How to use AI in logistics: what to automate first
In logistics, AI pays off fastest on documents and messages: orders received by email, CMRs, PODs and carrier invoices, extracted and checked against the order automatically, with a person reviewing the exceptions. Route planning and driver scoring are not the first step. Start with one flow, measured before and after, in a pilot.
Transport and logistics companies share one problem: good people spend a lot of time copying data from one document into another system. Orders arriving by email, scanned CMRs, carrier invoices, "where is my shipment?" questions. This guide is for the operations director or owner who wants to know what AI can do in practice, what it cannot, and where to start.
How do you use AI in logistics and transport?
You use AI where the work is repetitive, starts from a document or a message, and ends as a record in the TMS, WMS or ERP. In short, AI reads, extracts, compares and prepares. The dispatcher or operator checks and decides. The table below sums up the problems that show up in almost every logistics company.
| Problem | What AI can do today | What it does not do well |
|---|---|---|
| Orders arriving by email, PDF, Excel | Reads the message, extracts the data and drafts the order in the TMS | Cannot guess a field the order leaves out |
| CMRs, delivery notes, PODs, freight invoices | Extracts the fields and checks them against the order and the agreed rate | Cannot reliably read a faded scan or a signature |
| "Where is my shipment?" and exceptions | Answers from the TMS and flags delays | Cannot promise a time the system does not know |
| Damage and claims | Classifies photos and prepares the claim file | Does not decide liability |
| Route planning and forecasting | Explains plans and prepares data for optimisation | Does not replace a route optimiser |
1. Orders that arrive by email
Many customers send orders however they like: a free-text email, a PDF from their own ERP, a spreadsheet with 40 lines. Someone keys them into the TMS by hand. A language model reads these formats without a fixed template, pulls out addresses, loading dates, cargo type and pallet count, and drafts the order. The operator confirms it with one click or corrects it.
What it cannot do: fill in a field that is not in the message. If the unloading time is missing, the system has to ask, not invent it. You design for this from day one: every field carries a confidence score, and uncertain fields go to a person.
2. CMRs, delivery notes, PODs and carrier invoices
This is the highest-volume document flow. AI extracts the fields from a CMR or a freight invoice and checks them against the order and the rate agreed with the carrier. Differences (an extra surcharge, a different quantity, a missing POD) land on someone's list; the rest goes through. We describe the full pipeline in how to automate document processing with AI.
What it cannot do: reliably read a document photographed at an angle in a truck cab, with the stamp over the text. Source quality matters more than the model. Often the first gain comes from asking drivers for a better photo, not from a bigger model.
3. Customer questions and delivery exceptions
"Where is my shipment?" is the most common question a dispatch desk gets. An AI agent connected to the TMS can answer by email or chat with the real status and flag exceptions: a delay, a refused delivery, a wrong address. More on agents on our custom AI agents page.
What it cannot do: promise a time your system does not know. If the TMS does not receive the truck's position, the agent has no way of knowing where the goods are. An agent is only as good as the data it can reach.
4. Damage and claims
At goods-in, someone photographs a damaged pallet. A computer-vision model can classify the type of damage, and a language model can assemble the file: the photos, the CMR with its remarks, the order, the correspondence. The person handling the claim starts from a complete file, not from scratch. The visual side is covered in how to use computer vision to automate inspection.
What it cannot do: decide who is liable for the damage. That stays a human decision, based on the contract and the documents.
5. Route planning and forecasting
This is where most promises are made and where you should be most sceptical. Route optimisation is a mathematical problem that specialised software has solved for years. A language model does not solve it better. What AI can do: gather the constraints from emails and orders into a form the optimiser can use, and explain the plan in plain language. Demand forecasting needs clean history over at least a few seasons; without it, any model is guessing.
What should a logistics company automate first?
Start with a document flow, not with planning. The reasons are practical: the volume is high, the rules are clear, today's cost is easy to measure (minutes per document, errors per month), and a person already checks the result, so the risk is low. The order we usually recommend:
- Orders received by email, if you re-key them into the TMS.
- Carrier invoices checked against the order and the agreed rate, if you work with many subcontractors.
- CMRs and PODs, so you can invoice sooner after delivery.
- The customer-question agent, once the TMS data can be trusted.
