuously as demand shifts is the part manual processes never manage.
• Pick-path optimisation: routing a picker along the shortest viable path rather than in SKU order.
• Wave planning: grouping orders so the picking effort per order falls.

StowPilot combines these — re-slotting by velocity, planning pick waves, and routing pickers on the shortest path. The measurement discipline matters here: walking distance per order is the honest metric, because picks per hour can improve for reasons that have nothing to do with the system.

The yard and the dock

Detention charges and driver waiting time are among the most avoidable costs in the chain, and they come from a scheduling problem rather than a transport one. Letting carriers self-book appointments, assigning the right door on arrival, and tracking turnaround makes the constraint visible — which is most of the fix. DockQueue is our build in this area.

Matching and pricing

Freight brokerage is fundamentally a matching and pricing problem with abundant historical data, which is close to ideal conditions. HaulBoard matches open loads to best-fit carriers, prices each lane against live spot-rate data, and tracks margin per load.

The subtlety is that the objective is not "fill the load" — it is margin against the probability the carrier actually performs. A cheap carrier who fails costs far more than the spread saved, and a model optimised on price alone will learn to make that mistake.

What consistently fails

• "Supply chain intelligence" dashboards with no decision attached. Interesting once, ignored by week three.
• Anything requiring drivers to enter data they do not already enter.
• Systems assuming clean master data. Address quality and SKU dimension accuracy are usually the real project, and pretending otherwise wastes a quarter.
• Demand forecasting sold as an AI project when the operational constraint is actually warehouse capacity.

The scoping method in how to scope an AI project applies particularly cleanly here, because logistics decisions are so easy to state concretely. More of our work in this space is on the logistics and mobility page.

• logistics
• supply chain
• operations
• industry