Custom AI agents for repetitive data work.

02 / 05

Which business processes to automate first

The first thing worth automating is work that someone does the same way every time and could describe in a few rules. A typical example is an accountant who matches rows by hand in the same Excel file every month. Decisions that call for judgment or a personal relationship with a customer are a poor fit for a first agent.

Not every such process needs AI. Where data arrives in a fixed format, such as exports from your bank or online store, fixed rules that can be tested are enough. I only add AI where something has to read emails, attachments or other text without a fixed format.

When I assess which processes to automate, I check each one against four criteria. The more of them a process meets, the better a candidate it is:

• It repeats daily, weekly or monthly
• It follows rules you could write down on paper
• The data is already digital: in an email, a spreadsheet, a PDF or a system
• A mistake would cost money, and the result can be checked against the source

03 / 05

How I keep AI agents from making costly mistakes

I keep agents from making costly mistakes with fixed rules around the AI model: when an agent isn't sure or an input changes format, it stops and waits for a person.

In invoicing, this means that if a partner changes its report format, invoicing stops before a wrong invoice is created. The agent invoices only the items that clearly match, never calculates or estimates missing figures, and takes every value that drives the amount from a single source. A person then only has to check the exceptions and discrepancies the agent finds.

I built my own Python tool on the same principle: it works with Slovak public company registers and adds financial data and website addresses to about 2,000 company records. Its language model returns the data as structured output, in a fixed shape rather than free text, and a hard limit caps API spending. Every agent I build has these safeguards:

• Limits it must not cross, including a cap on AI spending
• Answers from the model in a fixed shape, not free text
• A log of every step, so you can trace what it did and when
• Human approval for steps where a mistake would cost money
• Fixed rules wherever the AI doesn't need to estimate anything

04 / 05

Which systems an AI agent can connect to

An AI agent can connect to most tools businesses commonly use: email, CRM, accounting software, an online store or shared spreadsheets. Where a system allows it, I connect through its API; otherwise I use file imports and exports or connect directly to the database of your own system. With accounting software, API access is often a paid add-on; Money S3, for example, offers it as an extension module. At the consultation we'll find out whether your system can be connected and whether you need to buy anything. We'll also agree in advance which data the agent may access.

I build agents in TypeScript and Node.js with the Vercel AI SDK, and I use Python for data-heavy jobs.

For companies in Slovakia, e-invoicing changes what an agent needs to read. From 1 January 2027, VAT payers established in Slovakia must, with some exceptions, issue invoices to Slovak businesses as structured XML files, which software can read directly with no AI extraction needed. AI still helps with documents that arrive as PDFs or scans, for example from foreign suppliers or from businesses that aren't VAT payers.

05 / 05

How much does an AI agent cost?

An agent for a single process starts at around €500. The price depends mainly on how many different formats the agent has to read, how many systems it writes to and how people approve exceptions: whether email is enough or you need a dedicated app. If an agent grows into a system your company runs on every day, it becomes custom software development: a project of several months that I price in parts.

Running costs for hosting and AI model calls range from tens to low hundreds of euros a month, depending on volume. By default I run the finished agent on my own accounts for a monthly fee, and a move to your own accounts can be arranged. I always fix its bugs for free; for the other terms, see how we work together.

After the first consultation, you get a quote split into parts, each with a specific price and timeline. If you approve it or some of its parts, I build a free prototype on a sample of your real documents or exports, so you can check that it will work before you make a final decision. For that, I need someone who knows the process and can tell me when the result is right. Automating a single process is then a few days of work. The full deployment with documentation and training usually takes 6 to 12 weeks, depending mainly on how many of your systems need to be connected. To start, use the form at the bottom of this page to tell me which process slows you down today.