– CyberNet AI currently has offices in Uzbekistan, Kyrgyzstan, the United States, Mexico and Egypt. Could you tell us more about what your product is designed to do?
– We develop platforms for voice agents. These are enterprise-grade platforms designed for large organisations that require reliability, regulatory compliance and integration with internal systems. Our clients include banks, telecommunications operators and government agencies.
For example, our platform is used by the 1414 unified contact centre. It enables eGov to automate responses to citizens’ enquiries about accessing government services. Automation already covers more than 100 services.
The platform is integrated with internal knowledge bases and government services. When citizens call with questions about obtaining certificates, documents or digital signatures, our agents explain the necessary procedures.
The effectiveness of voice agents becomes particularly evident during peak hours at contact centres. They reduce the workload on human operators and help serve all callers promptly without increasing staffing levels. According to CyberNet AI's data, our agents replace the equivalent of at least 6,000 operators across Central Asia.
Many customers still ask to speak to a human operator. But this behaviour dates back to a time when voice agents were less capable: they could not answer questions properly and struggled to understand context. They would often respond to callers by saying, "Could you please repeat that? I didn't understand."
Today, agents are becoming more intelligent, and their responses provide real value. Customers now have a choice: wait on hold for three, five or ten minutes, or receive assistance immediately. When an agent successfully resolves a problem, people's attitudes towards it begin to change.
In sales, agents can also offer greater flexibility. They can instantly access various databases maintained by our partners and suggest the most suitable options to customers. The goal of a voice agent is to guide customers effectively through the sales funnel, increasing conversion rates.
Many of our clients have seen conversion rates increase by 10–15%. That is a very good result. At the same time, an AI agent costs a company at least half as much as a human operator, whose employment involves expenses for workstations, salaries, bonuses, taxes and software licences.
– In Kazakhstan, callers often switch between languages during a conversation. How do voice assistants adapt to this?
– This was a major challenge for our company.
We were the first in Kazakhstan to develop a model capable of understanding two languages within the same conversation: mixed Russian and Kazakh speech. We subsequently developed solutions for mixed Russian-Kyrgyz and Russian-Uzbek speech.
Hundreds of thousands of hours of audio recordings were used to train these models. This expertise later proved valuable in Mexico, the United States and Egypt.
Our team is now developing a solution with ready-made models and templates for small and medium-sized businesses. Using a no-code builder, customers will be able to select suitable options, add information about their business, give their agent a name and launch it quickly.
Our goal is to create simple, intuitive templates that can be used by clinics, schools and SMEs.
– One application of voice agents is debt collection. How do you design their interactions with vulnerable customers?
– For people working in debt collection, the job carries a risk of burnout. Operators regularly communicate with customers who are facing difficult financial circumstances and are unable to repay their debts.
We adopt the same procedures and processes used by our partners, including banks and microfinance organisations. They establish the rules for communicating with customers, determining when a polite reminder is sufficient and when a more persistent approach is required.
Based on these rules, we configure our agents to follow the same communication scenarios as human operators.
– With telephone fraud becoming a widespread concern, people are increasingly cautious about calls from unfamiliar numbers. How do you help customers verify that they are genuinely being contacted by a bank or another organisation?
– We understand these concerns. Our agents call from the official telephone numbers of banks, telecommunications operators or government agencies. We only use numbers belonging to our client organisations.
If a caller suspects fraud, we advise them to contact the bank directly using its official telephone number to verify the information.
– What options do you offer customers for using your platform?
– CyberNet AI offers several pricing and deployment models. Customers can purchase the platform and create agents independently within their own IT infrastructure. Alternatively, they can use the platform through a cloud-based subscription. The product can also be purchased for a specific project.
For example, we had a client who needed to contact 350,000 people ahead of the Paris Olympics. Using an AI agent, we completed the task within a week. We obtained valuable insights for the client that they would not have been able to gather independently within such a short period.
– What are the main technical differences between the previous generation of AI agents and the current one?
– The new generation of AI agents requires graphics processing units (GPUs). These are expensive, and access to them is limited. We therefore reserved GPU capacity at a data centre in Kazakhstan three to four months in advance and signed a one-year contract guaranteeing access to these resources.
The contract does not allow us to cancel the arrangement, but we are satisfied with the terms. At present, this computing capacity costs us less in Kazakhstan than it would in the United States.
Previous-generation agents run on conventional servers using CPUs, while newer agents powered by generative AI require GPUs.
These newer agents are more expensive but offer customers greater capabilities. They are more flexible and provide a wider selection of voices. They can also be created in minutes, a capability already available through the platform we are providing to customers for pilot testing.
– How frequently do voice platforms need to be updated?
– Updates are continuous. A voice platform consists of four main components.
The first is the orchestrator, which manages call parameters. The second is the speech recognition module, which converts the caller's speech into text and passes it to a large language model (LLM). The LLM generates a response, which is then spoken aloud by the fourth component, the text-to-speech synthesis module.
All these components require continuous updates. Some processes are automated, while others require manual analysis and configuration.
For example, it is essential to monitor the delay between a customer's question and the agent's response. According to our estimates, CyberNet AI's platform has one of the fastest response times on the market, at less than one second.
We continuously update both our models and their configurations. CyberNet AI's engineers and developers have specific KPIs, such as achieving 90% speech recognition accuracy in a particular domain or keeping response times below 0.7 seconds.
– What proportion of CyberNet AI's work is performed by AI?
– At present, we write approximately 60–70% of our code with the assistance of AI, but testing is carried out by our specialists. When large amounts of code are generated, it is important to verify that the results meet the requirements. That is why human testers remain essential to our processes.
We also use AI agents in sales. On one occasion, we needed to contact 400 companies. Doing this manually would have taken considerable time, so we created a voice agent.
By running two or three iterations each day, we developed a sales approach that increased the conversion rate into requests for commercial proposals to 25–35%.
During the initial tests, this figure was around 1%, possibly even lower. However, rapid iterations make it possible to refine a sales pitch until it captures the customer's interest and keeps them engaged in the conversation.
– Which operations are you not prepared to entrust to AI?
– I meet with clients personally because we operate in the B2B market, where it is important to convince customers of the value of our solution.
At the same time, AI significantly accelerates sales preparation. Previously, creating a simple demo took around two weeks. Now I can prepare a fully functional demonstration in a single day.
All of this remains interesting as long as it delivers value. Today, AI reminds me of the internet in the 1990s. We are experimenting with new products, searching for markets, gaining experience from users and gradually discovering what people really need
