Lessons learned from Svea: Leadership, competence, and technical adaptations

"The technology itself is not the biggest obstacle. What is required is large-scale change management. We must have leadership that encourages the use of AI. We need specific, broad-based training. Investments, resources, and substantial efforts are needed for us to get started and learn to use AI in the Swedish public sector."

However, behind the scenes, the technology and its adaptation are central nonetheless. There is a wide range among public sector users regarding needs, competence levels, and contexts where the tools can be helpful.

"For example, a great many people in the public sector work in places other than at a desk, with the mobile phone as their primary tool. That is an aspect to take into account. At the same time, there are very high demands that information must be correct and that confidential data must be handled in ways that comply with Swedish legislation," says Jonatan Permert.

He also expresses a certain concern about parallel development currently happening in the public sector, where many organizations are running their own development projects for, among other things, chatbots and transcription solutions.

"The fact that so many public organizations are investing in self-developed tools clearly indicates that there is a demand for Swedish solutions. But in our dialogues with municipalities, regions, and government agencies, we see that the same needs recur over and over again. It is not sustainable for 600 public organizations to build their own, customized AI tools," says Jonatan Permert.

Presentations from the Svea Seminar (in Swedish)

Background, vision, and successes

Jonatan Permert on the background and vision for the project. He describes the challenges of introducing generative AI in public organizations and presents solutions that the initiative has developed, including a prototype for a digital assistant that can relieve employees and increase the quality of public services. The presentation provides an overview of the initiative's goals, successes, and challenges, and demonstrates the potential for a common digital assistant in the public sector.

Change management in practice

Shirin Henare highlights the need for change management to realize the value of AI and how the Svea initiative works with change management and training to strengthen the public sector's ability to use AI in a confident, courageous, and learning-oriented manner. The focus is on building engagement, competence, and trust in a time of rapid technological change.

Design as a driving force

How can design make complex systems simple, create security, and build trust? In this presentation, Maja Haak, UX and product designer, shares insights and examples from Svea – where clear structure, plain language, and transparency are central. And how design can be a powerful driving force in the public sector, and how AI can be adapted to the user – not vice versa.

Svea's technical architecture

How can one build a powerful AI assistant that simultaneously meets the public sector's high demands for security and sovereignty? Tobias Wilhelmsson provides a broad technical overview of Svea, the AI platform built from the ground up for the needs of the Swedish public sector. The presentation covers the central parts of the system, from the secure and scalable foundational architecture on Swedish soil to the privacy-focused design choices. It offers an insight into how the platform's intelligent AI engine is built to deliver relevant answers and how the system is designed to be flexible and adaptable for the future.

RAG technology in Svea

Organizations' knowledge is often scattered across countless documents, difficult to access for those who need it. Lea Cornelio describes how Svea's RAG pipeline (Retrieval Augmented Generation) solves this. By converting unstructured documents into searchable knowledge, the assistant can retrieve relevant facts. The generated answers are then grounded in the organization's own sources instead of solely in the model's general training data.

Training data for AI

In today's rapidly changing world, we face a fundamental challenge for artificial intelligence: static AI models and datasets limit AI models' ability to evolve as the world around them evolves (i.e., the data they work with). Adam Ek, AI lead in the Svea project, describes an innovative solution to this problem, where language models are used to dynamically generate and evaluate new data. This approach not only makes it possible to specialize AI for specific organizations and use cases but also significantly improves the accuracy of information retrieval.

Evaluation of AI models

To be able to trust an AI assistant in the public sector, it must be reliable, correct, and secure. Ali Alladin explains how the data and AI team systematically works with evaluations to quality-assure Svea. Hear about different methods for testing everything from the model's ability to follow instructions to the RAG system finding the correct information. Evaluation is not just a way to produce a number, but a diagnostic tool for driving continuous improvements.

Collection of user intent and data collection under secrecy requirements

How can an AI assistant be improved without access to user data? Danila Petrelli describes the challenge of data collection under strict secrecy requirements, where developers cannot view users' chats. The solution is a new annotation task to understand the user's intent and preferred response format. In parallel, a secure function is being developed where specially trained users can voluntarily donate harmless chats, a work carried out in close collaboration with legal experts. 

Lessons learned and insights from the Svea project

Jonatan Permert on important insights and lessons learned from the project. He touches upon, among other things, the importance of AI transformation, the need for customized AI services for the public sector, and challenges such as data readiness and legal interpretations. The presentation provides an overview of the opportunities and challenges the public sector faces regarding the implementation of generative AI.