TL;DR: TRUSTBANK deployed an AI agent powered by OpenAI's GPT-4.1 series to transform how users discover donation destinations on Furusato Choice, one of Japan's largest hometown tax donation platforms. By solving choice overload across over 760,000 return gifts, their multi-agent system delivers personalized recommendations that significantly outperform traditional approaches, driving sustainable regional development.

Here's a uniquely Japanese solution to rural decline: Furusato Nozei, the "hometown tax donation" program. The problem? As workers flock to Tokyo, rural tax bases collapse. The fix? Let taxpayers redirect their taxes to any municipality they choose.

The mechanics are elegant. Donate to a town. Get a tax credit covering most of it (up to income-based caps). Receive "thank-you gifts"—premium local products from that region. Everyone wins: donors get Hokkaido crab or Kyoto sake, small towns get revenue.

But here's the catch: choice overload. How do you pick from thousands of municipalities and hundreds of thousands of gifts?

Enter platforms like Furusato Choice. Operated by TRUSTBANK, it's one of Japan's largest, featuring over 760,000 thank-you gifts from over 1,700 municipalities. It simplified discovery for millions of first-time participants. Yet even with an intuitive interface, the sheer scale created new challenges.

Picture this: You open an app with over 760,000 gift options—premium beef from Hokkaido, citrus from Ehime, accommodation vouchers from Kyoto. Where do you even start?

Welcome to the reality facing Furusato Choice. Despite partnering with over 1,700 municipalities (approximately 95% nationwide), the platform faced a critical challenge: concentration. Donations kept flowing to the same well-known regions and popular gifts. Lesser-known municipalities? Invisible.

The math was brutal. With hundreds of thousands of options, users faced paralysis. Small towns struggled for visibility. Donors missed discovering regions that truly matched their interests. Something had to change.

Enter Choice AI. TRUSTBANK partnered with Recursive—an AI startup whose co-founder is a former senior research engineer at Google DeepMind—to build an in-app chatbot that actually understands what donors want. Recursive leveraged its technical collaboration with OpenAI to develop the solution.

The genius part? Choice AI uses a multi-agent architecture featuring specialized agents that work together like a recommendation dream team. A routing model coordinates three key players: Search Agent (finds candidates), Recommendation Agent (ranks options), and Greeting Agent (handles conversation flow). Sub-agents handle specialized tasks.

But here's the thing: this isn't just smart matching. When you type something vague like "a gift for my parents," the AI doesn't panic. It asks clarifying questions, progressively refines understanding, and surfaces options you'd never have found scrolling.

The secret? Controlled randomness. Choice AI deliberately varies recommendations across prefectures using donation data, ensuring small municipalities get their moment alongside the big names.

Recursive leveraged its technical collaboration with OpenAI, announced in July 2025, gaining early access to cutting-edge models.

The Stack:

• GPT-4.1 series: Defaults to GPT-4.1 mini for speed and efficiency, with dynamic switching to GPT-4.1 nano (faster) or larger models (more accurate) based on real-time latency and accuracy requirements.
• Multi-agent architecture: Routing model orchestrates specialized agents—Search, Recommendation, Greeting—plus sub-agents for complex tasks conventional systems can't handle.
• Smart pipeline: Ranks and reranks gift candidates using user history and behavioral patterns.
• Security-first: LLM-based data masking, encrypted channels, strict access controls, vulnerability management, and content moderation protect sensitive information.

These choices enable personalized, secure experiences at massive scale.

The results: Before launch, TRUSTBANK and Recursive tested Choice AI using LLM-based simulated users that mimic real donor behavior patterns—from crystal-clear requests to completely vague queries.

Here's what changed:

• 78% accuracy rate in recommendation relevance—crushing the existing system that relied on popular weekly/monthly items
• Higher conversion rates for Choice AI users versus standard on-site search
• Over 1,700 municipalities (approximately 95% nationwide) now benefit from balanced donor engagement
• Over 760,000 return gifts made discoverable through natural conversation
• Fair distribution achieved through controlled randomness across prefectures

The broader impact? Donors discover hidden gems. Small municipalities gain visibility. The platform drives app downloads and user satisfaction while strengthening financial support for sustainable regional development. Economically useful connections that weren't happening before? Now they are.

• Multi-agent beats monolith for complex personalization. When you're matching diverse users to hundreds of thousands of options, specialized agents collaborating (Search + Recommendation + Greeting) dramatically outperform single-model approaches. Division of labor works.
• Simulate real behavior before launch. LLM-based personas representing actual user patterns—clear requests, vague queries, everything between—let you validate AI performance across difficulty levels before real users arrive.
• Fairness requires intentional design. Controlled randomness and prefecture-based variation prevent winner-takes-all dynamics, ensuring lesser-known regions compete alongside popular choices.

TRUSTBANK envisions expanding AI beyond gift selection to deepen connections between donors and communities. The roadmap includes helping users discover regional characteristics and local challenges through the recommendation experience. The collaboration may extend beyond Furusato Nozei to support broader municipal services and regional revitalization efforts.

🐯 Tiger Takeaway

TRUSTBANK proved that solving choice overload at scale requires architectural sophistication, not just better search bars. By deploying multi-agent systems powered by GPT-4.1 mini with dynamic model switching, they're transforming how millions discover regional Japan. The lesson? When you're operating at massive scale—760,000+ options, 1,700+ partners—traditional recommendation engines hit their limits fast. Multi-agent AI doesn't just suggest. It guides discovery, ensures fairness, strengthens communities, and enables sustainable growth.

🤖

AI-Powered Newsletter