AI in practice - Compliance processes automated with LLMs et al.
For many companies, compliance is a constant source of noise: questionnaires, audits, documentation requirements, recurring requests – often with enormous overlaps, but each time in slightly different language. It becomes particularly frustrating when the same content has to be manually reassembled again and again: from PDFs, reports, certificates, codes of conduct, and process documents – often under time pressure and with high liability relevance.
This podcast episode is not about the usual complaints about regulations, but about a specific practical case: how we worked with a software provider to further develop a compliance platform in such a way that AI radically accelerates the process – without losing reliability, traceability, and responsibility. The core: human-in-the-loop, but with maximum relief.
Key Points
- “The absurd thing is that every company ultimately answers the same question hundreds of times.”
- “In the end, a senior employee acts as a human copy-paste machine.”
- “In compliance processes, there is no room for ‘It was only roughly correct.’”
- “Source references are the difference between interesting and auditable.”
- “You should definitely start with the workflow and not with the technology.”
- “With each use, the system becomes more valuable – a self-learning knowledge repository.”
- “Large language models remain tools – we are miles away from full autonomy.”
- “We don't relieve people in order to replace them – but to eliminate monotonous work.”
Summary
1) The real problem: repetition, inconsistency, and tied-up expertise
Compliance questionnaires are largely redundant in terms of content, but expensive to operate: companies answer similar questions over and over again – often manually, often with the risk of inconsistencies. The hidden costs arise not only from the time involved, but above all from the fact that expensive experts are tied up in monotonous tasks.
2) The client's starting point: Good workflows – but no scaling yet
The client already had a solid platform: data room, document management, questionnaire import, structured workflows.
This eliminated the chaos of email and Excel – but the core work remained human: searching through documents, formulating answers, linking sources.
3) The lever: Use AI where it is reliable – and leave responsibility to humans
Instead of “just throwing a chatbot at it,” the process was first analyzed: What steps do humans actually take? What artifacts are created?
Where are the bottlenecks? The result: automation yes – but auditable and with humans in the loop, because responsibility and liability cannot be delegated to a model.
4) Technical implementation: Stability comes from clean input and retrieval
The breakthrough came through consistent preprocessing during document upload:
- Intelligent OCR for real document quality (scans, distortions, “not all PDFs are created equal”)
- Semantic embeddings at the paragraph level
- RAG (Retrieval Augmented Generation): Relevant passages are found for each question, rather than entire documents being “dumped into the prompt.”
This reduces response times and increases quality because only relevant context is processed.
