AI that does the repetitive work,
while the decision stays in your hands.

Three things decide the price: how much data we start with and how well ordered it is, how many systems it has to connect to, and whether the model runs in the cloud or on your own server. A narrow pilot can usually be covered in a few weeks, and that already shows the scale of a full rollout. Before quoting we check whether there is a realistic return, because we do not take on work that will not pay for itself.

Yes, every language model can be wrong, which is why we never let unchecked output go straight to a customer where a mistake would do harm. We narrow the answers to the company own documents and require a source alongside, so it stays verifiable where a statement came from. If the model finds no backing, it does not invent an answer but hands the case to a person.

At the start of the project we put in writing which data leaves the system and which does not. For a cloud model we choose a provider that contractually rules out use for training and offers EU-based data handling. With sensitive data the personal fields are pseudonymised before anything is sent, and where everything must stay in house we use a locally running model.

For most tasks no, because cloud models are cheaper and more capable. Where the data must not leave the company, though, we run a local model on our own Ubuntu VPS server. That means slower and more expensive operation, but in return no data leaves the building.

A well defined pilot is typically working within 2–4 weeks; the full rollout and wiring in takes more like 2–3 months. The longest part is almost always gathering and tidying the data, not the development itself. That is why we start with one narrow process, so there is something to hold on to early.

What it takes over is the repetitive preparation: typing, searching, summarising, copying data. The decision and the approval stay a human task, because that is where the responsibility sits. In practice colleagues typically spend more time on the work that genuinely needs human judgement.

To a webshop, an ERP or invoicing system, e-mail, a database and most services that have an API. Besides our own framework we are at home in WordPress, WooCommerce and Magento 2 environments too. Where there is no API we usually find another route, such as an export and import based sync.

MCP is an open standard for letting an AI assistant reach a company systems safely: to query, summarise and prepare drafts, with precisely defined permissions. Without it the data has to be copied and pasted every time, which is slow and error-prone. A connection like this runs live on our own framework, so we will not be experimenting on your system.

Yes, today models write and understand Hungarian convincingly as well. Experience shows, though, that without a glossary and style rules the text stays recognisably machine-made. So for every project we put together the company own vocabulary and its list of words to avoid, so the output speaks in the voice of the brand.

We build the solution so that the model can be swapped, so it is not tied to a single provider. When a new model arrives we re-run the measurements and only switch if it genuinely brings better results or cheaper operation. This is part of the operations package.