Bank apps in Central Asia are already more convenient than the banks themselves: the client side is digitized, while manual operations still remain in the middle and back offices. What part of AI in banking is hype, and what yields measurable results? How do you calculate the cost of a resolved ticket, and why will part of daily banking shift to personal agents? We discussed this with Roman Nadein, co-founder and CEO of Rocket Tech and Ready Bank.
How do you assess the current level of digital transformation in Central Asia’s financial sector? Have we passed the “basic digitization” stage?
The client side is digitized much better than the banks themselves. In Kazakhstan, according to Q1 statistics, over 90% of cashless card payments went through mobile banking. This is an established model, not an experiment.
But a convenient app and a digital bank are not always the same thing. A great interface can still hide manual actions and data transfers between systems. The middle and back offices are not fully digitized: simple operations are automated, but some processes run semi-manually.
So, on the client side, the basic stage is complete. The next step is not adding new features, but refining existing ones: removing the unnecessary and making processes seamless, simple, and convenient. Inside the banks, there are processes that can and should be automated, and AI enables exactly that.
Comparing key markets—like Kazakhstan and Uzbekistan—what are the fundamental differences in how local banks approach innovation? Where do you see the highest potential for AI adoption right now?
You can’t say Kazakhstan and Uzbekistan are at the same stage. The gap between two banks in the same country is sometimes wider than between the countries.
Kazakhstan. Digital services are already built here, so AI is needed for efficiency: making operations cheaper, faster, and more convenient for the client.
Uzbekistan. The market started from a low base and is now catching up. Banks can build AI directly into new digital services and move faster because of it. State participation slows things down: as of mid-summer, about 62% of sector assets belonged to state-owned banks, where changes are harder to implement. Some banks are being privatized, which is the right move. There are already good cases: this year, TBC Bank launched an in-app AI assistant that works in Russian and Uzbek and can perform some banking operations.
For the entire region, the quality of local language processing is crucial. A client should be able to explain a problem in their own words—for example, “I was charged twice”—and the assistant must understand the request, find the transaction, and resolve it. Agents are moving to voice, as recent Meta and OpenAI presentations show, which introduces accents, speech speed, and noise. Quality should be tested on real requests from local clients.
The potential exists in both markets, but it is measured not by a “we launched an agent” presentation, but by the value to the end client.
What main internal barriers do financial institutions face when trying to transition from classic banking to AI-driven ecosystems?
The main barrier is security and accountability for the result. It has two components.
Development security. Banks already actively use AI agents to write code and tests. It is important that the AI development cycle is built through application security practices. The question is how the bank monitors this: does it just buy subscriptions for developers who push whatever they want to the shared repository, or does it have a platform with checks?
Such platforms usually have guardrails—limiters that check what data goes into the model and what comes back. They prevent the leak of personal data, secrets, and other sensitive information. There is now a term for this: harness, the entire framework built around the model.
Client data security. If a client interacts with AI inside the app, or AI processes tasks within the app, you need to know which models receive the data. Usually, it comes down to hardware. If a bank has money, it deploys models within its own perimeter, and the regulator monitors this. But some requests might still go via API to an unprotected perimeter, to models from giants like Anthropic or OpenAI. Even if anonymized, the question of how correctly this is implemented remains.
The problem is known, and tools for it already exist. We also help implement anonymization solutions: the context is preserved, real client data does not go to the model, and when the response is received, the necessary data is substituted back.
AI in banking—where is the line between marketing hype and real, measurable business results right now?
For me, the boundary is the result. If an assistant resolves a ticket, cuts transaction time, and reduces the load on support, it has business value. If the client still has to go through the same manual path after the conversation, the effect needs to be proven.
AI helpers and assistants in banking apps are more likely hype. Essentially, these are wrappers over existing models that you can chat with. But it is unclear why you would do this inside a banking app if the same thing is available outside.
Anti-fraud and scoring deliver real results. AI anti-fraud systems monitor transactions and stop fraudulent ones in time. Scoring becomes faster and more effective, helping the bank know its client better.
Processing client requests has great potential. Suppose a person asks why a payment failed. In a good scenario, an assistant with the necessary permissions checks the transaction status, explains the reason, and helps execute the next step. If an employee is needed, it passes on the already collected information so the client doesn’t have to explain everything again. The effect can be verified: how many tickets were closed, how long it took, and how many clients returned with the same problem.
Another application point is analytics. I have yet to see a single banking app with convenient income and expense analytics. This is proven by the very existence of separate financial tracking services like CoinKeeper, where users export their data. Banks try to do this internally, but it looks raw and inconvenient. With AI, this can be made genuinely user-friendly.
