Powering tax donations with AI powered personalized recommen…

日本的“故乡税”捐赠计划，即 Furusato Nozei ，允许纳税人通过捐款支持他们关心的地方自治体。随着人口向东京等大城市集中，农村城镇的税基不断缩小，该计划的出发点是让纳税人将部分税款定向用于他们想支持的社区。实际上，它通过抵税机制运作：在按收入设定的上限内，捐款的大部分会在次年的所得税和居民税中抵扣。作为回报，自治体会寄出“返礼品”，通常是当地特产，既让捐赠者能享受地域物产，又能支持地方发展。

不过，面对数量庞大的自治体和海量的返礼目录，许多捐赠者觉得难以抉择。为简化流程、方便按地区或主题比较选项，出现了专门的平台。由 TRUSTBANK 运营的 Furusato Choice （日文名 ふるさとチョイス ）是日本最大的此类平台之一，约有 760,000 件返礼品上架。其直观的界面帮助自治体与捐赠者建立联系，也为大量首次参与的用户提供了支持。

为了进一步改善用户体验， TRUSTBANK 将目光投向人工智能，认为 AI 能在面对过多选择时帮助用户决策。公司使用 OpenAI API 开发了 Choice AI ，帮助用户找到符合偏好的返礼品。

与 Recursive 合作构建 Choice AI

在 Furusato Nozei 推出 15 年多之后，仍有不少纳税人难以充分利用这一制度。 Furusato Choice 业务总部产品总经理 Yuki Tateyama 指出，“很多人觉得故乡税制度复杂或令人生畏。为此，我们在 Furusato Choice 应用中加入了 AI 功能，让流程更好上手。” TRUSTBANK 的第一步是引入 AI 驱动的搜索，帮助用户检索返礼品。

Tateyama 还说，“在 Furusato Nozei 里，人们不是像在一般电商平台上买即时需要的商品，而是考虑如何把捐款额度用到最合适的地方。返礼品目录的规模不亚于大型在线商城，找到合适的选项一直很有挑战性。”

由于缺乏内部 AI 专长， TRUSTBANK 借助外部力量，与现为官方 OpenAI 服务合作伙伴的 Recursive 建立合作。产品部平台推广负责人 Issei Hirano 解释说，他们选择 Recursive 是基于对方在 AI 方面的先进能力和国际经验。

他描述了双方的协作方式：“从规划阶段起， Recursive 就提供了技术支持，设计并实现了对话式 AI 代理，还搭建了 RAG 系统。我们负责准备返礼品数据库、定义功能需求，并将这些功能集成到我们的应用中。”这种协作促成了 Choice AI 的顺利上线，使用户能在 Furusato Choice 应用内通过交互式对话发现推荐商品。

以多智能体架构推动个性化

Choice AI 的核心是多智能体架构。一个路由模型会分析用户输入以识别意图，并把任务分派给相应的智能体。在路由层之下，有专门的智能体如 Search Agent 、 Recommendation Agent 和 Greeting Agent 等，各智能体还可调用子智能体和工具，实现流畅的编排与基于意图的精确结果。

个性化也体现在提示词设计上。带领智能体开发的 Recursive 软件工程师 Matthew Whalley 说明：“我们根据用户的具体信息动态组合智能体。例如既有老用户走一条交互路径，也有首次用户走另一条。我们动态生成提示词来管理这些路径。”

目前 Choice AI 运行在 GPT‑4.1 系列上。Whalley 说，“默认使用 GPT‑4.1-mini ，但在测试中我们会根据延迟和准确性在 nano 版本或更大模型之间动态切换。”

对真实用户行为的分析也带来了新见解：许多用户像使用搜索引擎一样与应用互动，向大模型提供详尽的商品信息并期待即时推荐；另外，设计用于启动对话的简短内置提示词被频繁使用。基于这些发现，团队对 Choice AI 进行了多项改进，例如将推荐流程提前呈现，并扩大推荐商品的多样性，以增加用户接触到不同选项的机会。

通过随机化搜索结果扩展选择范围

Choice AI 针对 Furusato Nozei 的两大痛点给出了解决方案：

• 通过交互式对话提供个性化推荐，减轻海量返礼品带来的用户困惑；
• 通过定制化建议帮助用户发现更广泛的地区和返礼品，避免过度集中在少数自治体或热门商品上。

借助多智能体架构，即便用户缺乏搜索技巧或不了解具体商品，只需自然对话，甚至模糊的需求如“给父母准备的礼物”，也能得到合适推荐。

为避免集中化偏向， Choice AI 在搜索结果中加入了可控的随机性。Whalley 解释说，“我们引入随机性，并根据捐赠数据在都道府县间调整推荐，以支持公平和地区多样性，除非用户明确表明偏好。”这有助于让较小的自治体和小众商品被更多用户发现，营造更丰富的体验。

使用 Choice AI 的用户转化率高于仅依赖站内普通搜索的用户。Hirano 解释原因是，AI 能挖掘出用户自己难以言明的模糊需求，比如偏好和预算，甚至能直接推荐具体返礼品。

走向为 Furusato Nozei 提供礼宾式服务

目前， Furusato Choice 主要利用 AI 优化返礼品检索环节，帮助用户更快找到合适选项。展望未来，公司计划把 AI 扩展到更多场景，提升服务整体价值。

Tateyama 设想， Furusato Choice 不仅要成为一个经济上有益的平台，更要通过真诚的连接把用户和地方自治体联结起来。为实现这一愿景，团队将持续打磨 AI 推荐质量，个性化用户的全流程体验，最终打造贴近个体需求的礼宾式服务。

Japan’s hometown tax donation program, known as Furusato Nozei, allows taxpayers to support municipalities they care about by making a donation. As people move to large cities like Tokyo, local tax bases in rural towns shrink, so the program was designed to let taxpayers redirect a portion of their taxes to the communities they want to support. In practice, it works through a tax credit system: up to an income based cap, most of the donated amount is credited against the donor’s income and resident taxes for the following year. In return, municipalities send donors “thank-you gifts,” typically local specialty products, so donors can enjoy regional offerings while contributing to local communities.