A hypothetical example: a freight forwarder receiving a few hundred orders a month by email would start with one large customer and one order format. It measures today's time per order, runs the pilot for a month with operators checking every draft, then compares. That is the only way to learn the real saving: measured in your own pilot. The steps of a pilot are in how to run an AI pilot.
What to leave for later: any system that scores drivers or allocates work based on their behaviour. Beyond how sensitive it is for people, it is an area with strict rules under the AI Act (see below).
Data and compliance: GDPR, the AI Act and transport rules
A few things to know before the first pilot. This is not legal advice; for your case, talk to your lawyer or DPO.
- GDPR: drivers' names, phone numbers, number plates linked to a person and location data are personal data. A pilot on real documents needs a legal basis, restricted access and a clear retention rule. More in how to stay GDPR-compliant when using AI.
- AI Act: extracting data from documents and answering order-status questions are not, as a rule, high-risk systems. Systems used to monitor and evaluate workers' performance, or to allocate tasks based on their behaviour, are listed in Annex III (Regulation (EU) 2024/1689), with obligations that apply from 2 December 2027 after the deferral in Regulation (EU) 2026/1744.
- Transparency: if an AI agent talks directly to your customers, they have to know they are dealing with an AI system. The obligation (Article 50 of the AI Act) has applied since 2 August 2026.
- Public systems: if you file in Romania's e-Transport system or invoice through RO e-Factura, AI can prepare the data, but submission stays a controlled step in your own system, with a person in charge.
How do I start?
- Tech Call, free: a short conversation where you show us the flow and we tell you honestly whether it is worth automating. Book it here.
- Tech Audit: a short, paid audit where we look at real documents, your TMS and your ERP, and give you in writing the use case, the data it needs and what "working" means in numbers.
- Pilot: one flow, on real data, with your people checking every result. We measure against the audit's numbers.
- Production: integration with the TMS, WMS or ERP, monitoring, and a person approving what is sensitive. The code and documentation stay with you.
What we can show: at Sapio we built ai-aflat.ro, the AI assistant over 220,000+ Romanian legislative acts, updated daily. It is not a logistics project, but it is the same kind of work: a large corpus of mixed documents, read correctly, with the source cited. What we build for operations is on the AI automation page, and every published project is on the projects page.
Sources
- Eurostat, enterprises using AI by economic activity (isoc_eb_ain2), reference year 2025
- Eurostat, statistical report KS-01-26-009, published 26 March 2026 (Table 8, reasons for not using AI)
- Regulation (EU) 2024/1689 (AI Act), Annex III and Article 50
- Regulation (EU) 2026/1744, the new dates for high-risk systems
In 2025, only 11.2% of EU transport and storage enterprises used AI, and 69.1% of those that considered AI and did not proceed named a lack of expertise (Eurostat, isoc_eb_ain2 and KS-01-26-009, Table 8).
Frequently asked questions
Can AI read scanned or photographed CMRs?
Yes, if the image is legible. Modern models read CMRs without a fixed template and extract the sender, consignee, goods and remarks. Problems come from skewed photos, stamps over the text and hard-to-read handwriting. That is why every field carries a confidence score and anything uncertain goes to a person.
Do I need to replace my TMS to use AI?
Usually not. AI connects to the TMS, WMS or ERP through an API, a file import or, when needed, automation over the existing interface. What matters is that the system holds the data the agent needs, such as delivery status. If the data is not in the TMS, no AI can invent it.
How much will I save by automating transport documents?
We do not publish a percentage, because it would be an invented number. The saving depends on volume, on how varied the documents are and on how much time they take today. The only honest figure is the one measured in your own pilot: time per document and errors before and after, on the same flow.
Can AI plan routes better than my planning software?
A language model cannot. Route optimisation is a mathematical problem solved by specialised software. AI helps before and after: it gathers constraints from orders and emails and explains the plan in plain language. If someone promises better routes from a chatbot alone, ask how they measure it.
Is an AI system that scores drivers high-risk under the AI Act?
Systems used to monitor and evaluate workers' performance, or to allocate tasks based on their behaviour, are listed in Annex III of the AI Act, with obligations from 2 December 2027. That means documentation, human oversight and informing workers. This is not legal advice: the exact classification depends on the specific system.
Want to discuss a project?
Book a free discovery call with the Sapio team.