AI in Logistics, Separated From the Marketing

Separate the optimisation from the prediction

“AI in logistics” covers two different things. Route optimisation is a constraint-solving problem with decades of mature software behind it. Prediction — how long will this actually take, which delivery will fail — is a machine learning problem and depends on your own history.

Confusing them leads businesses to commission a model when they should buy a routing product, or vice versa.

Document handling is the safe first project

Delivery notes, consignment paperwork, customs documents, supplier invoices. Extraction into structured data removes a large amount of keying and is straightforward to verify.

In international freight the document burden is the administrative bottleneck, and it is the least contentious place to start.

ETA prediction from your own data

Carrier estimates are generic. Your own history knows that deliveries to that industrial estate take longer on Fridays, that this customer's goods-in closes early, and that this route is slower in winter.

A model trained on your own completed jobs usually beats the carrier's estimate substantially, because it knows your specific customers and routes. That improves customer communication more than any tracking page.

Exception triage

• Classify failed deliveries by cause from driver notes and photographs
• Route each to the right resolution path automatically
• Predict which shipments are likely to fail before they are attempted
• Surface patterns — a customer whose failures cluster on one day, a site with recurring access problems

Pattern surfacing is where the recurring value sits, because it turns individual incidents into fixable causes.

What to buy rather than build

Route optimisation, telematics and standard track-and-trace. These are mature markets with capable products, and building them is rarely justified.

Build where the requirement is specific: your own ETA model, your own exception classification, and the integration between the products you have bought.

Data quality decides everything

Prediction needs accurate timestamps of what actually happened — arrived, unloaded, departed — not just planned times. Many operations record the plan and not the reality.

If that is your position, capture comes first. A year of accurate event data makes prediction possible; without it, no model will help.

Frequently asked questions