How Ukrainian artificial intelligence speeds up licensing in e-Permit: the experience of the Ministry of Economy

A new “smart” module has been launched in the e-Permit system, which analyzes documents using a large language model (LLM). This is not just analytics, but a complete preliminary check of all materials, which was previously done entirely manually: the validity of diplomas, the consistency of data, changes in surnames, the relevance of certificates, compliance with requirements for education, terms, names, and dates. AI does not make decisions, but it prepares the groundwork for them. It gives people what takes the most time — preliminary verification with specific comments. AI does this without bias, in a closed environment — on Ukrainian servers, without transferring data abroad.

The pilot launch concerns veterinary licenses. But this is only the beginning. In fact, we are talking about an architecture that can be scaled. About a stack that can combine document scans and structured analytics. And also about a team that was able to implement this under strict legislative restrictions, a lack of data, and high responsibility. Let's take a closer look at what exactly they did and how it works internally.

Why introduce AI into the licensing system?

Licensing is not an industry where anyone expects technological breakthroughs. No one renews their permits with an iPhone in their hand or dreams of UX. Everything here is conservative. But that is precisely why it has become the ideal field for the implementation of artificial intelligence.

The problem is simple. To obtain a license, an entrepreneur usually prepares an application, attaches documents, and sends them through Diya. Next, the application enters the system, where an official manually opens it and begins to read through the documents. One by one: diploma, certificate, reference, signature, date. They compare, check, and question. If something is wrong, the application is rejected, which is not always clear to the applicant. While they are figuring out the reasons for the rejection, the queue has already moved on. And so it goes, round and round.

A veterinary license is just such a case. It is complex, multi-component, with clear requirements for professional qualifications. There may be several employees in the application, and each of them has a package of documents that must be read, verified, and approved. Before the introduction of AI, this was done manually. Even a perfectly completed application took a certain amount of time — for one person, it took minutes in ideal conditions, without interruptions or distractions. But what if there are many people in the application, or the data is inconsistent? What if the certificate has expired? What if the surname has changed, or the diploma is old? All this significantly slowed down the data processing process.

Now everything has changed. The application is submitted through the interface in “Dii,” and then the documents are automatically sent to the “eDovizil” system, where the AI module is activated. All materials undergo a complete check for logic, integrity, and compliance with requirements. Any errors or inconsistencies found are highlighted. A recommendation for a decision is generated. The official no longer opens a “blank table,” but rather the result of the analysis, receiving one of two recommendations: “recommend issuing a license” or “reject.”  In case of rejection, the AI report indicates the reasons and inconsistencies.

“In the public sector, there are tasks that no one wants to do and no one has to do manually, but they have a critical impact on the result. The AI module takes on this layer of work and allows experts to make decisions faster and with more confidence”, said Oleksandr Tsybort, Deputy Minister of Economy, Environment, and Agriculture of Ukraine for Digital Development, Digital Transformation, and Digitalization.

Important: the model does not make decisions on its own. It provides guidelines. It highlights what is worth paying attention to. The final decision is always made by a person.

How the system works: from document processing to recommendations

The AI module in eDozvil is not just another LLM integration; it is an infrastructure designed from scratch with clear logic, a controlled pipeline, and a complete rejection of foreign cloud services. When a user submits an application, the system receives two data streams: the application structure (it comes in digital format) and attached documents (most often PDF or images). This is where the first module based on Qwen 2.5 7B VL comes in. It extracts text from files: scans, recognizes, and breaks it down into logical blocks. It works like OCR, but is much more flexible. In particular, it records discrepancies: if the surname on the diploma does not match the data in the application, this discrepancy is automatically transferred to a separate line.

Qwen 2.5 7B VL is a multimodal model with approximately 7 billion parameters, capable of working with text and images simultaneously. It can analyze pictures, diagrams, videos, identify objects and their location, and convert visual data into a structured format. In addition, it demonstrates the best results among small open models on text recognition benchmarks.

The next step is processing plain text. This is where Gemma 3 27B comes into play — a large multimodal language model from Google DeepMind with approximately 27 billion parameters, capable of working with text and images. It has an extended context window, allowing it to process even very long documents or large data sets at once. The model analyzes the context, cross-checks the data, and verifies compliance with the conditions. For example, has the applicant reached the age specified in the licensing conditions? Or is the diploma valid, and is its date within the acceptable range? The model does not work with facts as such, but with logic: how complete the application is, which fields raise doubts, whether there are any inconsistencies.

