In short
Process automation and assistants built on your documentation — that is what we deploy most.
One payroll automation cut monthly HR work from over 2 days to about 3 hours.
An assistant answering from company documents: over 80% of answers usable.
Personal data is minimised and pseudonymised before anything reaches a model.
We work on model provider APIs, and on models hosted in your own infrastructure when the data requires it.
What we deploy
Process automation
Repetitive work on documents and data that somebody currently does by hand every month. This is usually where the largest saving sits.
Assistants on your documents
An assistant that answers from your documentation, procedures and knowledge base — with a reference to the source rather than an invented answer.
When AI pays off, and when it is just a cost
AI is not the answer to every problem and we will not pretend otherwise. A deployment makes sense when the process is repetitive, based on text or documents, and costs real people real time.
It does not make sense when:
- the process happens a few times a year — the build will never pay for itself,
- the input data is in a state that has to be cleaned up first,
- full certainty is expected on decisions where a mistake is expensive and nobody checks the output.
In those cases we say so on the first call, instead of selling a pilot that ends up in a drawer three months later.
Results from deployments
Numbers from projects actually running in production.
Example: payroll automation
The situation. Preparing payroll each month took the HR team over two days — retyping data between systems, checking it by hand and reconciling discrepancies.
What we did. We automated the processing and validation of the input data, leaving the human decision only where it is genuinely needed rather than at every step.
The result. The same process now closes in around 3 hours a month. The HR team did not disappear — they stopped retyping data.
Example: an assistant on documentation (RAG)
For the tax sector we built an assistant that answers questions from company documentation, in a RAG architecture — the model answers only from retrieved fragments of your documents, not from its own memory.
That is the essential difference from a general-purpose chatbot: the answer can be checked against the source, and the model does not fill in what the documents do not contain.
Over 80% of answers are usable. The rest go to a person — and that is by design, not a shortfall. In a field where a wrong answer is expensive, the assistant is there to take load off people, not to impersonate an expert.
How a deployment runs
We start with one process and with the question of how we will know it worked.
Choosing the process
We look for one that is repetitive, costs real time, and can be measured before and after.
Data review
We check what state the input data is in and what has to be cleaned up before anything gets built.
A narrow pilot
We build a working version on part of the process and measure accuracy against your real cases.
A decision based on numbers
If the pilot does not deliver, we say so. We widen the scope only when the result justifies it.
Deployment and upkeep
Production, monitoring of answer quality, and corrections when the data or the process changes.
Data, GDPR and what reaches the model
Before any data reaches a model it goes through preprocessing in which we minimise the scope of data and pseudonymise personal data. Only what is necessary to perform the task is sent to the model provider.
If the nature of the data means it cannot leave your environment in any form, we run the model inside your own infrastructure. The data then does not leave your environment at all.
Which option is right is settled during the data review — a decision taken before the build, not after it.
As an EU company we are subject to GDPR directly, and we sign a data processing agreement covering what we process on your behalf.
Provider API or a self-hosted model
A provider API — this is how we work most often. Faster, cheaper to maintain, and it gives access to the strongest available models without investing in hardware.
A model hosted with you — when data cannot leave your infrastructure, or a regulator requires it. It costs more and needs maintenance, but it is sometimes the only acceptable option.
We have no ideology here. We pick after reviewing the data, not before.
Scope of work
Process and data review
Before anything is built — what repeats, what it costs, and whether the data allows it at all.
Process automation
Document and data processing with a human decision where it is genuinely needed.
Knowledge-base assistant
RAG over your documentation, with a source reference on every answer.
Integration with your systems
Deployed where people already work, rather than as yet another separate tool to open.
Measuring the outcome
We measure answer quality and process time before and after. Without that there is no way to tell whether it was worth it.
Let us check whether it pays off for you
Tell us about the process that eats the most time. We reply within 24 hours — including when our view is that AI is not the right tool here.
