In Finland, a single legal consultation can cost up to 250 euros per hour, and public legal aid is restricted by strict eligibility rules. Many people facing housing disputes, employment problems, or family conflicts therefore go without any guidance at all. This research set out to investigate whether an AI-powered chatbot could provide reliable legal information to those people, free of charge and at any time of day.

The outcome of the thesis is a functional prototype of a conversational legal information service, built on Large Language Models (LLMs) and evaluated through empirical research. The prototype is not a replacement for lawyers. It is a tool for preliminary orientation, helping users understand their situation and find appropriate next steps before, or instead of, seeking formal legal help.

How service design shaped the research

The research was structured through the Triple Diamond framework, which extends the classic Double Diamond model with Implementation and Scale phases. This guided the work through six stages: Discover, Define, Develop, Deliver, Implement, and Scale. Throughout each stage, service design tools were used not simply to document findings, but to generate design decisions.

User interviews with fifteen prospective users revealed the real nature of the problem. People did not lack legal information so much as they lacked the vocabulary to find it, and the confidence to trust what they found. This insight, produced through structured interview analysis, directly shaped how the chatbot was designed to communicate: in plain language, without legal jargon, and always with a visible reference to the source legislation.

Personas and user journey maps translated interview findings into concrete design targets. The journey map of one key persona, an immigrant receiving an eviction notice, identified two critical pain points: difficulty articulating the legal problem without knowing the correct terminology, and uncertainty about what to do after receiving information. These findings led to specific design decisions, including a query reformulation feature and a structured escalation pathway to human legal services.

The service blueprint was particularly productive. By mapping the full interaction flow across user actions, interface behaviour, AI processing, and technical infrastructure, the blueprint made visible a key architectural insight that would not have emerged from technical development alone: citation generation had to happen in the background, but the resulting legislative reference had to appear prominently to the user. This single finding shaped the entire Retrieval-Augmented Generation (RAG) architecture, in which every answer is grounded in and linked to the specific Finnish law it draws from.

What the prototype delivers

The working prototype (at lagen.fi) demonstrates that an LLM-assisted service can provide accurate legal information in Finnish consumer, housing, and employment law domains. The citation-first architecture substantially reduces the risk of incorrect AI-generated answers. Privacy is protected through data minimisation design: no persistent user profiles, optional account creation, and user-controlled conversation history with immediate deletion.

Six rounds of usability testing confirmed that users respond positively when the service communicates its limitations clearly and provides transparent references. Expert consultations with practising lawyers verified the accuracy of prototype responses and shaped the boundaries between general information and legal advice, a boundary that Finnish and EU regulation makes legally significant.

Privacy as a design constraint

Privacy was not an afterthought. Legal matters inherently expose sensitive personal information, including financial difficulties, family conflict, and housing insecurity. Early concept versions followed common commercial AI patterns and included persistent chat history and user profiles. Critical examination showed this to be inappropriate: unlike commercial services where users trade privacy for convenience, a legal aid service must protect users who may already be in vulnerable positions.

The final design therefore avoids aggressive data collection and gives users explicit control over their conversation history, including immediate deletion. This meant accepting reduced personalisation in exchange for privacy protection appropriate to the user group.

What comes next

The prototype remains at proof-of-concept stage and is not ready for public deployment. It does not yet support multiple languages, integrate case law, or handle complex query disambiguation. However, the research provides design guidelines applicable to any organisation considering LLM-based legal information tools, including courts, legal aid providers, non-governmental organisations, and legal technology companies. The central finding is that responsible AI-assisted legal information is achievable, provided that accuracy, privacy, and scope limitations are treated as non-negotiable design constraints from the outset.

Sources

Paasonen, J. (2026). Feasibility of a LLM assisted Legal Aid chat. Master’s Thesis. Turku University of Applied Sciences.
Image: Paasonen, J. (2025). Stable Diffusion, Version XL