Kazakh Founders Secured 7 Million Tenge From Clients Before Even Launching Their Product and Built an AI Platform for Businesses

Arthur Krivtsov and Nurlybek Maratov were running an IT studio and building custom projects for clients. One day, a client asked them to train an AI bot on a large volume of corporate data.That request made Arthur and Nurlybek realise this was not just a one-off job, but a common problem for businesses. So they went on to create Barion AI, an AI platform that helps companies launch AI agents quickly.

As part of the special joint project by Digital Business and Astana Hub, “100 Startup Stories from Central Eurasia”, Arthur and Nurlybek shared why companies struggle to bring AI assistants into their workflows, how hundreds of documents can be turned into an “AI brain”, and what happens when a robot sales rep calls thousands of customers in just half a day.

“The product didn’t even exist yet, but the money was already in the bank”

— How did you two meet, and what were you doing before Barion AI?

Arthur: — Nurlybek and I studied software engineering together at Narxoz University in Almaty. We even wrote our thesis together. It focused on development in the field of augmented reality.

Right after finishing our bachelor’s degree in 2021, we built a small prototype for an AR project aimed at furniture stores. It let people ">see how a piece of furniture would look in their own home. We applied to Astana Hub’s Seed Money programme with that solution and managed to secure funding to build an MVP. So, straight out of university, we had our first startup.

We landed several major clients among furniture stores. Before long, the project had paid for itself and even made a small profit, but we never managed to achieve explosive growth in Kazakhstan. We went to Dubai and tried to break into the UAE market, but because we lacked sales experience, we eventually had to shut the company down.

Still, we had a team, experience and contacts, so we reshaped the business into a classic IT studio and started taking on outsourced projects.

Arthur Krivtsov

— How did you decide to build a new startup? 

Nurlybek: — One day, we got a request from an education company. They wanted us to train an AI bot that could check homework and answer students’ questions. But there was a huge amount of information involved: 50 courses, each with its own lessons, homework and tests. To train just one AI assistant, we would have had to feed it a massive amount of data, write hundreds of prompts, and spend a lot of time and money on it. At that point, we did not have a ready-made product or a clear understanding of how to make it work, so we said no at first.

Arthur: — Later, we came back to that case and thought, “What if we actually took it on?” We realised it was inconvenient to train a separate AI assistant for every course, department or task. But if we built a single platform where a business could upload all its information, including documents, policies, sales scripts, training materials and knowledge bases, the system could structure that data and turn it into one shared “AI brain”. After that, the client could simply choose the role they needed for the bot: a sales manager, a tech support specialist or a teacher checking assignments. That was the idea that kicked off Barion AI.

– What did you do at the very beginning?

Arthur: — We sketched out the technical architecture and worked out who on the team could actually build it. Then we called the client, said yes, and the very next day they paid the full amount upfront. So the product did not exist yet, the idea was only a day old, but the money was already sitting in our account.

Then we thought of a few other companies that had come to us with similar problems before. We called each of them and offered them the chance to get in early on better terms. That’s how we brought in almost 7 million tenge in prepayments, which became our development budget.

“High-quality voice AI in Kazakh needs to be built in Kazakhstan”

— LLM models and voice engines already exist. So where does Barion AI’s know-how come in? What exactly have you automated?

Arthur: — Usually, businesses have to manually pull together documents, policies, sales scripts, FAQs, messages and product databases, then structure all of it, write prompts, test the answers and set up separate scenarios. That is why, according to McKinsey, launching a single AI bot in a large organisation can take up to six months.

Over time, we realised that our killer feature was auto-training AI agents. You can upload your files, and the system structures the information itself, then turns it into a knowledge base for the bot.

We use different LLM models, including Gemini, Claude and ChatGPT, depending on the task. But the algorithm that collects the data, structures it and trains the bot is our own development.

Nurlybek: — Overall, we see Barion AI as a platform where a business can choose the scenario it needs, click a button and create a bot without having to write any prompts.

Right now, we’ve decided to focus on voice AI agents because, for businesses, this is one of the easiest use cases to understand. A bot can call customers, qualify leads, handle upsells and pass already processed requests on to managers.

We also connect voice engines through third-party APIs. But in the future, we plan to build our own voice model. It makes sense because global models reproduce Kazakh very poorly and unnaturally. I’m sure that high-quality voice AI in Kazakh needs to be built in Kazakhstan, and we hope to play a part in that.

— Why did you choose this direction?

Arthur: — This year, we were selected as one of six startups from Kazakhstan to travel to Silicon Valley for Hero Training, an international accelerator programme run by Astana Hub and Draper University. We came to the mentors and investors in the US with one clear question: how do we turn our technology into a product that can scale? Because selling the tech on its own is tough. A system that builds AI bots for automation can sound too complex and a bit too geeky.

So now, we’ve decided to go to market with ready-made scenarios, where the client buys a clear solution to a specific problem. We started with voice calls.

“People can mess with AI”

— How does the call scenario work?

Nurlybek: — You log in to your account on the platform, upload the information, choose the call scenario, set it up, upload the contact list and launch the campaign. Afterwards, you get a detailed report. The whole thing can be done in just a few minutes.

Let me explain using a recent case with a consulting company that provides legal and accounting services. They had a database of cold contacts, so we set up a voice robot to make the calls, address people by their first name and patronymic, and ask follow-up questions. In some scenarios, once the client gave consent, the bot would send a presentation via WhatsApp. In the end, the client received a detailed report with call recordings, transcripts and leads sorted by how “warm” they were.

The campaign covered 5,000 contacts. We kicked it off around lunchtime, and by the evening the client already had a full report. Just imagine how many call centre agents, and how many hours it would take to get through that volume manually in half a day. The campaign led to around 30 signed contracts for the client, and just one of those sales covered the full cost of the whole campaign.

