From Scanned Paper to System Record:
AI-Powered Dangerous Goods Entry.
Anyone who has processed a dangerous goods shipment knows the routine. A document arrives, sometimes printed, often scanned, occasionally handwritten, and the work begins: opening it on one screen, the TMS on the other, and transcribing UN numbers, packing groups, flashpoints, and shipper details into the system one field at a time. If you have twenty of them waiting, that is a significant chunk of your day gone.
The new AI capability built directly into TMS replaces that process. A user uploads the document, and the system handles the rest.
The Problem With Conventional Parsing
Structured document parsing, the kind that looks for a field in a fixed position on a page, breaks down quickly when applied to dangerous goods forms. These documents come from dozens of different shippers, each with their own template. Some are partially handwritten. Scan quality varies. Columns shift. A rule-based parser has to be told exactly where to look; when the layout changes, it fails.
The solution requires a different approach: a reasoning layer that understands what it is looking at, rather than one that pattern-matches against a fixed schema.
A Thinking Layer Over Your TMS
The capability is built around a large language model paired with a set of tools connected to the TMS backend via an MCP service layer. When a document is uploaded, the system identifies the document type automatically, extracts the relevant fields: UN numbers, packing groups, flashpoints, reference data, and matches the record to the correct booking using container numbers or reference IDs. Extracted values are then validated against the dangerous goods master data in the TMS, with OCR errors resolved before anything is written to the system.
The user does not need to check for transcription errors. The system already has.
“At the moment, it takes a lot of time for users to enter this information manually. They open the PDF on one side, open the system on the other screen, and read it across. Imagine having twenty of these documents to enter.” – Orkun Orbay, Product Lead TMS
Handling the messy reality of scanned documents
Scan quality is one of the subtler challenges in any document automation workflow. A UN number misread by a single character is enough to cause a booking mismatch or a compliance error. The system addresses this by cross-referencing every extracted value against the dangerous goods master data in the TMS. Where the extracted text is close but not exact, the model identifies the most likely correct entry and resolves the discrepancy automatically. The result is that validation happens as part of the extraction process, not as a separate step that falls back to the user.
Processing time drops from minutes to seconds per document, regardless of scan quality or layout complexity. The approach scales without additional user effort twenty documents are handled exactly as easily as one. For organisations with specific requirements around data control, the LLM provider is configurable: Claude, GPT and Gemini models are all supported with customers’ own API Keys, so the data can be controlled within the customer domain rather than TMS domain.
A foundation, not just a feature
The dangerous goods workflow is the first concrete application, but it is built on infrastructure designed for much more. An orchestration layer sits between the interface and the model, managing each step: proposing actions, requesting user confirmation before any changes are made, and only then executing them, in a way that allows new capabilities to be added without rearchitecting the whole system. This keeps users firmly in control: the AI suggests, but the human decides.
The roadmap includes batch processing support, open-ended Q&A across all TMS pages, and expansion to RM, Marketplaces, and Fleet modules in subsequent quarters. A beta launch for the dangerous goods scenario in TMS is planned for mid-April 2026.
“The milestone here is not just one feature. It is that we now have an infrastructure for working with AI tools, an orchestration layer, MCP connectivity, a chat interface, that we can build every future capability on top of.” – Osman Akdemir, Co-CEO & Co-Founder
The dangerous goods scenario is the proof of concept. The platform is the product.