However, many donors find the program difficult to navigate, given the sheer number of municipalities and the enormous catalog of thank-you gifts. To simplify the process and help donors compare options by region or theme, dedicated platforms have emerged. Furusato Choice, operated by TRUSTBANK, is one of Japan’s largest Furusato Nozei platforms, with roughly 760,000 thank-you gifts listed. Its intuitive interface has helped municipalities connect with donors, and it has supported many platform users, especially those participating for the first time. 

To further improve the experience, TRUSTBANK looked to AI, recognizing its potential to help users make decisions when the range of options feels overwhelming. Using the OpenAI API, the company developed the Choice AI feature, which helps users discover thank-you gifts that match their preferences.

More than 15 years after Furusato Nozei was introduced, many taxpayers still struggle to make the most of it. Yuki Tateyama, Product General Manager at Choice Business HQ, explains, “Many people find the hometown tax donation system complicated or intimidating. To address this, we added AI powered features to our Furusato Choice app to make the process easier to navigate.” TRUSTBANK’s first step was to introduce AI powered search to help users explore thank-you gifts.

Tateyama adds, “In the Furusato Nozei system, people are not simply shopping for products they need immediately, as they might on typical ecommerce platforms. Instead, they focus on how to make the best use of their donation limit. With a catalog of thank you gifts on par with major online marketplaces, finding the right option has always been challenging.”

Providing personalized recommendations for thank-you gifts, based on each user’s information and intent, is an area where AI can be especially effective. However, TRUSTBANK did not have internal developers who specialized in AI, so external support was essential. To address this, TRUSTBANK partnered with Recursive, now an official OpenAI services partner.

Issei Hirano, Head of Platform Promotion in the Product Division, explains the decision, “We selected Recursive as our partner because of their advanced AI expertise and their global experience.”

He describes how the teams worked together. “Recursive provided technical support from the planning stage, designed and implemented the conversational AI agent, and built the RAG system. We prepared the thank-you gift database, defined functional requirements, and integrated these features into our app.” This collaboration enabled a smooth build and rollout of Choice AI, allowing users to discover recommended items through interactive conversations within the Furusato Choice app.

The core of Choice AI is its multiagent architecture. A routing model analyzes user input to determine intent and delegates tasks to the appropriate agents. Beneath this routing layer, specialized agents such as the Search Agent, Recommendation Agent, and Greeting Agent operate. Each agent can invoke additional subagents and tools, enabling smooth orchestration and accurate, intent based results.

Personalization is also built into the prompting. Matthew Whalley, a software engineer at Recursive who led agent development, explains, “We combine agents dynamically based on user specific information. For instance, existing users follow one interaction path, while first time users follow another. We dynamically generate prompts to manage these interaction paths.”

Choice AI currently runs on the GPT‑4.1 series. Whalley explains, “By default, we use GPT‑4.1 mini, but we are experimenting with dynamically switching to either the nano version or larger models based on latency and accuracy during testing.”

Whalley also notes that analyzing real user behavior led to new insights: “Our analysis revealed that many users interact with the app similarly to how they would use a search engine. They provide extensive product information to the LLM and expect immediate recommendations. We also found that short, built in prompts designed to start conversations were used frequently.” Based on these findings, the team incorporated a range of improvements into Choice AI. For example, the recommendation flow was adjusted to surface suggestions earlier, and the variety of recommended products was expanded to increase exposure to more diverse options.

Choice AI addresses two challenges in the Furusato Nozei experience:

1. Reducing user confusion caused by the vast number of thank-you gifts, by offering personalized recommendations through interactive conversations.
2. Avoiding concentration on a small set of municipalities or popular items, by helping users discover a wider range of regions and thank-you gifts through tailored suggestions.

With the multiagent architecture inside Choice AI, users can find relevant thank-you gifts even without search skills or detailed product knowledge. Natural conversations, or even vague requests such as “a gift for my parents,” can be enough to surface suitable recommendations.

Choice AI reduces bias toward specific municipalities or items by adding controlled randomness to search results. Whalley explains, “We introduce randomness and vary recommendations across prefectures based on donation data to support fairness and regional diversity, unless users clearly indicate their preferences.” This helps users discover smaller municipalities and niche products, creating a more diverse and engaging experience.

As a result, users who used Choice AI saw higher conversion rates than those who relied on standard on-site search. Hirano explains the reason, saying, “Because the AI could draw out vague needs, like preferences and budget, that users themselves often struggled to put into words, and could even recommend specific thank-you gifts.”

Today, Furusato Choice primarily uses AI to improve the experience of searching for thank-you gifts, helping users quickly find options that fit their needs. Looking ahead, the company plans to expand AI into additional areas and further enhance the overall value of the service.

Tateyama envisions Furusato Choice becoming a platform that connects users and municipalities through genuine goodwill, rather than focusing only on economic benefit. To support that vision, the company aims to refine the quality of AI driven recommendations and personalize the end to end user experience, ultimately creating a concierge style service that is closely attuned to each individual user.

Generated by RSStT. The copyright belongs to the original author.

Source