Skip to content

AI Agents in Logistics: Shipping, Warehousing, Customs and Planning

Most manufacturers do not run logistics as a separate business with its own software stack. Shipping, goods receipt, customs and dispatch planning are spread across a handful of people who move between the ERP, carrier portals, email and spreadsheets all day. This guide looks at logistics from that shipper's point of view: where an AI agent helps, where transport and warehouse systems remain the better tool, and how to begin.

What is an AI agent in logistics?

An AI agent in logistics is software that connects a language model to tools. It reads emails, delivery notes, carrier portals and ERP records, matches them up and completes narrow tasks such as writing a status reply or preparing a carrier booking. When something is unclear, it asks or hands over to a person.

What sets it apart from conventional automation is how it copes with messy input. An ERP rule can post a delivery when the order number sits in the right field. An agent can find the order number on a supplier's PDF delivery note, wherever that supplier happens to print it this time, and still prepare the match.

The perspective matters. Carriers and third-party logistics providers optimise routes, load space and cross-docking. A manufacturer has different concerns: getting its own goods to the customer on time, with the right paperwork, at a sensible cost, and booking incoming goods correctly. Those concerns are the focus here.

How does a logistics agent work?

A logistics agent has four parts: a language model that understands text and documents, tools that give it access to the ERP, mailboxes and portals, rules about what it may do on its own, and a log of every action. The tools are what make it useful. Without read access to orders and deliveries, it is only a text generator.

In practice:

  • Inputs: shared mailboxes for shipping, customs and purchasing; carrier portals; EDI messages; scanned delivery notes and consignment notes.
  • System access: reading sales orders, deliveries, purchase orders and item master data; writing only where you explicitly allow it, for example as a draft or a note.
  • Knowledge base: customer routing guides, packing instructions, delivery terms, internal work instructions and past decisions.
  • Approval rules: what the agent may close on its own (a status reply based on tracking data), what it only prepares (a customs line) and what it never touches (postings with financial impact).
  • Audit trail: every reply, proposal and source is stored so it can be traced later.

The model can run in the cloud or on your own servers. If customs and customer data must stay in-house, a locally hosted model is an option, described on the page about sovereign AI.

Which logistics tasks can an agent take over for a shipper?

Plant logistics breaks down into five areas: outbound shipping, goods receipt, customs and dangerous goods, dispatch planning, and returns. Each involves a lot of reading and cross-checking that an agent can prepare. How far it should act on its own depends on how costly a mistake would be in that area.

Outbound shipping: status requests and carrier bookings

Two tasks keep coming back. The first is "where is my order?" from customers and the inside sales team. An agent can pick the order number out of the request, find the delivery in the ERP, fetch the status from the carrier's portal or tracking interface and draft a reply, or send it directly when the situation is clear-cut. The second is booking the carrier: deliveries, packages, weights and delivery terms become a draft booking for the portal or an email to the forwarder.

The delivery term is more than a text field. The ICC's Incoterms rules, currently Incoterms 2020 with eleven terms, set out which tasks, costs and risks fall on seller and buyer. An agent can check that the term on the order matches the planned booking and flag mismatches, rather than quietly booking freight at your expense that the buyer should have arranged.

Goods receipt: delivery note against purchase order

In receiving, the value lies in matching. The agent reads the delivery note, whether PDF, scan or EDI advance shipping notice (the technical side is covered in documents, PDF and OCR for AI agents), links it to the purchase order and compares items, quantities and dates. If everything matches, it prepares the goods receipt; if there is a short delivery, a wrong item or a missing certificate, it raises a query for purchasing. The posting itself stays in the ERP and is confirmed by a person.

Customs and dangerous goods: prepare, do not declare

With few exceptions, exports from the EU require an electronic declaration. In Germany, according to German Customs, that runs through the ATLAS export procedure, either via certified participant software or an online declaration. An agent does not replace that channel. It fills the gap before it: collecting and sanity-checking goods description, commodity code, origin, weights and values.

Commodity codes follow the Combined Nomenclature. The European Commission describes it as an eight-digit code whose first six digits follow the international Harmonized System, and it is revised every year. An agent can propose a code with its reasoning, based on the item description, drawings and past classifications. Responsibility for the classification stays with your customs team or broker.

For dangerous goods the line is tighter still. The ADR covers classification, packaging, marking and documentation of dangerous goods by road; the German transport ministry notes that the latest international amendments took effect on 1 January 2025, with the next set announced for 1 January 2027. An agent can check whether an item is flagged as dangerous goods and whether the required details appear on the shipping order. Classification and the dangerous goods safety adviser's responsibilities are untouched.

Dispatch planning: what has to ship today?

In dispatch planning, an agent helps with the overview. It gathers due dates from open orders, completion messages from production and pickup confirmations from carriers, and produces a morning list: what is ready, what is running late, where a pickup is still missing. It can draft chasers to carriers or delay notices to customers once someone approves the wording.

Route planning and load optimisation are a different matter and belong in specialised software. A language model is not an optimisation solver. At most, the agent sits where results are explained or passed on to the people involved.

Returns: triage and follow-up

Returns arrive through many channels: complaint emails, web forms, handwritten notes on a delivery note. An agent can read the reason for return, find the original order, check whether a return authorisation exists and route the case to quality, sales or the warehouse. Credit note, replacement or repair is decided by a person.

Agent or TMS/WMS: where is the line?

Transport and warehouse management systems handle structured processes: freight rates, route planning, storage locations, picking and stock. An agent does not replace them. It works at the edges, where information arrives unstructured in emails, PDFs and portals, and turns it into input for those systems.

