Choice AI: как OpenAI помогает японцам платить налоги и получать подарки

Choice is always stressful, especially when it directly affects your money and tax obligations. In Japan, there is a unique and extremely popular system called Furusato Nozei. It allows citizens to direct part of their taxes to the budgets of rural prefectures, and in return receive local delicacies, artisan goods, or vacation certificates. The idea is wonderful, but in practice it has turned into a logistical nightmare for users. When you open a catalog of tens of thousands of items — from Kobe marble beef to sets of organic vegetables from Hokkaido — your brain simply refuses to choose. TRUSTBANK decided it was time to stop torturing clients with endless filters, and turned to OpenAI technologies.

Together with startup Recursive, the bank launched Choice AI — an intelligent system that acts as a personal concierge. This is not just a chatbot in the traditional sense that delivers pre-prepared answers. The developers built a full-fledged multi-agent system. Inside it work several specialized AI agents, each responsible for its own part of the task. One agent analyzes the user's profile and past preferences, another filters the huge gift database in real-time, and a third ensures that recommendations comply with the specific person's tax limits. As a result, instead of exhausting search, the user simply has a relaxed conversation, gradually clarifying their wishes.

Why is this important right now? We are observing a fundamental shift in how business interacts with customers. Previously, companies competed on who had more convenient filters and faster search. Today, that is no longer enough. Choice AI demonstrates a transition to an "intention interface" model. Users no longer need to know how the database is structured or what product categories exist. They simply need to express an intention — for example, "I want something unusual for dinner for a family of four, to support farmers affected by the rain." OpenAI's models have enough contextual understanding to connect this abstract request to specific items in the system.

Recursive did not choose multi-agent architecture by accident. A single LLM is often prone to hallucinations or oversimplifications when working with large data arrays. Division of responsibilities between agents allows for much higher accuracy. While one agent is "being creative," suggesting options, another acts as a strict censor, checking product availability in stock and price validity. This makes the system suitable for use in the banking sector, where any numerical error can lead to reputation losses.

For TRUSTBANK, this project became a way to stand out against competitors who still offer users traditional web storefronts. In a market saturated with options, the winner is not the one with more products, but the one who faster and more accurately helps the client make a decision. We are seeing how artificial intelligence gradually becomes that very "smart layer" between humans and complex bureaucratic or financial processes. The Japanese experience with taxes is only the first sign. Soon such agents will help us choose insurance, plan investments, or select educational courses, eliminating the need to spend hours studying fine print in descriptions.

The main point: The era of independent catalog browsing is coming to an end. Choice AI proves that multi-agent systems can turn even tedious tax payment into a personalized and pleasant experience. Will AI assistant become a mandatory standard for any fintech application in the next two years?