The Challenge Is Finding the Right Information, not More Information
Linnunmaa Lex helps organisations manage statutory requirements. The service is built around four key tasks: organisations must be able to identify the legislation that applies to them, monitor changes in that legislation, assess their own compliance and demonstrate compliance when required by authorities, customers or other stakeholders.
The service is typically used not by legal professionals, but by business leaders, managers, specialists and supervisors. Users range from occupational safety and environmental professionals to HR specialists and ICT and information security experts. For many of them, the legal text itself is not the primary concern. What matters is understanding what specific requirements mean in practice and how they affect their own operations.
The challenge is that there is a great deal of information. A single organisation may be subject to hundreds of laws containing thousands of individual sections and requirements. Although Linnunmaa Lex already narrows down the legislation visible to a customer during implementation based on the nature of the organisation’s operations, a substantial amount of content remains.
Traditional keyword searches do not always solve the problem. A broad search term may return dozens or hundreds of results, leaving users to determine which law or legal provision is most relevant to their question. In addition, the everyday language used by users does not necessarily align with the terminology used in legislation.
This created a need for a new kind of search experience.
A Search Assistant, not a Legal Assistant
The solution was to build an AI-powered search assistant. Users can ask questions in natural language through a familiar chat interface. The assistant identifies the intent behind the question, searches Linnunmaa Lex content for the relevant information and generates an answer together with its sources.
The goal is not to replace experts or provide legal advice. This distinction was intentionally built into the project: it is a search assistant, not a legal assistant.
The assistant helps users find the right information, summarises the most relevant content and provides links to the original sources. When a response includes legislative content, it is presented in its original form, and users are encouraged to verify the information directly from the legislation itself. The solution also distinguishes between legally binding content and supplementary material produced by Linnunmaa Lex experts, such as comments, summaries and compliance-related guidance.
If a user’s question is unclear, the assistant can ask for clarification. If the question falls outside its scope or expertise, the user is directed to Linnunmaa Lex specialists.
This keeps the role of AI tightly focused on the original problem the project set out to solve: finding accurate and reliable information as quickly as possible.
High-Quality, Up-to-Date Information Is Essential Behind the Scenes
The conversational interface is the most visible part of the solution, but considerable work is required behind the scenes to ensure the assistant can find the correct information.
A general-purpose language model does not automatically know the content produced by Linnunmaa Lex experts, nor does it necessarily reflect the latest legislative changes. For this reason, information relevant to the user’s question is first retrieved from Linnunmaa Lex’s own service before the selected content is provided to the language model for response generation.
Legislative content is prepared for search by breaking it into smaller components, enriching it with additional metadata, structuring it and removing inconsistencies, helping ensure that the correct content can be found as reliably as possible. When a user submits a question, the system also takes the organisation, subscription and access rights into account before retrieving suitable sources.
Data quality plays a central role in this process. It is not beneficial to provide the language model with as much information as possible. Instead, the objective is to provide the smallest possible set of information that is relevant to the specific question being asked. This improves answer quality, speeds up response times and reduces operating costs.
Information must also remain up to date. Whenever legislative content is updated within the Linnunmaa Lex service, those updates are also made available to the search assistant, ensuring that the information behind its responses remains as current as possible.
Legal Content Places Higher Demands on AI
The project revealed that the complexity of legal content also places demands on the language model itself. Lighter models did not perform as well because generating a response may require synthesising information from multiple different sources.
At the same time, AI introduces a new challenge for testing. In traditional software, the same input generally produces the same outcome. Language models do not behave in the same deterministic way. Responses may vary, and quality cannot be evaluated solely by checking whether a technical error occurs.
To address this, the project team developed a lightweight quality evaluation tool based on predefined example questions, expected answers and the source material that search operations should retrieve. This makes it possible to evaluate both search performance and response quality, while also monitoring how changes to the application affect the overall experience.
Automated evaluation does not replace real users, however. Ultimately, the most important question is whether the service creates genuine value for users, and that can only be assessed reliably through real-world usage and user feedback.
A Clearly Defined First Version Accelerates Learning
One of the key lessons learned during the project was the importance of carefully defining the scope of the first version. During development, numerous ideas emerged for additional features and future enhancements, but these were intentionally deferred.
The objective is to bring a focused solution into the hands of real users as quickly as possible and determine whether it genuinely solves the identified problem. A narrower use case is also easier to monitor and evaluate than a solution attempting to address multiple challenges simultaneously.
Linnunmaa Lex began with a lightweight proof of concept that provided the foundation for further refinement towards production use. The next important step is to release the solution to customers, collect feedback and continue development based on real user experiences.
There are several possible future directions. The same technology could be applied to more routine customer service questions, such as login issues, password-related support or assistance in finding content within the service. The assistant could also identify situations where customers might benefit from content not included in their current subscription and guide the conversation accordingly.
Better Customer Experience, More Efficient Service
From Linnunmaa Lex’s perspective, the objective of the AI search assistant is ultimately very practical: customers should be able to find the information they need more quickly and resolve everyday questions without unnecessary searching.
The benefits are also visible on the service provider’s side. When specialists no longer spend time guiding users through information discovery or answering simple navigation questions, their expertise can be focused on situations where human expertise genuinely adds value.
The project also demonstrates how important it is to define the role of AI appropriately. In a legal context, AI does not need to, and should not, know everything or make interpretations on behalf of users. Value can be created simply by understanding a user’s question, finding the correct content within a large body of information and helping users reach reliable information more efficiently.
