Artificial intelligence projects often start with an ambitious idea, but only a small fraction of them reach the stage where the solution becomes a stable part of the system. This is where R&D (research and development) comes in, enabling the transition from “might work” to “works reliably.”
For organizations seeking real value, it is not enough to apply a standard model or combine a few technologies. A systematic process is required, in which the solution is tested under real-world conditions, with real data, real processes, and real errors.
“The biggest mistake is assuming that a model that works in a demo environment will automatically work in an organization. This is usually where the problems begin,” says Aistis Raudys, an artificial intelligence solutions developer. 
Why does artificial intelligence require context?
Artificial intelligence is not a one-size-fits-all solution that can be “implemented” identically everywhere. The same model can yield completely different results in different organizations. This is determined by data structure, processes, user behavior, and even how the solution is used. The R&D phase allows these differences to be identified in advance. Instead of guesswork, there is testing, hypothesis verification, and a clear understanding of where the solution works and where it does not.
“Artificial intelligence is always about data. If you don’t understand its limitations, you won’t understand the model’s limitations either,” emphasizes A. Raudys.
R&D and Reliable AI Solutions
One of the most common challenges is models that perform perfectly in a testing environment but begin to fail in real-world conditions. Seasonality changes, new data types emerge, user behavior shifts, and the solution becomes unstable. In the R&D process, such scenarios are modeled in advance. We test how the system reacts to changes, where its limits lie, and how those limits can be managed.
Another important aspect is explainability. It is not enough for organizations to know that the model “works.”
They want to understand why it makes one decision or another.
“If you cannot explain the model’s decision, you cannot fully control it,” says A. Raudys.
From Problem Definition to a Working Solution
A mature R&D process always begins with a very clear question: what problem are we solving, and how will we measure success? Only then do we move on to data analysis, quality assessment, and preparation.
At this stage, it often becomes clear that:
- some of the data is misleading;
- important information is missing;
- hidden biases exist.
Only after resolving these issues can we move on to model development and experimentation. Here, not only accuracy is important, but also speed, stability, and the ability to integrate the solution into real-world processes.
How does R&D work in practice?
This principle is clearly evident when working on projects in the healthcare sector, where the cost of errors is particularly high and processes are complex. One example is the “Medicall” solution developed by AAI-Labs, which is designed to optimize call center operations in high-volume medical facilities. The problem here is very specific: thousands of calls per month, long wait times, and a heavy administrative burden.
The solution was not built “from the technology up,” but rather from an analysis of the real-world situation. They assessed how calls were handled, what recurring questions arose, and where operators spent the most time.
“In projects like this, the most important thing isn’t the complexity of the model, but whether it actually reduces the human workload,” notes A. Raudys.
The system combines several components:
- real-time speech recognition;
- contextual response generation;
- automated appointment scheduling;
- creation of conversation summaries.
Technologically, advanced models (such as “Whisper” or “Wav2Vec2”) are used, but the real value comes from adapting them to a specific context: the Lithuanian language, medical terminology, and real-life patient scenarios. The result is not only faster service, but also a reduced administrative burden and a better patient experience. 
AAI Labs R&D Projects
These principles come to life through concrete projects that AAI Labs carries out together with research institutions and business partners.
We develop technology across a range of projects:
- AIDER – Autonomous Drone Delivery (Eureka-II, 2026–2028) – AI-controlled system for heavy-lift drones delivering critical supplies to hard-to-reach areas.
- Large Language Models for Medicine (ERDF, 2024–2026) – Specialized medical language model for healthcare documents, deployed in compliance with sector regulations.
- Automated Caller and Assistant System (ERDF, 2024–2026) – Automated call assistant operating in three languages, real-time prompter for operators, and emotional voice analysis module.
- 5G Innovation in Transport – Kautra (NextGenerationEU, 2024–2025) – Integrated intercity transport platform: event detection, ticket scanning, discount card recognition, and dynamic pricing.
- 5G Drone Base Stations for Urban Monitoring (NextGenerationEU, 2024–2025) – DBOX M2 5G and DBOX Mini 5G systems for infrastructure monitoring; the world's first 5G station with an automatic drone battery swap function.
- 5G and AI Public Transport Optimization – JUDU/Vilnius (NextGenerationEU, 2024–2025) – Transport optimization system from 5G equipment research through to IoT model deployment.
- FUN-KB – AI Accounting Assistant (EU Funds, 2025–2027) – LLM-based accounting assistant in Lithuanian, targeting over 10,000 accounting firms in Lithuania.
A more complex challenge is understanding medical language
Another project demonstrates how R&D enables the creation of more advanced solutions—specialized large language models (LLMs) designed for the analysis of medical documents. This project, implemented in collaboration with Santaros Clinics, addresses a complex challenge: how to automatically understand medical language, which is highly specific, context-dependent, and error-prone.
The system is capable of:
- structuring medical documents;
- generating summaries;
- transcribing doctor-patient conversations;
- recognizing medical terminology in real time.
The technological solution combines LLaMA architecture models with language recognition technologies such as "Whisper" or "wav2vec 2.0." However, the real breakthrough was not in the technology, but in the data (the models were trained on real medical records, allowing them to understand context rather than just words).
“In medicine, it’s not enough to recognize language, you need to understand its meaning. A single inaccurate term can completely change the interpretation,” emphasizes A. Raudys.
Solutions of this type allow for a significant reduction in documentation time, faster analysis of information, and a lower probability of human error.
FAQ
What is R&D in AI projects?
In the context of AI, R&D (research and development) refers to activities aimed at creating new or significantly improving existing artificial intelligence solutions, algorithms, or technologies.
When is an AI project considered R&D?
When an unclear, technologically complex problem is being solved and a new solution is being created, rather than simply implementing an existing tool.
What types of AI projects most commonly fall into the R&D category?
These may include the development of new models, the creation of unique data processing methods, complex forecasting systems, or innovative automation solutions.
Is chatbot development considered R&D?
Generally, no. If off-the-shelf solutions are used, it is not considered R&D. However, if a unique model or algorithm is developed, it may be classified as R&D.
What are the benefits of R&D in AI for businesses?
You can take advantage of tax incentives, increase your competitive advantage, and create unique technologies that your competitors do not have.
What expenses can be classified as R&D?
Employee salaries, technology development costs, testing, prototyping, data collection, and analysis.
How can R&D activities be justified in an AI project?
You must document the problem, the research process, experiments, and results, and demonstrate that the solution is not standard or readily available on the market.