The Czech Financial Administration is testing artificial intelligence in the Dialóg project, which aims to simplify work with interrogation recordings, but strictly in a secured on‑premises environment. The reason is the protection of sensitive tax data and internal procedures, which is why the cloud is not used for this data. The experience aims to inspire Slovak institutions as well.
Security above all: why not the cloud
The Financial Administration in the Czech Republic has drawn a clear line: data from tax proceedings and other sensitive information do not belong in cloud-based artificial intelligence. The decision stems from requirements to protect tax data, tactical procedures, and cybersecurity. AI is therefore operated exclusively on‑premises for these purposes, even though it is more costly than the cloud. Cloud tools may be considered only for less sensitive digital services and taxpayer support.
The context is extensive: in 2023–2024 alone, approximately 15,000 testimonies and oral hearings took place. Interrogations last several hours, and processing them requires additional hours of work on the records. They are emotionally demanding, with a risk of verbal attacks and potential corrupt behavior. This is precisely where AI can help as a tool that speeds up search and surfaces relevant passages.
Project Dialóg: transcription, search, and assisted conversation
Dialóg allows uploading an interrogation recording to the internal system, where the AI identifies speakers and generates a transcript. The inspector then searches the text for key names, events, or time segments and can quickly play back selected passages. A chatbot operates on top of the transcript, answering questions like “when was a particular company mentioned” or “how did the parties respond to each other.” The human remains the decisive element: AI is an assistant, not a judge.
The solution also includes working with documents — a staff member can upload an internal document and ask about its content without needing to read it in full. This is especially important for methodologies and procedures that cannot be shared with public AI services. The project runs in a secure environment and is carried out in cooperation with IBM and a state-owned enterprise. It involves validating technologies and procedures as part of a pilot operation.
Practical benefits and next steps
Thousands of employees are working in the field today, and up to 3,300 of them take part in interrogations, while often only a short portion of the recording is relevant. Dialóg saves time in finding key moments, reduces workload, and helps standardize differing regional approaches. It also increases transparency — the recording and transcript enable multiple verifications and subsequent retrieval of context. Thanks to this, inspectors can focus on the burden of proof and the quality of records.
An on‑premises solution is indeed more expensive than the cloud, but it brings control over data and meets strict security requirements. Alongside this, other AI possibilities for less sensitive use are being explored: internal assistants, administrative automation, or predictive models, which are being developed in interesting ways, for example, in Austria. The key remains gradual testing, a realistic assessment of capacity, and an emphasis on ensuring that AI supports rather than replaces the expert. A strategy set up this way can be a useful guide for Slovak institutions as well.
