AI for Logistics Companies: Freight Documents, Customs, Exceptions and Shipper Data That Stays Confidential

How is AI used in logistics? A map by type of company

Logistics is not one business, and the work AI can take on differs by role. The table below sorts the common text and document tasks by who usually does them.

This guide covers the first column, where generative AI and agents are new. The second column is established territory for optimisation and forecasting software and does not need a language model.

Freight document processing: bills of lading, commercial invoices and CMR notes

Every shipment arrives as a bundle of documents in different layouts from different senders: commercial invoices, packing lists, bills of lading and sea waybills, air waybills, CMR consignment notes for international road haulage, certificates of origin and dangerous goods declarations. Someone keys the relevant fields into the transport management system and checks that they agree with the booking.

Modern models read scanned and digital documents well enough to extract parties, addresses, goods descriptions, HS codes where present, piece counts, weights, volumes, Incoterms and reference numbers into a fixed schema. The useful part is the cross-check: a model can compare the invoice, packing list and bill of lading for the same shipment and list every mismatch in weight, quantity or consignee for a person to resolve before it becomes a customs problem.

Three design rules keep this reliable. Extract into a defined schema rather than free text. Return the page and position for each field so an operator can verify it in seconds. And route low-confidence fields to a person rather than guessing. Our freight document processing use case describes that intake workflow in more detail.

Customs classification support and declarations, with the broker in charge

Customs work is where AI can save the most time and where errors cost the most, so the human role has to be explicit.

What a model does well. It can propose candidate tariff headings from a goods description, cite the heading text and explanatory notes it relied on, retrieve the company’s own past classifications and any binding rulings for similar items, flag descriptions that are too vague to classify, and assemble the declaration data set from the shipment documents.

What stays with people. In the United States, the importer of record must use reasonable care when making entry under 19 U.S.C. 1484, and licensed customs brokers file on importers’ behalf. In the EU, Article 15(2) of the Union Customs Code currently in force, Regulation (EU) No 952/2013, makes whoever lodges a customs declaration or an entry summary declaration responsible for the accuracy and completeness of the information in it. A model’s proposal is input to that judgement, not a substitute for it. The customs and trade compliance workflow shows how classification suggestions, evidence and broker approval fit together.

EU ICS2 release 3

The EU’s Import Control System 2 requires an entry summary declaration with safety and security data before goods arrive. Commission Implementing Decision (EU) 2023/2879 rolled out release 3 in three steps: maritime and inland waterway carriers from 3 June 2024, house-level filers in maritime traffic such as freight forwarders from 4 December 2024, and road and rail carriers from 1 April 2025, each with a deployment window that ended by 1 September 2025 at the latest. The Commission’s ICS2 page says that from 1 June 2026 every consignment entering the EU by any means of transport should have a valid ENS, filed directly in ICS2 or combined with a transit declaration in NCTS Phase 6 where a country has opted in.

ICS2 puts weight on goods descriptions and party data. The Commission publishes a list of stop words for ENS data, most recently updated with effect from 3 August 2026, and warns that customs may impose administrative sanctions for non-compliance with data requirements. Drafting precise descriptions from invoices and packing lists, and flagging vague ones before filing, is a well-suited task for a model with a human check.

The 2026 Union Customs Code reform

The European Parliament and the Council adopted the reform of the Union Customs Code on 16 September 2026, as Regulation (EU) 2026/2108. It creates an EU Customs Authority in Lille that will manage an EU Customs Data Hub. According to the Commission, use of the data hub to record imports and exports becomes mandatory for e-commerce on 1 July 2028 and for all traders from 1 March 2034, with voluntary use possible from 2031. The direction is clear: more structured data, submitted once, and more scrutiny of its quality. Firms that clean up their document data now will find the transition easier.

US broker confidentiality

Under 19 CFR 111.24, records pertaining to the business of a broker’s clients are confidential. A broker must not disclose their contents or related information to anyone other than the client, the client’s surety on a particular entry and authorised government officers, except by subpoena or court order or with the client’s written authorisation. Sending client invoices and entry data to an outside AI service is a disclosure question that brokers should resolve with clients in writing before it happens.

Exception and disruption management

Exceptions are where coordinators spend their day: a vessel rollover, a missed truck slot, a customs hold, a temperature excursion, a short delivery. Each one means reading the notice, working out which bookings and orders are affected, finding options and telling several parties.

An agent can take the first pass. It reads the carrier or terminal notice, queries the transport management system for affected shipments, checks alternative sailings, flights or trucks against service commitments, and drafts updates to shippers and consignees for a coordinator to approve. The coordinator decides on rebooking and cost. Over time, the agent’s drafts and the coordinator’s edits become a record of how each type of exception was handled, which helps the next shift. The exception and disruption management use case and our article on agent-driven logistics operations for documents and exceptions go further into the operational design.

Track-and-trace customer service

“Where is my shipment?” remains the most common customer question. A customer service assistant connected read-only to the transport management system and carrier tracking feeds can answer it with the latest milestone, the expected arrival and any open exception, in the customer’s language.

