AI for forwarders and logistics companies: less typing between mailbox and TMS, with people deciding

The change is less typing and fewer screens. For a single shipment, a forwarding clerk switches between the mailbox, the TMS, carrier portals, customs software and spreadsheets. A 2026 logistik-heute report on the software vendor Nemox put it this way: a clerk today works in five programs at the same time. AI takes over the reading and pre-filling, so the clerk checks and decides instead of copying.

Data entry is where the industry starts. In Descartes' 2026 benchmark of 300 European shippers and logistics providers, 75 of them from Germany, 41% use AI for automated data entry and processing unstructured data, 37% for route and load optimisation and 29% for freight forecasting. It is a vendor study, but it shows where the work is. Adoption across the whole sector is still low: in the IAB establishment panel 2025, transport and warehousing is among the sectors with the lowest use of generative AI, under 15%.

Agents, programs that carry out steps in your systems and present the results for approval, are still early. In Inform's 2025 survey of 114 specialists and managers, about 90% of them at companies headquartered in Germany, 12% use AI agents productively and 25% are testing them. 46% prefer a model in which AI makes suggestions and a person makes the final decision, 81% expect explanations and 61% want to be able to intervene. That is how we build.

The same goes for shipping departments at manufacturers, wholesalers and retailers. They deal with the same emails, documents, carriers and status questions.

In the flows where an email or a document turns into a record in the TMS. Language models do that kind of work reliably. Alice Kirchheim of Fraunhofer IML sees the greatest potential in company-specific generative AI built into processes, for example for pre-filling forms from existing company data (CHEManager, 2025). Each use case has its own page, linked below.

• Transport documents into the TMS. Transport orders, delivery notes, CMR consignment notes, air waybills and booking confirmations arriving as PDF or email are read, checked against master data and prepared as an order or booking in the TMS. The clerk sees the document next to the filled fields and confirms.
• Customer emails and status requests. Questions such as "where is my shipment?" are answered from TMS status and tracking data, as a draft the clerk sends or, for clearly defined cases you release, as an automatic reply. Complaints and anything unusual go to a person.
• Quotes. An assistant reads the quote request, looks up rates, surcharges and similar past quotes, and drafts the offer for your sales staff. At the Frankfurt air freight forwarder Cargo Movers, an assistant drafts offers in Outlook and staff check them. According to the vendor Nemox, as reported by logistik-heute in 2026, 54% of the offers went to the customer unchanged and 17% with one correction. Below a weight threshold the system may send on its own.
• Dispatcher assistant. Built into the TMS, it gathers what a dispatcher needs for a decision: open orders, capacity, carrier answers, delays and customer constraints, and proposes the next step. The dispatcher decides, and the assistant prepares and documents.
• Freight invoice checks. Carrier invoices matched line by line against orders and agreed rates in the TMS, with deviations passed to a person. For Cargo Movers, the vendor reports more than 28000 invoice lines reconciled, 97% of them without correction of the extracted values (logistik-heute, 2026). These figures come from the vendor. Cloudsail did not build that system.

Specialist products already cover parts of this. Mailbox assistants for forwarders exist as products, as the Cargo Movers case shows, and TMS vendors add AI features of their own. Route and load optimisation is a field for specialist planning software, and we do not build it.

A custom build makes sense when your flows run across several systems, such as a TMS, a customs tool, carrier portals and an ERP, when your customers send documents a product does not read well, or when shipment and customer data must stay on your own servers. If you want to compare, we test a product and a custom build on the same past shipments, so you decide on numbers.

Wrong prices and promises leave the company in writing. A quote with a wrong rate or a reply with a wrong delivery date binds you, so the approval step and its thresholds are the core design question: which quotes may go out automatically, if any, and which always need a person. You set them, and they are widened only when the measured results support it.

Customs and dangerous goods documents carry legal liability. The system can pre-fill them and check them for completeness, but a qualified person signs and submits them.

Data quality and integration are the main hurdles. In the Inform survey, 67% name data quality and availability as the biggest challenge and 58% integration with existing IT. We start every project with the master data the use case needs, such as customer, address, rate and carrier data, and where it is incomplete, you see that before we build anything.

Drivers and staff. Telematics, driver data and dispatcher activity are personal data, and we leave them out unless the use case needs them. Using AI to evaluate the performance of drivers or warehouse staff is a high-risk use under Annex III of the AI Act. From 2 December 2027 it brings duties such as human oversight and informing staff for the company using it. We do not build it. A dispatcher assistant that logs who did what can in principle monitor performance, which makes it subject to co-determination under § 87 (1) no. 6 BetrVG. We therefore design the logging with pseudonymised IDs and agreed retention, so the works council can agree to it.

Customers who chat with a status assistant or receive automatic replies must be told they are dealing with AI. That duty under the AI Act applies from 2 August 2026. Data protection and legal questions are yours to decide, advised by your data protection officer and lawyers. We provide the technical facts: what is stored, where it is processed and who can see it.

The agent runs where your shipment and customer data may go: in your own cloud account in an EU region, or on your own servers with an open-weight model, a model whose weights you run yourself. An LLM gateway, one controlled access point to all models with keys, budgets and logs, sits in front of the models, and every action is logged in your systems.

• Mailboxes: Outlook or Exchange inboxes for orders, quote requests and customer questions, read by a service account limited to those inboxes.
• TMS: orders, shipments, rates and status read through its interface, EDI messages or database views. Bookings and drafts are written only through tools that each do one business action, with approval.
• Documents: PDFs, scans and photos of delivery notes and transport documents, read and checked against master data, with the original kept next to the extracted fields.
• Carrier and customer portals, tracking data and, for invoice checks, the ERP and accounting system.
• Approval in the tools your staff already use, such as a queue in the TMS or a draft in Outlook, not one more screen.

Before go-live, the agent works through a set of your real past shipments: transport orders with the booking your team created at the time, quote requests with the offer that was sent, customer emails with the reply that was right. We measure field by field and case by case, and you set the threshold the acceptance test must meet.

Then the agent runs in shadow mode: it prepares bookings and replies alongside your team, nothing is sent, and we compare. Only when the numbers hold do drafts go to staff for approval, and only cases you release go out automatically. We measure hours per booking or per quote before and after, on your own shipments.

Explanations are part of the test. Each suggestion shows where it comes from, whether an email line, a document field or a TMS record, so a dispatcher can see why. After go-live the same tests run after every change to models, prompts or interfaces.

Start with the flow that floods your inbox: quote requests, transport orders by email or status questions. Bring a description and a rough count of cases to a free 30-minute first call with an engineer. The call needs no access to your systems or data.

Then we agree the first step and its acceptance threshold. Each step has a fixed price, agreed in writing after the first call. Before anything is built, we check your master data and the TMS interface, so you know whether the data carries the use case.