TRUSTBANK uses AI agents to personalize Furusato Nozei gifts

Key Actionable Insights

1

Design multi-agent systems with a routing layer that delegates to specialized agents based on user intent. TRUSTBANK's architecture uses a routing model for intent detection, then delegates to Search, Recommendation, or Greeting agents, each capable of invoking subagents and tools. This separation of concerns enables each agent to be optimized independently.

This pattern is especially effective when user queries span multiple distinct intent categories (search vs. recommendation vs. general conversation) that require different processing pipelines and data sources.

2

Use different AI model sizes strategically within the same system to optimize for cost, latency, and accuracy. TRUSTBANK uses GPT-4.1 mini as the default but experiments with dynamically switching to nano for simple tasks and the full model for complex reranking. This tiered approach allows you to allocate computational resources where they matter most.

Apply this when building production AI systems where latency and cost are constraints. Simple tasks like greeting detection don't need the largest model, while final recommendation quality benefits from more capable models.

3

Introduce controlled randomness in recommendation results to promote fairness and diversity, rather than always optimizing for popularity or relevance alone. TRUSTBANK varies recommendations across regions based on donation data to prevent concentration on a small set of popular items, helping users discover smaller municipalities and niche products.

This approach is valuable for any marketplace or recommendation platform where you want to balance user satisfaction with fair distribution across suppliers, regions, or lesser-known options.

4

Analyze real user interaction patterns after deployment to refine your AI system. TRUSTBANK discovered that users often treated their conversational AI like a search engine, providing extensive product information and expecting immediate results. This led them to adjust the recommendation flow to surface suggestions earlier.

Users rarely interact with AI systems the way designers expect. Post-launch behavioral analysis is essential to understanding actual usage patterns and iterating on the conversational flow accordingly.

5

Dynamically generate prompts based on user-specific context rather than using static prompts for all users. Choice AI combines agents differently for existing users versus first-time users, creating personalized interaction paths that adapt to the user's history and experience level with the platform.

This is particularly effective for platforms with diverse user segments where a one-size-fits-all conversational experience would fail to address the different needs and knowledge levels of various user types.

6

When your organization lacks internal AI expertise, partner with a specialized AI firm while maintaining clear division of responsibilities—domain expertise and data preparation internally, AI architecture and implementation externally. TRUSTBANK handled the gift database and functional requirements while Recursive built the conversational AI and RAG system.

This collaboration model works well when the domain knowledge (product catalogs, user behavior, business rules) is complex and company-specific, while the AI engineering can be externally sourced from specialists.