Service · AI Automation
The machine works. The human approves.
AI pays off not as a gimmick but where the same work recurs every day: summarising findings, sorting inquiries, matching participants, drafting texts. We build exactly these flows - with a human in the loop and, where needed, on your own server.
The problem
"We are doing something with AI now" often ends in a chatbot nobody uses. The real work sits in the processes behind it: where does the data come from, who approves, what happens on uncertainty - and may the data leave the building at all?
What we build
Process automation
Classifying inquiries, summarising findings, drafting letters, enriching data - one flow per task, with human approval.
Matching and allocation
Algorithms that allocate participants, inquiries or quotas by rules - like the community matching in mP Event. Traceable, not magic.
AI that never leaves the building
Local models on your own infrastructure when patient data, contracts or findings must not go to a provider. Privacy as architecture.
The analyzer as an example
Our own tool: deterministic rules, one measurement, one AI step with clear guardrails - and honestly labelled who works when.
How it runs
- 01
Find the task
What costs time every day and follows a pattern? That is the candidate - not what impresses.
- 02
Prototype with approval
Within days a flow that really runs, with a human in the loop. Measured in hours saved, not demos.
- 03
Harden and operate
Edge cases, privacy, hosting. We run it with you until it runs without us.
What we measure
The number that justifies an automation. Otherwise none.
How often the human accepts the proposal - if it drops, the machine is tuned wrong.
Which data leaves the building. For practices and law firms: none.
Common questions
Which AI do you use?
Anthropic's Claude models where data may go outside; local models on your own infrastructure where not. No OpenAI. The model follows the task and the privacy requirements, not the hype.Does this replace staff?
It replaces work nobody likes doing: sorting, summarising, retyping. Approval stays with the human - that is architecture, not politeness.What about privacy?
For sensitive data the model runs on your server, not a byte goes outside. For non-critical data we work with data processing agreements and EU hosting. We tell you beforehand what is possible and what is not.How long does a first flow take?
A prototype with approval in days, a hardened operation in weeks. We start with one task, not with a strategy.
Next step
Where do you stand?
The analyzer is such a flow itself: rules, a measurement, one AI step with guardrails. Try it on your site - and we talk about what eats your time every day.