TRUSTBANK Choice AI: Personalizing Furusato Nozei with Multi-Agent Architecture
TRUSTBANK has launched Choice AI, an AI-powered assistant integrated into the Furusato Choice app, to simplify the process of selecting thank-you gifts within Japan's Furusato Nozei (hometown tax donation) program. By leveraging a multi-agent architecture powered by the OpenAI API, the platform helps users navigate a catalog of approximately 760,000 items to find personalized recommendations based on vague intent or specific preferences.
Multi-Agent Architecture for Intent Routing
Choice AI utilizes a multi-agent system to orchestrate complex user requests and provide accurate, intent-based results. The system is structured with a routing model at the top layer that analyzes user input to determine the user's intent and delegates the task to a specialized agent.
Specialized Agent Roles
Beneath the routing layer, the system employs several dedicated agents:
- Search Agent: Handles specific product queries.
- Recommendation Agent: Provides suggestions based on user preferences.
- Greeting Agent: Manages initial user interactions.
These agents can invoke further subagents and tools to refine results. To enhance personalization, the system dynamically generates prompts based on the user's status; for example, first-time users follow a different interaction path than returning users.
Technical Implementation and Model Selection
TRUSTBANK partnered with Recursive, an OpenAI services partner, to design and implement the conversational AI agent and the underlying Retrieval-Augmented Generation (RAG) system.
Model Configuration
Choice AI currently operates on the GPT-4.1 series. The technical implementation details include:
- Primary Model: GPT-4.1 mini is used by default.
- Dynamic Switching: The team is experimenting with switching between the nano version and larger models depending on the required balance of latency and accuracy during testing.
Impact on User Experience and Regional Diversity
Choice AI improves the Furusato Nozei experience by reducing the cognitive load on users and preventing the over-concentration of donations toward a few popular municipalities.
Conversion and Discovery
Users utilizing Choice AI have demonstrated higher conversion rates compared to those using standard on-site search. This is attributed to the AI's ability to extract "vague needs"—such as budget and preferences—that users often struggle to articulate in traditional search bars.
Promoting Fairness and Diversity
To ensure regional diversity and avoid bias toward a small set of popular items, the system introduces controlled randomness into search results. Unless a user specifies a clear preference, the AI varies recommendations across different prefectures based on donation data, helping users discover smaller municipalities and niche products.
Future Roadmap
TRUSTBANK intends to evolve Choice AI from a search-improvement tool into a comprehensive concierge-style service. The goal is to move beyond economic benefit and create a platform that connects users and municipalities through genuine goodwill by further personalizing the end-to-end user experience.