Artificial intelligence

What it means

Artificial intelligence derives an answer from data rather than from rules written out in advance. That is what separates it from robotic process automation, which repeats one fixed path exactly.

In bookkeeping the real uses are narrow: reading documents with OCR, suggesting a posting, spotting outliers, drafting text. The output is probabilistic, so responsibility for the entry stays with a person.

A language model assembles an answer that sounds plausible, not one that has been verified. It can therefore confidently quote a section that does not exist or blend the old and new wording of an act. Certainty on a tax question comes not from the tool but from the law itself, guidance from the Financial Administration or a binding ruling.

How it works in practice

Duties graded by risk come from the AI Act. The most sensitive deployment is in HR, meaning the screening of applicants and the monitoring of employees while they work.

Once personal data goes into the tool, GDPR applies and the provider needs a processing agreement. Without one it is a disclosure to a third party with no legal basis behind it.

When we process bank movements, the tool proposes a posting from the payment text and the partner’s history, and the accountant confirms or corrects it. Every proposal is flagged in the system and the audit log records who approved it, so it can later be shown who is responsible for the entry.

What to watch out for

Pasting a client payroll file into a public chatbot. It is a disclosure of personal data to a processor with no contract, which is assessed as a personal data breach and, as a rule, notified to the Slovak data protection authority within 72 hours.

Confirming proposals in bulk without checking them. A systematic error, such as a wrong VAT rate for one supplier, repeats across hundreds of documents, flows into the VAT control statement, and fixing it means supplementary filings for several periods.