Got work that keeps repeating itself? We'll build an AI agent to take it over

It depends on how many systems the agent has to work with and how well the brief is prepared. A simple automation of a single task typically costs in the low thousands of euros; an agent with its own set of tools, an evaluation set and integrations with company systems costs more. During the initial consultation we go through your brief and send a written quote with scope of work, timeline and price. Neither the consultation nor the proposal costs anything.

A script is a program with a fixed, hard-coded procedure: it does the same thing every time and never improvises. A workflow is a predefined sequence of steps in which the model handles one part – for example it reads an e-mail and converts it into structured data. An agent is given a goal and a set of tools and chooses the order of steps itself, based on what it discovers along the way. Each of the three architectures has a different price, speed and level of predictability, so the choice is not cosmetic.

An agent pays off where the next step is decided only by what is discovered along the way, and where the list of possible variants would be longer and more fragile than the description of the goal. If the whole procedure fits into a one-page flowchart with no node that requires judgement over open-ended input, a workflow or a script will do the job – cheaper, faster and more predictably. We analyse the task in the very first step and tell you straight away if an agent is unnecessary.

We connect to e-mail inboxes via IMAP, ERP and accounting systems, e-shops, CRMs, internal databases, spreadsheets and cloud storage – anywhere there is an API or access to the data. Where an API is missing, we work with exports, imports or a direct database connection. What the agent can do is defined by the set of tools it gets from us; whatever is not in that set cannot happen.

You receive a proposal and a quote within five working days of the intro call. For smaller tasks, the first working version runs on your real data within a month. A more complex agent connected to several systems takes longer, but even then we work in stages so you can continuously see what already works and what does not yet.

We plan for it from the start, because an agent usually does not crash – it returns a plausible-looking piece of nonsense. That is why every step has a check against hard data: verification in the catalogue, a control calculation, a comparison with the database. Whatever fails the check goes to a person, together with the reason it was stopped. Every tool call is logged including inputs and results, so for every output you can trace back what it was based on.

That is your decision and it is set separately for every operation. By default we recommend a mode in which the agent completes cases that pass all checks on its own and prepares a proposal for approval for the rest. Operations that change data or go out to a customer stay subject to human confirmation until the measured numbers show it is no longer needed.

We work with commercial models such as Claude, GPT or Gemini as well as open models that can run on your own infrastructure. We build agents in Python and PHP on top of the providers’ APIs, with MCP servers for connecting tools, and – where that is enough – on automation platforms such as n8n or Make. The choice of model is a question of price, latency and data-handling requirements, not fashion.

That depends on what you choose. We can run the agent on your server or in your cloud, with the model running in the EU or directly on your premises. If we use a commercial model, we tell you in advance exactly what is sent to it, and sensitive values can be replaced with placeholders before anything is sent. The processing of personal data is covered contractually.

You do. The source code, prompts, evaluation sets and all data remain yours, and we hand them over together with documentation. We do not lock you into a platform of our own, so you can take the solution in-house or hand it to another supplier at any time.

Running costs consist of the model’s token fees, hosting and our oversight. An agent run consumes considerably more tokens than a single chat question, because at every step the model re-reads the whole history of previous steps. That is exactly why every run has a predefined cap on the number of calls and on spend, and it is terminated when the cap is exceeded. We measure costs from day one, and tasks that do not need an agent are moved to a cheaper workflow or script.

Twenty real examples of the task, including the unusual ones; a person who does the task today and can say when the result is correct; and access to the systems the agent is supposed to work with. None of it has to be prepared in advance – we collect it together in the first step.

We can run and monitor it for you, or hand it over to your IT team with documentation and training. During operation we track the error rate, costs and the cases the agent returned to a human, and fine-tune it accordingly. When the inputs change – say a partner reformats their e-mails – the checks reveal it before it shows up in the output.

Yes, and we recommend it. We start with one task, measure how it performs, and only then add more tools or let the agent loose on neighbouring processes. Extending is cheaper at that point, because the system integrations, checks and logging already exist.

In practice it takes routine off people’s plates and leaves them the decisions. An agent handles retyping, look-ups and preparing inputs, but exceptions and approvals stay with a person – for operations that change data or go out to a customer, that is also a technical necessity. Most often the same team handles a larger volume of work, and there is less overtime spent on routine.

Write to us via the form on this page or call +420 555 333 158. We get back to you within one working day, discuss what you want to solve on a fifteen-minute call, and within five working days you receive a written proposal with a price. Neither the consultation nor the proposal costs anything or commits you to anything.