AI automation

AI pays off first where someone today reads text and derives the same kind of decision from it every time: evaluating invoices and delivery notes, sorting and answering enquiries, enriching and reconciling data sets, pulling reports together from scattered sources.

It pays off less for tasks that come up rarely or where a mistake is expensive and nobody checks the result. We say that in the initial call too.

We run language models on our own hardware in Germany. For you that means your documents, customer data and internal texts do not go to OpenAI, Google or Anthropic, and they are not used to train anyone else's models.

That is not a footnote in the privacy notice but the reason these projects can be approved at all — with personnel data, health data or trade secrets, the route via a public API often fails at the approval stage.

One process first, not ten. We look for the single workflow with the highest repetition and the clearest outcome, build it and measure what it saves. Only then comes the next one.

The reason is experience: automations that touch ten processes at once are, three months in, neither finished nor switchable-off. A single workflow stands after a few days and produces a number you can go on calculating with.

Only what someone can describe can be automated. And what people describe regularly differs from what they actually do — not out of carelessness, but because nobody consciously notices well-worn routines any more.

That is why an automation project with us starts where the process runs, whenever possible: half a day sitting alongside, taking notes, asking questions. For that we come to you. From Bogel near Nastätten in the Rhein-Lahn district, Koblenz is about 25 km away; Limburg an der Lahn and Wiesbaden are part of our regular catchment area.

Everything after that — development, testing, operations — runs remotely, and we look after clients elsewhere in Germany in exactly the same way. Where an on-site visit is not possible, screen recordings of the actual steps take its place.

Projects start at 3,600 EUR and go up to 150,000 EUR for a platform with several processes and its own model infrastructure.

How long it takes depends on what “finished” means. A single process is up and running after roughly 3 days — far enough to produce a figure you can work with. Until that same process is tested, connected to your systems and in production, it usually takes 2 to 4 weeks. A platform with several processes takes several months.

We bill by time and materials at 150 EUR net per hour. If the solution runs on our infrastructure, operations come on top — from 10 EUR a month, depending on how much computing the model needs.

If the process comes up rarely, automation does not pay off — at twenty cases a year, doing it by hand is cheaper than any software.

If nobody can check the result and a mistake is expensive, a language model is the wrong tool. And if the process cannot be described because everyone does it differently, it should first be put in order and then automated. That costs a conversation, not a project.