AI for Logistics and Shipping

In logistics, the most common setting error is asking a generative AI to schedule turnarounds: it is a computing problem that solves dedicated algorithms better, not a language model. The true value lies in the words around the shipment: the note written by hand in the order, the phone call of those who ask where the package is, the reserve to be noted when a neck arrives bruised.

They are tasks of interpretation and documentation, not calculation, and that is where a system that reads and writes in natural language makes what a routing algorithm cannot give. It is worth saying immediately because it is the first misunderstanding that derails a project: those who propose it thinking about the routes are disappointed, those who build it thinking about the words around the expedition find a solid use case from the first month.

The document accompanying the goods, and what it must contain by law

Those who deliver goods without immediately issuing an invoice for each delivery rely on a specific document, born to replace the old accompanying note. It must contain the date, generality of the transferor and the transferee, any person in charge of transport, and the description of the nature, quality and quantity of the goods sold, as established by the Article 1 of Presidential Decree 472/1996. It does not have a mandatory pre-printed model: the same standard admits it in free form, provided it is issued before the start of transport.

This document is not only used to make the goods travel in order: the same rule makes it suitable to overcome the tax presumptions on disposals and to determine how and for what it should be kept, according to the general rules on the preservation of invoices. A system that digitizes transport documents, therefore, is not only speeding up logistics: it is also building the proof that serves to the deferred billing described in AI and Electronic Billing.

The written reservation on delivery: why it matters more than a phone call after

The carrier shall be liable for the loss and damage of the goods from the time it receives them to the time when they return them, unless it proves that the damage results from fortuitous occurrence, the nature of the goods or the behaviour of the sender or recipient, as determined by the Article 1693 of the Civil Code. The same rule adds a detail that changes a lot in practice: if the carrier accepts the goods unreservedly, it is assumed that the packaging had no apparent defects.

This means that a reservation written on the document at the time of delivery, even just "stained neck" or "wet packaging", is worth much more than a phone call made the next day when the neck has already been opened. An assistant who, reading the signed transport documents, automatically recognizes the presence of a reserve and connects it to the correct shipment turns a pen annotation difficult to find into a given consultable months later, when needed for a dispute.

Where Mathematics Ends and Interpretation Begins

Does the customer see an event or a promise?

When the shipping information arrives at customer care, it must remain clear what is verified. The AI assistant for order status connects source, update and affected lines, distinguishing partial delivery and full order. If the data is old or discordant, prepare the switch to the team without inventing an explanation of the delay.

The exception seen before the complaint

In the traditional flow a delivery problem is discovered when the customer writes. At that point it has already passed time, the customer is already unhappy, and the conversation starts uphill.

1. The system reads the states transmitted by the courier continuously.
2. Recognizes abnormal states: absent recipient, incorrect address, stock, reported damage.
3. Classify the type of problem and its urgency.
4. Notify the customer with a specific message for that type of anomaly, not a generic alert.
5. It opens an activity for those who have to intervene, with the context already ready instead of a naked signal.

The order notes: the field that no one reads again

There is a field, in almost every management software, where customers write important things that then remain there: preferred delivery times, indications for the intercom, limited traffic areas, floors without elevator.

• Extract time slots and days indicated in the free text
• Recognize recurring access constraints and limitations
• Report ambiguous notes requiring a human decision before shipment
• Pass to the courier the information in the format that his system accepts

It is an inconspicuous use case, because it does not produce a number to show in a presentation, but with a concrete return: every delivery that fails for an indication that there was but did not arrive to the courier costs a second trip, in addition to the discomfort of the customer who had written it thinking it was enough.

What to count before deciding if the project really serves

In one example: An online store with a steady stream of shipments receives more weekly order status requests than it receives on any other theme. An assistant connected to the management software and the courier absorbs the majority of these requests, and the time released is immediately seen on those who respond by hand today.

A B2B company with deliveries by appointment has a different problem: not the status of the shipment, but the coordination, time slots, contact persons on site, unloading constraints. Here the value lies in the interpretation of notes and in a proactive notice the day before, not in the answer to a question that the customer almost never asks.

The most honest way to choose between the two, or to understand that neither applies, is to count for a week the after-sales requests that really arrive and classify them broadly: order status, changes, delivery issues, other. If a category dominates clearly, the project has a precise direction from day one.

There is also a third case, less told but as frequent as the other two: the distributor who works with multiple different couriers depending on the area or the customer. Here the use case is clear from the start, but the cost grows with each extra courier to connect, because each exposes its own states with a different level of detail. The practical choice is to start with the two couriers that cover most of the volumes and treat the others as a manual exception until they are worth integrating as well.

Why an assistant on shipments almost always jams the same way

An assistant who answers on the status of shipments seems a simple project, and almost always fails for one of three recurring reasons, which is worth knowing before leaving rather than discovering them in production.

The first point is deceptive because the error is not obvious: a slightly misaligned response to reality is worse than no response, because the customer trusts and then discovers that it was not up to date. The second is the one that turns a useful assistant into an obstacle: those who have a delivery problem and find no way to talk to someone do not become a satisfied customer just because the automatic message was well written. The third one looks like a copywriting detail and is not: an "absent recipient" exception requires action other than an "incorrect address" exception, and a message that treats them the same way still generates a phone call, that is, exactly the work that you wanted to avoid.

The account that must include failed deliveries

The calculation of convenience for logistics has a peculiarity: in addition to the time of the people, there is a direct and measurable cost in failed deliveries, which involve a second trip and sometimes a return. If any of these depend on information that existed in the order notes but did not arrive at the courier, that part is recoverable with the work described above.

Before estimating the benefit, it is best to really count the failed deliveries of a recent period and look at their causes one by one. It often emerges that the main problem is not the missing information but something else, an unreliable courier on certain areas for example, and in that case no system of interpretation of the text solves it: the general calculation method, however, remains that described in how much does an AI agent cost.

Finally, there is an effect worth anticipating instead of finding out in the race: in the early days a system that communicates exceptions proactively increases conversations with customers, not diminishes, because problems that previously remained invisible now generate a message and often a response. It's not a sign of failure: the conversations are shorter, they start from a better position because it was the company that warned first, and they close with a solution instead of a complaint.

The connection to the warehouse from which the goods depart is natural, and is described in AI for stock and inventory management; the one with the suppliers that deliver to you, in the other direction of the chain, is in AI for procurement and supplier management. If you work in production or distribution and want to figure out which shipping flow to start from, Giallo Studio helps to automate communications and exceptions before even touching the warehouse.