The third stage is the formation of a conclusion. The model does not make decisions. It provides explanations: here is what has been found, here is why it may be critical, here is a recommendation. The official sees it in their own interface. If everything is in order, they approve it. If not, they return it to the applicant or initiate a manual check.

"The system uses a combination of models: Gemma 3 is responsible for text analysis and generating recommendations, while Qwen 2.5 is responsible for recognizing attached documents. It is the combination of several models and additional file processing that makes this solution unique. This makes it possible to verify the applicant's data against the submitted documents, check the validity of diplomas, and even confirm a change of surname if it does not match on the diploma and certificate of advanced training", explains Alexander Akulenko, head of the AI department at MK-Consulting.

All these steps are linked into a single system built on Dify, a tool for agent management and modular processing. Dify is what connects Qwen and Gemma, setting the order of operations, verification logic, and recommendation templates. This is where the most important thing lies: flexibility of updates. If a better model for working with the Ukrainian language appears tomorrow, it can be connected without rewriting the entire system.

The Ministry of Economy of Ukraine team was responsible for developing the system, from updating procedures to technical logic and AI module integration. The interface for business in Diia was created by the Ministry of Digital Transformation team. The AI part itself was created by the MK-Consulting team. The development was coordinated by the BRDO Office for Effective Regulation, and funding was provided by the European Union as part of the EU4DigitalUA project, implemented by FIAP. Consultative support for the preparation of regulatory acts was provided by the EU4Business: SMEPIS project.

Another important detail is the infrastructure

The entire process operates within the Ukrainian data center of De Novo, without any calls to third-party APIs. A cluster with H100 and two A100s was used in testing to ensure experimentation and stability. Two A100s are used in production mode: one serves Gemma, the other serves Qwen. This allows the system to withstand the load without reducing the processing speed, which is 2 to 15 minutes per request.

"The combination of NVIDIA A100 and H100 accelerators is a classic combination for powerful projects. The A100 is well suited for stable operation of large models in production mode, while the H100 allows you to experiment with new configurations and scaling. The combination of these resources allows you to withstand heavy loads and guarantee service quality. But most importantly, all computations take place in the Ukrainian cloud, which is directly related to digital sovereignty: the state receives world-class power without the risk associated with transferring data abroad", comments Gennady Karpov, Chief Technology Officer at De Novo.

What AI reveals and how officials see it

This is perhaps the most important thing. Because the model is not an independent expert or a closed tool, as in many commercial services. In the public sector, it is not only important what the model “thinks,” but also how it explains its conclusions. After completing the full verification cycle, the official sees a structured result on the screen.

They are presented with not just an “approve” or “reject” button, but context. What has been verified, in which field the error was found, and what exactly it is related to. If one of the diplomas contains a surname that does not match the one in the application, this discrepancy is highlighted separately. If the applicant's age does not meet the minimum requirement, the field will be highlighted. If the confirmation of qualifications is not valid, the expert will be notified.

The response format is not a rigid template. It is designed to leave room for official analysis. The model does not dictate, but suggests. It identifies weaknesses, explains the logic, and allows the employee to rely on refined, neatly organized analytics. There is nothing automatic in the decisive sense here. Every decision remains with the person. But thanks to the new module, this person receives a high-quality tool — not a substitute, but a digital magnifying glass that reduces the workload, increases accuracy, and reduces risks.

The results of the implementation are already noticeable. There are fewer resubmissions. There are fewer returns due to technical errors. There are fewer questions from businesses asking, “What's wrong with my application?” Because now the answer is obvious — both for the system and for the person working with it.

What's next and why this story is important

A veterinary license is just the first step on a long road. It is, so to speak, a benchmark implementation on which something bigger can be built. The team is now preparing to expand to other licensing procedures. The logic of scaling is already in place: modular construction, flexible orchestration, neutrality to a specific model, and the availability of a proven infrastructure. All this allows us to build a new generation of state decision-making systems.

At the same time, AI itself is developing. Experiments are being conducted with retraining models on Ukrainian scans (not all types of handwritten documents are currently processed successfully). Multimodal approaches are being studied for future more complex cases — when a single application may contain text, graphics, tables, stamps, and arbitrary data formats.

The idea is not to create a “permit bot,” but rather a new type of government interaction, where AI enhances the expert, reduces routine work, and frees up time for truly important tasks. The digital state operates at a deep level of the process without compromising security. This is exactly what we have managed to achieve with e-Permit: open models, local infrastructure, guaranteed data residency and sovereignty, transparent logic, and full legal compliance.

However, this is just the beginning. As one of the project participants noted, we didn't prove that we could do it. We just did it.