— Do people need to be told they’re speaking to a robot?

Arthur: — People naturally prefer speaking to other people. Once someone realises they’re talking to a robot, they might either hang up or start messing with the AI.

Our agents sound very natural, so only around 3 to 5% of people realise there isn’t a human on the other end of the line. And of course, from a sales performance point of view, it’s better not to make a big deal of the fact that the call is coming from software rather than a real operator.

In general, not everyone can tell AI apart from a real person. For our own company, we generated an avatar of a beautiful girl, and suddenly we started getting messages on our work WhatsApp at night with questions like, “Diana, are you married?” We had to remove all the emojis from her replies so she wouldn’t come across as too friendly. And then, of course, we had to disappoint clients who came to the office and asked, “So where’s Diana?”

— What if someone realises they’re talking to an AI agent and starts trying to persuade it to give them a big discount?

Arthur: — Our bots are programmed not to drift away from the purpose of the conversation. Because they are trained on the company’s own policies and previous chats, the robot sticks strictly to the rules it has been given. So if there were no cases in past conversations where a customer was given a discount for no reason, the AI will not give one either.

Nurlybek: — But in general, you do need to keep an eye on what AI assistants are writing. We had a funny incident when we were testing a text bot on WhatsApp. It invited a client to a meeting at our office at 11 am, but we had no idea, and there was no one in the office. The client arrived, sent us a photo from the waiting area and wrote, “I’m here, where are you?” And the bot instantly replied on WhatsApp, “Please wait, someone will come down for you now.” And then it did the same thing several times in a row. After that, we had to apologise personally, explain what had happened and smooth things over.

Why did that happen? Apparently, the bot had learned from our old conversations, where we used to arrange meetings with clients at the office and come downstairs to meet them. But these are all isolated technical bugs, and we fix them quickly.

“An AI agent searches a car dealer’s database with 20 branches in seconds”

— How many people are currently working on Barion AI?

Arthur: — Nurlybek and I are the project’s co-founders. We both have IT backgrounds, so we combine the roles of technical specialists and entrepreneurs.

But the foundation of our technology was built by our lead developer, Eldar Khaibullov, who is an experienced IT specialist.

So you could say there are currently three people actively working on the product.

— How are you selling your solution?

Arthur: — Most of our clients come through referrals. They’re either companies we’ve worked with since our first AR startup, or former clients from our IT studio.

For larger enterprise clients, we bring in a good contact of ours. He knows the IT industry inside out and has access to senior decision-makers at major companies in Kazakhstan. He was the one who helped us close our first big deals.

— What is your monetisation model right now?

Arthur: — At first, we worked on a contract basis: we would build a solution for a specific client and then roll it out for them. The company is now making over 100 million tenge a year in revenue, but we’re moving towards a pay-as-you-go model, where clients pay based on usage. In practice, the client uploads their data to the platform, connects the scenarios they need and pays for the tokens they actually use. We don’t have the final prices yet, as we’re still working on the pricing model.

— What business segment are you targeting?

Arthur: — We decided to focus on companies with 500 or more employees. Businesses of that size already have all the files needed to train the system. At the same time, they have a real need and a real pain point: training AI agents quickly. That is exactly what we can help with. We already have a ready-made platform, so they do not have to spend time and effort on testing and implementation.

We started out in education, but later realised the product was relevant for all kinds of businesses. For example, for a large Kazakh car dealer with 20 branches across the country, we integrated our technology with their huge database of used cars. Now, if a customer in Atyrau is looking for a car, the system can take their description and preferences and find a suitable vehicle in that city, or anywhere else in Kazakhstan, in seconds. It also recommends similar cars from other brands with comparable specs. On top of that, it automatically forwards the message to the right manager at the relevant branch.

We’d love to run a couple of cases in the oil production sector, as we know companies there deal with huge volumes of documents and regulations. Rolling out AI systems in that kind of environment can be tricky, and that’s exactly where Barion AI could help.

– How do you handle data security? For large businesses, that is usually one of the most sensitive topics.

Arthur: — First of all, the AI bot will use public information to communicate with clients anyway, things like prices, service terms and presentations. We don’t take or share customer contacts or data with third parties. They stay within the company’s own environment.

Secondly, if it’s absolutely critical to keep everything on your own servers, we can easily deploy our solution locally.

“We want Barion AI to reach the IPO stage”

— How are you funding the project?

Arthur: — We stick to a bootstrapping philosophy: we grow using the money we earn ourselves, without taking out loans or bringing in venture capital.

However, during the programme in Silicon Valley, we realised that if we want to grow quickly and scale into global markets, we will need investment. But we’ll only start looking for investors once we have fully fine-tuned the unit economics of our new business model, which is based on selling ready-made scenarios.

– What are Barion AI’s plans for the near future?

Arthur: – Beyond Kazakhstan, we’d like to work in the US market, as well as in the MENA region. Hero Training had a big influence on that. It showed us just how large and important the US market is for AI startups.

— How do you see the startup in the long term?

Arthur:— The government is doing a lot to support the tech sector. Kazakhstan is now among the top 30 countries for ChatGPT usage, and the country is also building its own national LLM. Sooner or later, Central Asia will produce a trillion-dollar startup that goes public on an international exchange. We’d love Barion AI to be that company.

Nurlybek: — Tech moves incredibly fast. Back in 2021, when we were working on augmented reality, Apple Vision Pro had not even been announced yet, and the market looked completely different. Now we’re building AI agents. It’s hard to say what the next five years will look like, but we definitely want our projects to grow in scale and our solutions to bring real, practical value to businesses.