AI in logistics: accurate and slow is the same as useless
Logistics is the sector where latency stops being a quality attribute and becomes part of correctness. Every operational decision here sits against a physical deadline: a sailing, a cut-off, a driver's hours, a dock slot. A customs classification that arrives after the truck has left is not a slow answer, it is no answer at all, and the cost is a missed departure rather than an impatient user. That inverts a trade-off teams normally make without thinking. A system that is ninety per cent right in two seconds is more useful here than one that is ninety-five per cent right in two minutes, which is the reverse of the ranking in almost every other sector on this site. The second thing that shapes a logistics project is document quality, which is worse than anywhere else because the documents come from third parties who chose their own format.
Logistics, in short
| The constraint | Operational decisions have hard physical deadlines, so a system that is accurate and slow is useless in a way it would not be elsewhere. |
|---|---|
| The distraction | Route optimisation, which is a well-studied operations research problem and usually not where the manual effort actually goes. |
| A first project that works | Exception handling on shipping documents: high volume, structured output, and currently done by people reading scans. |
| Where ground truth lives | Resolved exceptions with the resolution recorded, which operations teams keep because they have to. |
What shows up most often, not a description of any particular organisation. No named clients and no case studies: see editorial policy.
Designing to a deadline rather than to an accuracy target
The practical consequence of a physical cut-off is that the system needs a defined behaviour when it has run out of time, and that behaviour has to be designed rather than emergent.
The pattern that works is a fast path with an explicit escape. The system answers quickly where it is confident, and where it is not, it says so immediately and routes to a person with enough time left for them to act. A system that thinks for two minutes and then produces an uncertain answer has consumed the only resource that mattered.
This also changes what an agent design costs. An agent that takes twelve steps is twelve times the latency as well as twelve times the price, and in a sector with cut-offs that is frequently disqualifying on its own. Fixed workflows with a model in one step fit this industry better than most.
The documents arrive in whatever state the sender chose
Bills of lading, customs declarations, packing lists, certificates of origin. They come from carriers, agents, suppliers and authorities across dozens of jurisdictions, and nobody in the chain has authority to standardise them.
So a meaningful share arrive as scans of faxes, photographs taken on a phone, or PDFs whose text is an image. A pipeline that handles clean digital documents and treats the rest as an edge case will work on the pilot set and fail on the real one, because the real distribution is dominated by exactly those cases.
The design rule is to assume image extraction from the start and test it in week one on documents pulled at random from last month rather than on a set somebody chose. This single decision separates logistics projects that ship from those that discover an extraction problem in month three.
Exceptions are the work
Most shipments proceed without intervention. The hours go into the ones that do not: a mismatch between the declaration and the manifest, a missing certificate, a discrepancy in weights, a code that does not exist.
This is a strong fit for current technology because the exceptions repeat. A team resolving them is applying a small number of patterns over and over, and those patterns are recoverable from the record of what they did. Volume is high, structure exists, and being wrong is recoverable because a person is reviewing anyway.
It is also where the ground truth is. Resolved exceptions with the resolution recorded are kept, because disputes happen and because regulators ask. Few sectors hand over an evaluation set as readily, and almost nobody in the industry thinks of that archive as one.
What the operations floor will and will not tolerate
Adoption here has a distinctive shape. Operations teams work to a rhythm set by the physical world, and a tool that adds a step during a peak is abandoned within a shift regardless of how good it is.
The systems that stick appear inside the tool people already have open, answer in the time it takes to glance at a screen, and are silent when they have nothing useful to say. A system that interrupts with a suggestion during a cut-off window is worse than no system, and the team will say so quickly and accurately.
The compensating advantage is that feedback here is fast and honest. An operations floor tells you within a day whether something works, which is a shorter loop than stage four usually offers, and it is worth designing the rollout to take advantage of it.
