AI in logistics: nine examples and what they do not solve

AI in logistics is software that learns from data or can read text, and uses that to take over part of the thinking in the warehouse or on the road: a forecast, a layout of locations, a route, an order pulled out of a PDF. The nine AI applications in logistics in use now are demand forecasting, slotting, order picking with robots, route planning, planning flexible staff, document processing, order processing from email and PDF, invoice checking and customer queries. Each comes with what you need for it and what it usually does not solve.

Where is AI in logistics already used today?

AI in logistics runs in two places today: in the planning and the machines of warehousing and transport, and in the administration around them. The first part is the best known. The second is less visible, and there people retype orders and invoices every day.

According to Eurostat, almost 20 percent of European enterprises with ten or more persons employed used some form of AI in 2025. Logistics was the least common purpose: just over 6 percent of enterprises that use AI apply it there. Enterprises that considered AI and still did not start name lack of expertise as the first reason. More than 70 percent of that group says so.

The examples below follow the chain. Demand forecasting, slotting, robots, routes and staff planning are usually built by specialised software vendors. Documents, orders, invoices and customer queries sit in the administration around the WMS, and that is where Pakr builds.

How does AI help with demand forecasting and inventory management?

AI for demand forecasting works out from sales history, seasons and promotions how much of an item you will need in the coming weeks, so your inventory management runs less on gut feeling. A good model sees that an item peaks every year in the second week of December, and that a promotion by the client causes a dip the week after.

You need clean history per item for that, ideally over a few years. The promotions also have to be recognisable in it. In fulfilment that knowledge often sits with the client. The web shop knows when the promotion starts. You only see it when the orders come in.

A forecast stays an expectation with a margin. A new item without history, or a client running a promotion without telling you, is something no model sees coming. For a logistics service provider the stock usually belongs to the client, and so does the decision to order. The forecast then mainly helps you with staffing and space in the warehouse. This kind of software usually sits in a planning package or in the client’s ERP. In contract logistics with a few large clients who do their own planning, a second forecast on your side often adds little.

What does AI do for slotting and order picking?

In slotting, AI decides where an item sits in the warehouse, and in order picking it helps decide in which order and by whom or what it gets picked.

Slotting

Slotting is the allocation of items to locations. Software looks at which items often appear together in an order and how fast they move, and suggests what should sit closer to dispatch. You need order lines with location and time, the dimensions and weights of your items and a WMS that supports re-slotting.

The advice is as good as the master data. If an item’s weight is wrong in the system, the suggested location is wrong too. And every move costs hours on the floor. So you plan a re-slot in a quiet week, and certainly not in November.

Order picking and robots

In order picking, the most visible part of warehouse automation, you see AI in picking robots, in systems that bring the rack to the picker and in cameras that check whether the right box was picked. That comes with an investment that runs over years, a building that suits it and an order profile that does not change every quarter.

For contract logistics with contracts of a few years, that is a hard sum. The robot has to last longer than the client it was installed for. Whether it pays off, you work out with the term of your contracts, the use of the building outside the peak and the cost of maintenance over the whole period it stands there. The speed in a demo says little about that. And a robot does not fix an order that is wrong in the WMS. It picks exactly what it is asked for, even when that is wrong.

Can AI take over your route planning?

AI in route planning works out which stops go in which order in which vehicle, with time windows, load space, driving times and traffic taken into account. A planner approves the result or moves a few things.

You need correct addresses and time windows, the dimensions and weights of what has to go and a current picture of which vehicles and drivers are available. This is a mature market with specialised vendors, often as part of a TMS.

A planning package works with what it receives. If a shipment comes in as an email with the delivery address somewhere in the text, someone has to enter it first. The planner is still needed for what is not in the data: the customer who never opens before ten, the tail lift that often breaks down. Pakr does not build route planning. What we do build is the step before it: getting the shipment out of that email and into your system.

