The demo is built quickly. It gets hard afterwards: when real documents, real users and real approval processes come in. From our projects of recent years - from comparing specifications to a document pipeline handling 300,000 documents a month - we draw six lessons.
1. The value lies in documents you already have
Almost every successful project started not with a model but with a pile of documents nobody could read in full anymore: 100-page price lists, cross-cutting specifications, thousands of user comments. Retrieval-augmented generation and semantic search make this knowledge usable without anyone rewriting it.
Our rule: First ask which document costs the most time today. Then build.
2. Data preparation is half the work
OCR for tables inside images, deduplication, version comparison - the unglamorous part decides answer quality. In the document pipeline, eight models work together; only one of them is generative.
3. Independence belongs in the architecture
Models change every quarter. Separating prompts, data access and model cleanly lets you switch - including to an own model when confidentiality or cost require it. In a recommendation service, we evaluated exactly this option from day one.
4. Humans stay in the loop - visibly
A chatbot for sales advisors does not replace advisors, it makes them faster. An AI summary of changes does not replace assessment, it directs it. The projects that get adopted make the human role explicit: who reviews, who approves, who is accountable.
6. Gates, not hope
Every AI initiative of ours passes through the same stages - and at every gate a deliberate decision is made: continue, pivot or stop.
5. Governance early, not at the end
From regulated projects - software update approval, data protection in autonomous driving - we know that documentation, roles and approvals can be built as a workflow. With the EU AI Act, this becomes mandatory for many AI systems. Thinking about it from the start saves expensive rework.
Note: this article provides orientation and is not legal advice.
