Transition and AI: Äventyret 2024
The year cutbacks met AI as our new everyday, both inside the agency and in client projects.
How it went
2024 will be remembered at Äventyret as a year of transition. It was the year when an uneasy outside world forced us to make cutbacks. The year we introduced OKRs to steer and measure our work — but above all, the year when AI truly became an integrated part of our everyday, both as a tool and in our client projects!
In a short time we have seen a democratisation of AI services, where the technology has become more accessible than ever. This has made it possible for us as a product and service design agency to build powerful AI applications — without being an agency full of data scientists.
We are proud to be an agency that is more workshop than talk — a team of doers who make things happen. We are at our most effective in the middle of a project. Our delivery never consists of 70+ slides of recommendations; we prefer to deliver the end product our insights lead to. But historically we have been less good at sharing our practical experience and lessons learned. That is what we intend to change.
AI in practice: challenges and solutions
Most of our AI assignments have focused on employer organisations, where we have faced unique challenges. One of the biggest has been managing and structuring complex information from disparate data sources. We have developed automated solutions to split the information into manageable blocks (chunking), enabling efficient processing by the AI models.
We have also worked out new methods for indexing the data sources, improving vector search and thereby giving the AI models access to relevant and precise data. To reach the level of accuracy required — especially for legal questions — we have worked with fine-tuning. By fine-tuning the base prompts and adapting them to our clients' needs, we have been able to raise both efficiency and quality of their support, freeing up advisors to work on harder cases.
But AI is not just about speed. It is about making sure the information is correct and reliable. Together with our clients' subject-matter experts we have reviewed and quality-assured the solutions. That collaboration has not just secured the outcome, it has also increased knowledge and understanding of AI inside the client organisations — and served as anchoring.
Experimentation and innovation
Beyond support solutions, we have explored other creative applications of AI. One of these experiments was generative meditation. Using a RAG model as the base, we drew on existing meditations to create new tailor-made meditations from the user's input. The generated meditation texts are then turned into audio via Eleven Labs' text-to-speech API, giving the end user a personal and interactive experience in their headphones. Namaste!
The cobbler's children: a new website
2024 was also the year we finally launched a new website for our own operation. Here we use DiffusionBee to generate illustrations — a tool our designers can run locally on their laptops. By training the model on our graphic profile we can create consistent, on-brand imagery for our articles and projects.
The air guitar takes to the piano
For better and worse, AI has become an umbrella term for a whole range of different techniques and disciplines. Generative AI is stealing the spotlight right now, but machine learning has long been an important sub-discipline. So what is machine learning? It is about algorithms and models that let computers learn from data, rather than being programmed for each individual task. In other words: computers that get smarter with experience.
A brilliant example of machine learning is the air piano, an experiment our own Erik built (seriously, you should try it — it's ridiculously fun!). The air piano is built on a machine-learning framework called ML5. The framework is easy to get started with, but getting it to track the fingertips and simulate key presses was a real challenge. The result? An AI-driven piano you can play — without any keys.
Hidden in the snow
It has not escaped us that generative AI can write code. Even if the generated code still often raises a few eyebrows, it creates fantastic opportunities to go from idea or dream to a fast prototype. Like when Christoffer listened to a podcast about the space that forms between the snow cover and the ground in spring. With that inspiration he built a game where lemmings under the snow are hunted by, and dodge, foxes. It may not go down in history as game of the year, but what GPT-4o generated from the prompts made us laugh a lot.
One insight from generated code is that it can help experienced developers like Christoffer close smaller gaps in their understanding, at the same time as it can help beginners overcome giant obstacles. The difference is that the experienced developer improves and learns from the process, while the beginner often does not fully understand what has happened — only that something works for the moment, and can only hope that it keeps doing so.
Advice for those who have not yet started their AI journey
Start small
Pilot projects are perfect for testing, learning and exploring the possibilities.
Start from within
Build internal tools that make life easier for your experts. Involving subject-matter experts ensures the solutions are correct and functional, while creating learning and buy-in across the organisation.
Focus on your strengths
Identify areas where AI can build on your existing competitive advantages.
Needs-driven
Just as before, but consider how AI can help solve them. Keep focusing on user needs — and on your own — needs!
It is exciting to think about where AI will take us next year. The rapid pace of development is inescapable and full of possibilities. We are ready to take on more assignments in this category and to continue the innovation journey together with our clients.