Less manual work
more time for business

I will not name the price of the work upfront, because it depends on what it has to do — I name it after we talk, and then it is one fixed figure with no hourly meter. The hardware is a separate, one-off matter: a mini-PC without a graphics card is usually enough to start, so a few hundred euros, and it stays with you. More documents or heavier AI models need a stronger machine — I say so before the purchase, not after. Monthly support is optional and costs €149 or €349; what it covers is set out below. There is a second road as well: the small computer stays, and the heavy part is computed by an outside provider's model. You pay less up front and a monthly fee instead — and some of the content leaves the company. I tell you which, before you decide.

By default the solution runs on the company's server or in a private EU cloud. If data must not leave the company, I use local AI models that work without sending content to an external service. On a computer in your company, the working data stays in the company. On a machine rented only for you, it sits with that machine's provider, outside the company. In both cases the backup goes to a separate place. External APIs only when the client consciously agrees and the data type allows it. If I process personal data on your behalf, we put the necessary arrangements in writing. I describe where the data physically sits and which services are involved; the legal assessment stays with your advisers.

Marketing material often talks about a “training obligation from August 2026”. That is misleading. Article 4 of the EU AI Act has applied since 2 February 2025. Since 27 July 2026 it requires measures to support the development of AI literacy among the people who work with such systems on the company's behalf; it does not require guaranteeing a particular level of competence for each individual. The regulation neither prescribes a specific course nor says that one training session is always enough. During the rollout I document the technical briefing and the limits of the system. Which further measures your company needs is a question for your legal or data protection advisers.

It depends on whether you issue invoices or mostly receive them. If you run a shop and invoice businesses — your invoices may have to go out in a format the other side can read automatically. Whether that applies to you, and from when, depends on the country, the scale and the type of sales — settle that with your bookkeeping. Technically a PDF does not do it, however neat it looks. This usually means real work inside the shop itself, and that is my ground: I have worked with shops for years. If you mostly receive invoices and a bookkeeping office handles the accounts — your part is small. It comes down to not deleting the data file — that file is the actual invoice, not the printout. Sometimes it arrives as a separate XML file, sometimes it sits inside the PDF, so a neat-looking PDF tells you nothing about whether the invoice can be read by a machine. The office takes care of the rest, and usually does it well. Deadlines and thresholds depend on turnover and the type of business, and they keep changing over the next few years. What applies to your company is for your tax adviser to determine — I advise on neither tax nor law.

Whether the whole thing works depends on it, so I treat it as part of the job rather than an extra. The team gets a screen that looks like a normal app — a box for the question, a form or a chat. Not a panel for specialists. I walk people through it unhurried and leave a short guide in German or Polish, included in the price. The technical side I take on myself, within the hours of the package you choose; if more comes up, I give you the price before I do anything.

In the first week we talk, and you get 1–3 ideas in writing — or an honest “there is nothing here to do”. In the second I show the first working version live, provided I get the logins and the data is in usable shape. I do not promise a date before I have seen that; if something drags on, I say so straight away. Before you approve the paid part you can walk away with no obligations.

I give a date once I have seen the data, the access and the programs that are meant to connect — before that it would be guesswork. What I can say upfront: with an AI assistant most of the work goes into tidying the documents, not the AI itself. I always start with a small scope that shows quickly in practice, and only then expand.

First we organise the documents: instructions, PDFs, system exports, emails or an internal wiki. Then I build an index so the system can find matching passages even when the question uses different words than the document. Only then does the AI model write an answer and show the documents it searched. Technically this is called RAG, but for the user it should feel simple: ask a question, get an answer, check the source.

Search shows matching documents — like an internal Google that understands the meaning of the question, not just keywords. An AI assistant goes one step further: it reads the found fragments and writes a short answer. Starting with search is often useful because it quickly shows whether the documents are organised well enough.

I start with today's process: where the data comes from, who copies it, who waits for the result. Then I build a flow that connects existing systems and performs repeatable steps automatically. Error handling matters: if something fails, the system should notify a person instead of silently stopping. At the end you get documentation and handover.

You do not have to. You should get something that works, a description of how it works, and a clear statement of who looks after what — the tool names you can safely skip. I pick them to fit the problem, not the other way round, and I take ones with no licence to pay for. If somebody ever takes this over from me, the documentation holds the full list: open-source, among others Docker, Postgres, n8n, Prometheus and Grafana.