TRUSTBANK uses AI agents to personalize Furusato Nozei gifts

Furusato Nozei, Japan's hometown tax donation scheme, has always had a weird problem: it works too well. Millions of residents donate to municipalities outside their home city to lower their tax bill, and in return they get a gift, anything from wagyu beef to rice cookers to hotel vouchers. The catch is that the catalog has ballooned to tens of thousands of options across thousands of towns, and scrolling through it feels less like philanthropy and more like doomscrolling an e-commerce site with no filters that make sense.

TRUSTBANK, which runs one of the major Furusato Nozei portals, decided the fix wasn't a better search bar but a better conversation. Working with Recursive, a Tokyo-based AI shop, the company built Choice AI on top of OpenAI's models. Instead of typing keywords into boxes, users chat with the system about what they actually want, budget, dietary preferences, whether they're feeding a family of four or just themselves, and it narrows thousands of listings down to a handful that fit.

What makes this notable isn't the chatbot format itself, which by 2024 standards is almost boring. It's the domain. Furusato Nozei sits at this odd intersection of tax policy, regional economic development, and consumer shopping, and the gift-matching problem has real stakes for small municipalities that depend on donation revenue to fund local services. A donor who gives up in frustration and picks nothing, or picks the wrong thing and doesn't donate again next year, is a measurable loss for some town in Hokkaido or Kyushu that was counting on that money.

TRUSTBANK's pitch is that personalization at this scale was basically impossible with static category filters. You can tag a gift as seafood or dairy, but you can't easily tag it as good for someone who wants to try Kyushu without leaving Tokyo, which is the kind of half-formed request an AI agent can actually parse. Recursive built the system to handle that ambiguity directly, treating the shopping decision as something closer to asking a well-informed local rather than filling out a form.

Whether this meaningfully increases donations, or just makes the existing donations feel smoother, remains an open question OpenAI's blog post doesn't really answer. But it's a clean example of AI agents getting deployed into a very specific, very local piece of civic infrastructure rather than another generic customer service bot, and that's worth paying attention to as these tools spread past the usual Silicon Valley use cases.