Based on Rocket Tech’s product experience and cases, in which banking processes does AI adoption currently provide the fastest and most tangible “wow effect” that a user notices from the first click?
I would start with scenarios where the client has to navigate multiple screens for a simple operation. Transfers, statements, bill payments—here the effect is immediately clear. Then we verify it with data: how much time was saved, how many people completed the operation, and how many errors occurred.
For a first experiment, I would choose voice control: for transfers, statements, and sending invoices. You say, “send five thousand to mom’s card,” the agent prepares the transaction, and you confirm. It sees my transactions and knows my contacts, so I don’t need to go through five or six screens to make a transfer or pay for parking. This would cover my real daily tasks.
This scenario has a few caveats:
- Audience. You need to look at the bank’s user demographic: how young and active they are. The older generation is more conservative; they will do everything manually and double-check.
- Business metrics. This is unlikely to improve efficiency or increase the number of transactions. But it can capture some hype and audience: the bank is perceived as digital and AI-adopting.
- Security. This requires separate work so users don’t worry. Such features are easy to build on local models within the bank’s perimeter.
- Communication. PR and marketing support is needed to explain how it works and why it is safe.
AI solutions are often seen as expensive investments with an unclear payback horizon. How should a bank or fintech company correctly calculate the ROI of such initiatives in the early stages?
Yes, it is indeed an expensive investment with unclear payback. To calculate ROI, you need to separate two types of implementations.
AI in internal development processes. This is the product delivery cycle: development, hypothesis testing, test coverage. Efficiency is easier to calculate here. We look at the number of features released, their release speed, and quality: how many features were sent back for rework, how many bugs were found during testing and after launch. All this translates into financial benefits and development speed, which yields the ROI.
AI in client processes. Here I would start with a specific unit of result. For example, how much it costs to close one support ticket. The cost includes the model’s operation, infrastructure, maintenance, and the employee’s time if they have to check or correct the answer. Then we compare it with the current process at a comparable quality. If the ticket became cheaper, but the client has to return, the savings must be recalculated. I would define these conditions before the pilot.
Can you give an example of a typical mistake in evaluating the economics of AI projects that prevents great tech initiatives from reaching production?
A typical mistake is failing to calculate operational costs before release. A team invents a strong AI feature, but right before launch, it turns out to be too expensive.
You need to calculate in advance: how many tokens will be spent, and if it’s an internal model, how many resources are needed for it to handle the load properly.
I would calculate the total cost of a successfully completed task: the model, infrastructure, integrations, human review, and error correction. The pilot should show whether these economics hold up with real requests and at scale.
We see fierce competition in the region: classic banks are building their ecosystems, while big tech and telecoms are aggressively entering fintech. In your opinion, who is currently winning the technology race and adopting AI faster?
I wouldn’t pit banks against tech companies. The winner is the one who controls more of the client’s useful daily scenarios, not just one niche. It is more convenient when a transfer, a movie ticket, and grocery delivery are in one app: consumption habits consolidate around it.
But execution matters. In some apps, the user journey is short and clear; in others, services are piled up so much that even basic banking operations become difficult.
This is where AI comes in handy. A bank with multiple services knows more about the client and can shorten the path to a service: I say what I need using voice, and it selects options—for example, movie or flight tickets for the right dates. All I have to do is choose.
The benefit is especially noticeable when services work together. An entrepreneur needs to pay supplier invoices and figure out if there is enough money until the next revenue arrives. An assistant could gather the invoices, match them with payment deadlines and expected receipts, show a potential cash gap, and prepare payments for confirmation. Today, this requires multiple systems and manual work.
How will the market landscape change in the next 3–5 years? Will AI development cement the monopoly of a few large ecosystems, or will niche AI fintech products have room for explosive growth?
Ecosystems will stay ahead for now, but I see several processes happening at once.
Concentration at the infrastructure level. Large language models require a lot of hardware, and there are problems with that on the market. It is hard to get, expensive, and difficult to maintain.
Diversity at the application level. Launching an MVP has become much easier. Previously, this required large teams, but now a product can be launched solo. It might be an imperfect version, but it allows the idea to be realized. Competition is shifting toward exactly what you build and how you build it.
Such companies won’t be able to compete directly with banking functions and scenarios, but they can provide distinct value. In these cases, banks will look at M&A: buying products or implementing similar ideas internally.
I expect that in the next three to five years, a portion of daily banking will shift to personal agents. The client will instruct them to pay bills, prepare a transfer, or sort out expenses. For banks, this creates a serious question: who will maintain the relationship with the client and influence their choices? I would already be preparing banking services to work through agents—with clear permissions, limits, and action confirmations.
