Key barriers: poor search and scripted bots
Barrier 1: “Stupid” search and lack of tolerance for typos
A user types “komunalka” (utilities), and the bank search returns “Nothing found” because services are entered in the database under complex bureaucratic names.
Product insight: Full-text search must understand Uzbek and Russian, Latin and Cyrillic alphabets, typos, and everyday names of services—and index not only the catalog but also the transaction history. A user who fails to find what they need three times stops using search altogether.
Barrier 2: Toxic experience communicating with chatbots
Instead of a real AI assistant, users face a “tree of buttons” from which it is impossible to quickly reach a live operator during a non-standard failure.
Product insight: True AI in banking must work predictively. If a client experiences a transaction error, the AI bot should be the first to write a push message with a solution, instead of waiting for a complaint from the client.
Behavior models: searchers and callers
-
Searchers: Ignore icons on the main screen. They immediately tap the search bar, expecting a Google-level experience—and perceive every “Nothing found” as a breakdown of the entire app.
-
Disappointed callers: At the slightest problem, they bypass the chatbot, wasting time calling the call center. Every such call costs the bank money and subtracts a trust point from digital channels.
What it looks like in practice
A user from Namangan wants to pay for electricity. In the search bar, he types “svet”—in Latin script, as he is used to writing in messengers. The search is silent: the service is entered in the catalog under the official name of the regional supplier, in Cyrillic. The second attempt—“свет” (light) in Russian Cyrillic—returns an article from the help section. He eventually finds the service through a month-old transaction history.
Later, the same user encounters a stuck transfer and writes to the support chat. The bot offers a tree of six buttons, none of which are about his case; the word “operator” returns the same menu. Having reached the call center, he resolves the issue in three minutes—and concludes: the in-app chat exists just for show.
The linguistic reality of Uzbekistan—Latin and Cyrillic, Uzbek and Russian mixed together, everyday names instead of official ones—turns “smart search” from a pleasant option into a basic requirement. Until the search understands “svet”, “kommunalka”, and typos, the conversation about generative AI in a superapp does not even begin.
A separate layer is voice input: a significant part of the audience dictates queries in Uzbek, and the recognition, trained predominantly on Russian, regularly misses. For millions of users, this is not exotic, but the main way of communicating with a smartphone.
Why it matters
The more services a superapp gathers, the more expensive every unsuccessful attempt to find something in it costs. Search and assistant are not a “feature from the roadmap,” but the load-bearing structure of navigation. And it will be necessary to start not with a spectacular chatbot, but with the linguistic foundation: until the model understands the local context, any “AI” will remain a tree of buttons.
FAQ
What is most often searched for in superapps?
Provider receipts, transfers by name, and the section with limit settings. Slightly less often—exchange rates, ATM addresses, and “how to disable a paid subscription.”
Why are chatbots so ineffective?
Most of them are not integrated with the core of the banking system: the bot sees the question but does not see the client’s account, so it can neither check the payment nor cancel the transaction.
What should AI in the bank of the future be like?
Invisible. It should analyze cash flow and highlight the risks of a cash gap without additional requests.
Why is search in Uzbek superapps harder than it seems?
Queries come in two languages and two alphabets, mixed with transliteration and everyday names of services. The engine must understand “svet”, “свет”, and the official name of the energy company as the exact same query—without this, a catalog of thousands of services is dead.
Are Uzbekistanis ready to trust their money to an AI assistant?
Not yet—trust starts small. Users calmly accept “quiet” AI: auto-categorization of expenses, hints, pre-filled forms. The attitude towards autonomous actions with money is cautious, and it will have to be cultivated in stages.
Original source of the research: How Uzbekistanis use search and AI in banking ecosystems