A simple rule of thumb:

TaskTMS/WMSAgent
Compare freight costs, plan routesyesno
Storage locations and pickingyesno
"Where is my order?" requestsupplies the datawrites the reply
Read and match a PDF delivery noteonly with EDIyes
Gather and justify customs datasubmits the declarationprepares the proposal

If you run neither a TMS nor a WMS and rely on ERP modules and manual work, an agent can help early without a large system rollout first. If you already run a TMS or WMS, check which AI features your vendor offers before building your own agent. The decision usually comes down to how much of the work happens outside that system.

What does a logistics agent need?

A logistics agent needs three things: access to the data your team works with today, a set of documented examples for each use case, and one person who owns what the agent is allowed to do. Perfect master data is not a precondition, but missing weights and commodity codes quickly limit what it can achieve.

Data and systems:

  • Read access to orders, deliveries, purchase orders and item master data in the ERP, through an API or a database export.
  • Access to the relevant mailboxes and, where available, to carriers' tracking interfaces.
  • Examples: completed cases with their outcome, such as answered status requests or approved customs lines.

Organisation:

  • A subject-matter owner from shipping or customs who sets the rules and reviews results in the early weeks.
  • A clear list of which steps the agent may complete without approval.
  • Where works councils or employee representatives exist, early involvement if the agent logs how staff handle cases.

Regulation: paper is slowly leaving freight transport. According to the European Commission's page on the eFTI Regulation, from 9 July 2027 authorities in EU member states must accept freight information shared electronically through certified eFTI platforms, covering road, rail, inland waterways and air. The more transport data arrives in structured form, the easier it is for an agent to read and check it.

How do you introduce a logistics agent step by step?

Start with one use case that is frequent, low-risk and easy to measure, usually shipment status requests or delivery note matching. Widen the agent's permissions only once results have been stable for several weeks.

  1. Count the work: how many status requests, receipts, exports and returns arrive each week, and who handles them?
  2. Pick the first case: frequent, explainable by rules, errors easy to spot.
  3. Connect the data: read-only, and only the systems that case needs.
  4. Shadow mode: the agent drafts, your team works as before and compares.
  5. Approval mode: the agent proposes, a person approves or corrects with one click. How to design these approval levels is explained in human-in-the-loop for AI agents.
  6. Partial autonomy: clear-cut cases are closed by the agent, everything else goes to people.
  7. Next area: only when the first case is stable, move on, for example from status requests to customs data.

In our own operations, collecting the corrections from steps 4 and 5 has proved the most useful habit. They are the raw material for improving the agent's rules and examples, and they show honestly where it is not ready yet.

What are the limits and risks?

A logistics agent can misread, mismatch and phrase something wrong very convincingly. That matters most where documents have legal effect: export declarations, dangerous goods paperwork, proofs of origin. In those areas the agent stays an assistant, and every output needs a traceable source.

Common risks and how to handle them:

  • Wrong match: similar order numbers or consolidated shipments lead to the wrong delivery. Have the agent check several attributes (customer, item, quantity) and hand over on any mismatch.
  • Invented details: a language model can produce a plausible commodity code that is simply wrong. Always require reasoning and a check against the current tariff.
  • Outdated rules: the tariff, the ADR and customer routing guides change. The knowledge base needs an owner and a review date.
  • Privacy and confidentiality: delivery addresses, contact details and prices are personal or confidential data. Set the system up so it can be operated in line with the GDPR, with clear access rights and logs.
  • Portal dependency: when a carrier redesigns its portal, browser-based access breaks. Use official APIs wherever they exist.

An agent is a poor fit when your flows already run fully structured over EDI and the ERP handles them cleanly. In that case it adds little beyond handling the exceptions.

How do you get started?

The quickest start is a week of counting: which logistics tasks occur, how often, through which channel and how long each takes to handle. That almost always reveals a first candidate you can test in shadow mode without touching existing systems.

It helps to look at other departments too. Many building blocks, such as reading a mailbox, querying the ERP and drafting replies, are the same as for agents in procurement or a B2B customer service agent. What changes in logistics is the addition of carrier portals, customs data and dangerous goods details, and the fact that many documents carry legal weight. For an overview of agents by department and how they fit together, see AI agents for companies; more articles are in the guides section, and terms such as tool use or shadow mode are explained in the glossary.

If you would rather frame the topic with your team first, a workshop where you walk through your own logistics tasks and agree on the first use case is a practical way in. After that, the question is no longer whether AI helps in logistics, but where it makes sense to start.

Frequently asked questions

Do we need a transport management system before an agent makes sense?

No. Shippers without a TMS often run logistics through email, carrier portals and spreadsheets, which is exactly where an agent can help. If you do have a TMS or WMS, the agent works through its interfaces rather than duplicating it.

Should an AI agent file export declarations on its own?

It could be wired up, but it should not. The declarant carries the liability, so the agent prepares commodity code, origin and values as a reasoned proposal, and a trained person reviews and submits through the certified customs software.

What data does a logistics agent need to get started?

Usually a mailbox or portal login, read access to orders and deliveries in the ERP, and a few dozen completed cases as examples. Clean item master data such as weights, dimensions and commodity codes makes shipping and customs work far more reliable.

Can an agent replace the dispatcher?

No. It collects due dates, readiness and carrier confirmations and suggests what needs to ship today. Sequencing, partial shipments and expedited runs remain the dispatcher's call.

Sources

  1. European Commission: eFTI Regulation (electronic freight transport information) transport.ec.europa.eu
  2. German Customs (Generalzolldirektion): IT procedure ATLAS export zoll.de
  3. European Commission: Combined Nomenclature taxation-customs.ec.europa.eu
  4. German Federal Ministry of Transport: Dangerous goods law, road (ADR) bmv.de
  5. ICC: Incoterms rules iccwbo.org