AI Assistant – Designing a Conversational Service for a Finnish Ministry

Overview

A Finnish ministry serves people around the world who need information about its services, often in situations where the right answer depends on their personal circumstances.

The ministry wanted a virtual assistant that could answer these questions reliably in Finnish, Swedish and English. I took over as project lead mid-project and worked as its conversational AI designer, steering the design work, the team and the collaboration with the client.

Challenge

The information is detailed, legally sensitive and changes over time. A wrong or overly confident answer can have real consequences for the person asking.

The challenge was to design an assistant that would:

• Give accurate answers across a broad range of topics.

• Recognise when a question depends on the user's situation, and ask about it.

• Be maintainable by the client after launch.

At the same time, many stakeholders had to agree on what the assistant should do, and just as importantly, what it should not do.

Approach

I designed the assistant's content structure so that each topic area could be handled and expanded on its own, and facilitated workshops to build a shared understanding of the service, its tone and its boundaries.

Before launch, I planned and ran pilot testing with people who know the questions best, and used what we learned to improve the assistant where it did not yet serve its users.

To make the service sustainable, I worked with the client to agree how the service is maintained and developed after launch.

Outcome

• The assistant went into production as a public service.

• The client can run and develop the service independently.

• Improvements are based on real use rather than assumptions.

  
  

Impact & Reflection

This project showed me that designing an AI service is as much about people and responsibilities as it is about the technology. The assistant only works if someone owns its content, notices when it goes wrong and knows how to fix it. It also deepened my conviction that an AI service should be honest about its limits, and that knowing when not to answer is part of good design.