Conversational Donations:
Designing Trust-Centered Interaction Patterns in LLM-Powered Fundraising

This project presents a conversational assistant that helps people move from donation interests to suitable charity options, with transparent recommendations and visible sources.

Illustrative image generated by ChatGPT

Initial situation

TrueTurn makes digital donation opportunities for around 210 charitable organisations available through the TWINT marketplace. Users can search and browse the catalogue, but choosing an organisation remains difficult when they know what they care about but not whom to support. Conversational AI can make this search easier by asking about a donor's interests and presenting relevant options with clear explanations.

Problem statement / Project goal

This thesis investigated how an LLM-powered donation assistant can support charity selection with transparent recommendations, clear user control, and source-backed explanations. The goal was to identify donor trust concerns, translate them into design implications, and define trust-supporting interaction patterns, linguistic design choices, and interface cues for a functional prototype. The assistant acts as a neutral decision-support tool: it asks focused questions, presents suitable options, explains each match, and shows the sources and limits behind its information.

Solution developed and its benefits

The project combined a literature review with interviews involving donors and fundraisers. The findings shaped a web prototype based on structured NGO data and retrieval-backed sources. The prototype was evaluated with 13 participants.

The solution helps donors move from a broad intention to a better-informed choice:

• Find relevant charities faster based on personal interests
• Compare several suitable organisations side by side
• Understand why each organisation was recommended
• Check sources and see where information came from
• Stay in control of the final donation decision

Participants especially valued the neutral tone, the non-pressuring experience, and visible source cues. Their trust depended on sufficient organisation information, varied options, and a clear distinction between facts and recommendation logic. The main conclusion is that donation assistants work best when they support orientation and verification while leaving the final decision to the user.

Technologies

The prototype was built with TypeScript, React, TanStack Start, LangGraph, PostgreSQL with pgvector, and OpenAI models. Structured NGO data and retrieval-augmented generation connect the assistant's answers to verifiable organisation information.

Key terms

• Conversational Donation Assistant LLM-powered dialogue interface that helps users discover and assess charitable organisations
• Donor Trust How users evaluate the assistant, its recommendations, the organisation information, and the donation platform behind it
• Trust-Centered Interaction Design Interaction patterns, language, and interface cues that make recommendations transparent and keep users in control
• Charity Recommendation Preference-based matching between users and charitable organisations from an available catalogue
• Conversational AI Natural-language interaction for clarifying donation interests and exploring suitable organisations
• Digital Fundraising Use of digital platforms and AI-supported interfaces for charitable giving
• Large Language Models (LLMs) Generative AI models used for dialogue, intent recognition, recommendation generation, and organisation lookup

Customer

TrueTurn AG
Swiss fundraising technology provider
trueturn.ch

Team