Werden kleine KI-Modelle die Zukunft dominieren?

We have organized an expert conference about The rise of AI in costumer communication in collaboration with our partner Feedyou. It took place at the Prague Microsoft building on June 17, 2025. More than 150 managers from contact centers and call centers of mostly large companies and corporations attended. Both companies, SentiSquare and Feedyou, specialize in handling customer communications using artificial intelligence, and the companies' teams include analysts and machine learning engineers with decades of experience in the field. Together with clients O2, Bezrealitky.cz and Pelikán Travel, who are already using AI in their work, we showed at the conference how far customer communication processing using AI has come and where it is heading.

Didn't get a chance to get to the conference? No wonder! Tickets have been gone within 2 weeks of registration opening. Dive into the conference highlights and breakthrough ideas with us now!

Both organizers of the event are dedicated to the field of artificial intelligence. SentiSquare works as an artificial intelligence for evaluating customer communications. It focuses on advanced analytics and automation of customer interactions using NLP, a field that uses AI algorithms to process and understand text. Feedyou are experts in connecting voice technology and AI. Thus, it brings AI assistants such as AI voicebots or AI chatbots to the market.

Voicebot can have your own voice now

The first presentation was about AI voicebots. Vojtěch Dlouhý from Feedyou introduced their second, quicker to learn, generation of voicebot, and he impressed the audience by claiming that 2025 is the year of the rise of AI voice agents. The introduction of AI voicebots into companies is greatly accelerating communication. Modern voicebots are adaptable and advanced in emotion, tone, laughter, accent, whispering, or even responses when the speech is being interrupted. You can even generate a voicebot based on your own voice. As a result, the caller experiences communication that feels more human and natural than ever.

The lecture was complemented by two case studies, from Bezrealitky.cz and Pelikán Travel, to present what information can such virtual AI assistant process. Specific examples showed that thanks to more than 10,000 queries handled even outside working hours, the work for operators is easier and managers can shorten their shifts. The speakers presented the entire journey of deploying AI assistants and agents necessary.

But real contact is still a must in customer service centers. Vojtěch Dlouhý says "Quality outcomes will only come with quality data, the right approach and a long-term strategy," proving that technology still won't fully replace humans.

Artificial intelligence built on data, research and experience

The second presentation was delivered by Tomáš Brychcín, CEO and CTO of SentiSquare. Tomáš stands behind the idea and development of the No-Code NLP platform SentiSquare. Our R&D team has more than 20 years of experience in natural language processing (NLP), with over 200 publications and more than 4000 scientific citations under their belts. In just a few years, we have successfully transformed the development of artificial intelligence in natural language processing into a business. The result of our research is AI that understands and analyzes text in any language in depth.

Tomáš Brychcín says: "AI models are and will be a commodity. The future lies mainly in the efficient production, training and operation of specialized AI models. I believe that there will be a great development of so-called No-Code and Low-Code platforms to facilitate working with AI without the need for expert knowledge."

Small specialized AI models are ideal for in-depth analysis.

You can visualize a language model as the brain of an artificial intelligence. Its job is to understand the text and predict the next word in the sentence. The breakthrough came in 2017 with the emergence of so-called Transformers, special neural network architectures. Since then, bigger and bigger language models capable of understanding more complex patterns in text have started to emerge.

During his talk, Tomáš Brychcín presented the differences between LLM (large language models) and SLM (small language models). It is the large language models that dominate the media space today, but their development has reached a stage where the differences between them are becoming negligible.

Nature magazine reported that the AI race in 2025 is tighter than ever before. At the moment, the best large language models are only minimally different from each other. The differences between the top 10 players are less than 5%.

But the field of AI is no longer just about the big language models. The trend has resulted in smaller specialized AI models, SLMs, that focus on a selected specific area where they are faster and more efficient. “Small, task-specific models provide quicker responses and use less computational power, reducing operational and maintenance costs,” says Sumit Agarwal, Vice President Analyst for Gartner. Andrew Ng, a world-renowned AI expert, says, “The future isn’t about bigger models, but better-aligned ones.”

Small task-specific AI models (SLMs) are ideal for applications that involve a high volume of interactions, low latency requirements, and data-sensitive scripts. Specialized SLM models can be used on-premise, in systems already in use, or stored on a private cloud.

Gartner, an IT research company, states: “By 2027, organizations will use small, task-specific AI models three times more than general-purpose LLM.”

The development of specialized SLMs is what we focus on at SentiSquare. Our models can gain macro insights in the areas of general operations, CX, communications quality, sales or compliance. Top large companies can rely on the models for quantification, trend monitoring, context finding or control.

This extensive analysis enables our global and innovative clients such as Kiwi, FTMO, E.ON or Kooperativa to get valuable feedback. Companies are thus able to increase customer satisfaction, reduce churn rates to competitors and minimize other threats that a thorough analysis can save them from.

How O2 uses the SentiSquare Analytics Tool to Boost Operators’ Sales Performance

At O2, as part of a successful collaboration, the SentiSquare platform helps analyse big amounts interactions per month, giving them 100% control over more than 5 million customer interactions. This NLP enables fast and accurate data processing without the need for technical expertise. O2 is now using our platform to create their own specialized SLM models without coding. It c

urrently uses over 30 custom SLM models. This is a local installation of the platform that does not require specialized hardware due to the SLM approach. It is the local operation that allows to analyze the quality of all calls in almost real time. A practical view of the use of AI directly in O2 was presented by Senior Manager Digital Customer Care O2 Roman Křivka.

"At O2 we use small AI models for classification tasks where annotated data is available. Large models are then used when we need to understand text in depth or generate summaries," Křivka explains.

Flexibility and the ability to customize these small models are key for evaluating specific questions related to O2 operators' calls with their customers. The customer care manager thus has a thorough overview of each individual operator's approach to the customer, but also monitors, for example, such specific situations as the keeping of a promise to a customer. Thanks to comprehensive and in-depth analysis, O2 can evaluate specifically what products are selling well, but also which operators are performing well and how they respond to objections. At the same time, they also measure customer satisfaction.

Roman Křivka presented evident benefits brought by cooperation with SentiSquare. The results speak for themselves. 100% of interactions under control. 9-point increase in NPS. 7% improvement sales performance. Thanks to the SentiSquare No-Code NLP platform, O2 monitors all operator communications and leverages them to the company's advantage.

Generative models will no longer be enough for large companies

The programme concluded with a panel discussion, during which a number of questions were asked about AI models tailored to the specific needs of companies. It thus became clear that there is a growing interest in the smaller language models that we create at SentiSquare.

The conference confirmed that using only one AI model will not be enough for large companies. The ideal is to combine both kinds of models. Gartner notes: Neither model, big or small, is self-saving. It will be necessary to combine the strengths of LLM and the speed of SLM to meet different needs. Going forward, linking generative and non-generative AI will become increasingly common. Often, where one model is stronger, the other is weaker. This makes them complementary.

Sharing specific results from O2's customer centre and showing how advanced their analytics are was a valuable contribution. O2 presented how they combine their models, how they use AI to improve NPS and how AI analytic tools have improved their sales performance. Previously 100% call and communication processing was unrealistic, now it's a necessary component. Large language models provide broad coverage and general text understanding, while small specialized models excel in speed and accuracy in select tasks. The two technologies complement each other, allowing companies to leverage the strengths of each language model. And that's exactly what we've been doing at SentiSquare for years. This is the future!

The role of artificial intelligence in business in 2025 is no longer a question, but a statement. Tomáš Brychcín adds that AI/data analyst will be the desired profession of the future.

‍