Can AI plan your flexible staff?

AI can propose the number of flexible workers per day and per shift, based on the expected order volume and the average time each task takes.

You need an order forecast and standard times per task, such as picking, packing, loading and processing returns. To test those standards you use the hours from your time and attendance system. The proposal then goes to your staffing agency.

Who actually turns up, such a proposal does not say. Sickness, an agency that supplies fewer people than requested and a client who announces an extra container on Thursday stay work for your planner.

Mind the rules as well. According to the European Commission, AI systems for employment and the management of workers fall into the high risk category under the AI Act. If you let a system assess or assign people, find out first whether it falls under that. What the act means for a warehouse is in the AI Act in logistics. Pakr does not build staff planning.

Where does AI help in the administration around the WMS?

In the administration around the WMS, AI reads the documents, orders, invoices and emails that are now copied over by hand. It then puts the result ready in the system you already have. That needs no machine and no rebuild.

The four examples below share the same setup. The system reads the item and puts the result ready. What it is not sure about goes to a queue where a person looks at it. Before you start, you measure how much time the process takes now and how often it goes wrong. After a few weeks you can see whether the difference is real.

Document processing

Document processing means reading consignment notes, CMRs, packing slips and customs papers. A language model reads a scan or PDF roughly the way a person does: it finds the shipment number even when every sender puts it somewhere else. You need examples of every layout you receive, and an agreement on what happens when in doubt. A blurry photo of a crumpled consignment note stays hard, and a missing signature does not appear by itself. More on this in document processing in logistics.

Orders from email, Excel and PDF

In order intake, the system reads an order from an email, Excel file or PDF, checks it against the agreements with that client and puts it in the WMS. You need each client’s item codes, a connection to your WMS and a queue where doubtful cases land. An order the client passes on wrongly goes in just as wrong. How that works is in orders from email into your WMS.

Checking invoices

AI reads the invoices of carriers and staffing agencies and sets each line against what your own systems say happened: the shipment in the TMS, the hours in the time and attendance system. You need the rates and agreements, plus the data from your systems over the same period. The system does not decide who is right. It puts the lines that differ at the top and someone in accounts payable looks at them. With staffing agencies it is often about allowances and breaks that differ between the invoice and your clock. With carriers it is surcharges for fuel, waiting time or an extra stop you never ordered.

Customer queries

For customer queries, AI reads a client’s email, looks up the status of the order or shipment in the WMS or TMS and drafts a reply. You need access to those statuses and a clear line for what the system may answer on its own. A complaint about a damaged pallet or a question about a rate goes to a person. Start with the questions your customer service gets most often now and where the answer is literally in your system.

Why does the administration around the WMS cost a 3PL so many hours?

The administration around the WMS costs a 3PL many hours because every client, carrier and staffing agency delivers its own format, while the WMS only works with what is entered cleanly.

One client sends EDI. The next sends an Excel file with its own columns, the third a PDF from an old system. Carriers each invoice in their own way, and so do staffing agencies. What sits in between is done by a person: retyping, checking, calling back. That work is in no planning and in no quote, and it grows with every new client. A client who sends orders as PDFs from a twenty year old system brings its own retyping work with it.

Reading text that looks different every time and turning it into fields is what current AI is good at. Pakr builds this into the WMS, TMS and ERP you already have, one process at a time. First we measure how it goes now, and the next step only comes once this one works. What we build and what we do not is on what we do.

Which example fits your company?

That depends on where the hours and the errors are now. If your office retypes orders from email, Excel and PDF every day, start with order processing. If money leaks away on carrier and agency invoices, start with invoice checking. If your contract logistics runs on a few large clients who do their own planning, a second demand forecast adds little.

Not sure where the hours are? For one week, count how many orders someone retypes, how many consignment notes are entered by hand and how many invoice lines get checked. That number is your baseline. With that count and a few of those documents, you can decide in an introductory call which process goes first.