--- title: AI that grounds its answers description: Knowledge assistants, document processing and computer vision running in real operations. Every answer traceable to its source, all on Dutch infrastructure. url: "https://rual.nl/en/ai" locale: en alternate: "https://rual.nl/ai" updated: 2026-08-03 source: "https://rual.nl" --- # AI that grounds its answers instead of inventing them. We build agentic AI that runs in real operations: knowledge assistants, document processing and computer vision. With a human in the loop wherever judgement is needed, and everything on Dutch infrastructure. [Book a call](https://rual.nl/en/demo) [See customer cases](https://rual.nl/en/customer-cases) [More on security](https://rual.nl/en/security) ## What we build with AI Not pilots that end up in a drawer, but systems that take work off your hands. This is where we can help. - **Knowledge assistants** — Ask questions of your own files, mail and regulations in plain language. Every answer cites its sources, and stays inside what the person asking is allowed to see. - **Document processing** — Invoices, packing slips, orders and contracts: read automatically, matched against your data and passed on. Anything doubtful goes to a person. - **Computer vision** — Recognising, counting and locating things in photos and camera feeds: products, objects, defects. From quality control to stocktaking. - **Agentic automation** — AI that picks its own steps and uses tools inside your flows, on a tight budget and within clear limits. Autonomous where it can be, human in the loop where it must be. - **Forecasting and planning** — Purchasing, stock and planning grounded in your own history: what you are going to need, and when. - **[AI on the floor](https://rual.nl/en/robotics)** — Vision and AI wired into robots and hardware: from pick and place to quality control on the line. ## CIMS Atlas: knowledge that answers back The first of three systems running in production today. A coordinator in the industrial waste chain handles hundreds of matching questions a day: which licensed processor can take this stream, and what does the regulation require. That knowledge sat scattered across mail, legal publications and case files. Atlas turns it into checkable, cited answers in plain language. [See the CIMS case →](https://rual.nl/en/customer-cases/cims-netherlands) - **Reasons until it knows enough** — Where a standard retrieval system searches once and then answers, Atlas fetches more context while it is composing: additional claims, domain terminology. Within a tight cost budget per question, the model decides for itself when it knows enough. - **Three search channels, one ranking** — Semantic, keyword and graph-based search run in parallel, each with its own weighting and failure isolation. A cross-encoder ranks the results, and dedicated channels for case files and regulation join in when the question calls for them. - **New knowledge teaches itself** — New operational records are distilled automatically into structured claims, with plausibility checks and prompt-injection detection. Borderline cases go to a review queue rather than straight into the knowledge base. - **The data stays the customer’s** — All processing on Dutch infrastructure, Zero Data Retention at the LLM layer, and authorisation enforced before a claim ever reaches the model. Every employee sees only their own organisational scope, on a single audit trail. ## CIMS Flow: invoices that process themselves For CIMS we built an intelligent invoice processing system that handles hundreds of invoices a month for ship and cargo related waste, from the ports of Rotterdam, Amsterdam, Moerdijk, Vlissingen and Terneuzen. - Documents are read automatically and matched to the right order - Differences in amount or content go to a review queue, not into the ledger - The team does not grow with the invoice volume, the system does **Invoice processing** — today · 34 documents - `09:41:07` invoice-8841.pdf · data extracted - `09:41:09` matched to order 2026-1187 · port of Rotterdam - `09:41:10` booked · passed on to finance (completed) - `09:43:22` invoice-8842.pdf · data extracted - `09:43:25` amount differs from order · sent to the review queue (held for review) What is unambiguous gets booked. Anything doubtful becomes a proposal for a person. ## CityCube: thousands of objects found without leaving the office Centercom manages thousands of advertising objects across the Netherlands. Their exact position and facing direction used to be fieldwork. We built an autonomous service that does it from a desk. - A vision model looks for the object in street-level panoramas from several angles - The real coordinate is triangulated from two independent sightings, and the facing direction derived geometrically - Confidence gate: an unambiguous find is written straight through, while a doubtful one becomes a proposal a person approves in one click - Every run records what was seen and why it was decided that way [See the Centercom case](https://rual.nl/en/customer-cases/centercom) ![Vision detection marking objects with confidence scores](https://rual.nl/site-assets/_opt/vision-detection-960.webp) ### Anthropic is built into Core Claude is available as a native block type in every environment, including a ready-made memory system and document processing. Adding AI to your flows is a building block, not a separate project. [Read the docs →](https://docs.rual.nl/block-types/anthropic?group=anthropic) ## Where is the knowledge in your company stuck? In heads, mailboxes or folders: we are happy to look at where AI genuinely pays off. And where it does not. [Book a call](https://rual.nl/en/demo) [See customer cases](https://rual.nl/en/customer-cases)