Two limits apply. The assistant must only reveal shipments the customer is entitled to see, which means checking the requester’s identity and account before every lookup. And it should say plainly when tracking data is stale or missing rather than inventing an estimate. Anything that changes a booking, such as a delivery date change or address correction, should create a task for staff rather than happen automatically.

Rate analysis, tenders and carrier contracts

Procurement teams at shippers and logistics providers run tenders with hundreds of lanes and dozens of bidders, each submitting in a slightly different format. A model can normalise bid sheets into one structure, flag missing lanes and unusual accessorial charges, and summarise how each bidder’s offer differs from the incumbent.

Carrier and subcontractor contracts benefit from clause extraction: liability limits, fuel surcharge mechanisms, payment terms, minimum volume commitments, termination rights, and insurance requirements. Putting them into a table across the carrier base shows quickly where terms are out of line. The decisions on award and negotiation stay with the procurement team, and the pricing data involved is commercially sensitive for both sides.

Fleet maintenance knowledge

Carriers that run their own fleets keep maintenance manuals, inspection checklists, defect reports and workshop histories, often across several vehicle makes. A maintenance assistant that answers from those documents helps a technician find the right procedure or the history of a recurring fault. It should cite the manual section and never override manufacturer instructions or the workshop’s safety procedures. Predicting component failure from telematics is a numerical modelling task that sits alongside, not inside, the language model.

AI in logistics and supply chain: the 2026–2027 regulatory frame

Several EU rules are changing the documents logistics companies produce, which affects what AI is asked to read and write.

• ICS2. Release 3 is fully in place for sea, inland waterway, road and rail, and the end of NCTS Phase 6 derogations on 1 June 2026 closed the remaining gaps.
• eFTI. Regulation (EU) 2020/1056 on electronic freight transport information obliges competent authorities to accept regulatory freight information that operators make available electronically through certified platforms. The Commission says this applies in full from 9 July 2027, and that from January 2026 platforms could prepare and authorities could start accepting data stored on certified platforms. Article 6 of the regulation requires authorities, eFTI service providers and operators to protect the confidentiality of commercial information.
• Customs reform. The new Union Customs Code and the EU Customs Data Hub phase in from 2028 to 2034, as described above.

The common thread is structured, accurate data supplied once. AI that turns unstructured shipper documents into clean, verified fields helps with all three.

Shipper data confidentiality decides where the models run

A logistics provider holds other companies’ commercial information: what they buy, from whom, at what price, and where it is going. Service agreements with large shippers often restrict disclosure to third parties and set rules for subcontractors. US broker records carry the statutory confidentiality in 19 CFR 111.24. eFTI adds its own confidentiality duty for freight data in the EU.

Those terms do not forbid AI, but they make the location of processing a contractual question. Before sending shipper documents to an outside model provider, a logistics company should check whether its customer agreements allow that provider as a subcontractor, whether the provider retains prompts or files, and whether it can show a shipper where every copy of its data went. Many firms find it simpler to keep document extraction, customs preparation and customer data on infrastructure they control, and to reserve cloud assistants for internal drafting that involves no shipper data.

AI for logistics companies: choosing tools and platforms

Transport management, warehouse and customs software vendors now ship AI features, and for a firm that runs mostly inside one suite those features are the natural first step. The limits appear when a process crosses systems: a forwarder’s exception affects bookings in the TMS, emails in shared mailboxes, carrier portals and customs filings at once.

Our transportation and logistics solution overview describes how VDF AI handles those cross-system workflows, and the supply chain and operations agent catalogue lists ready-made agents for exceptions, inventory and maintenance. For port and maritime administrations, the maritime authority playbook shows how agentic services can run on-premises. The logistics use case hub collects the full set of logistics workflows.

How VDF AI fits a logistics company

VDF AI is installed in a logistics company’s own data centre, a private cloud region of its choosing, or an isolated network at a terminal or depot. When it runs with local models only, shipper documents, customs data, prompts, retrieval indexes and logs remain inside that environment, which makes customer confidentiality terms and broker record rules easier to meet. VDF AI holds no customs authorisation or certification of its own; brokers and declarants remain responsible for what they file.

Document extraction, customs preparation and exception handling run as governed agents. Each agent receives only the tools its role needs, such as read access to the TMS or a draft-only customs connector, and stops at a human approval step before a filing, booking change or customer message goes out. Retrieval is permission-aware, so a customer service user sees only the accounts they serve and a broker team’s client files stay with that team. The audit trail records the requester, the model, the documents consulted and every tool call, which is the evidence a shipper’s auditor or a customs authority may ask for.

The VDF AI Router puts open-weight models served through vLLM, Ollama or another OpenAI-compatible endpoint behind one policy layer, with budgets per team and per-call logging, and can add an approved cloud model for work that involves no shipper data. Dispatchers and back-office users log in with Microsoft Entra ID, which is supported natively; a firm whose identity provider is Okta or Keycloak routes sign-in via a reverse proxy that understands SSO, and role-based permissions are included at every plan level. A forwarder or 3PL can test the platform on one lane or one customs team during a three-month trial before buying seats.

To find out whether document data, system integration or governance will hold back a first rollout, take the free AI readiness assessment. The enterprise AI procurement guide lists the questions to put to any AI vendor that will see shipper data. This guide is general information, not customs or legal advice.