We are passionate about machine learning and it’s potential to create value for organizations as well as to improve the quality of life for people. However, working in the field you soon realize that there are some challenges that prevents the potential from materializing.

In the early 2000’s this was largely due to a question of inadequate performance. The models where simply not good enough to make predictions or identify patterns to make it worthwhile to implement them. The performance issue where however vastly improved, if not solved, with the advent of deep learning and the access of more digitized data. Additionally, the world of open source allowed for very cost-efficient exploration through “proof-of-concepts”.

But still organizations and companies keep failing to walk the last mile, with surprisingly few implemented solutions. So, we kept asking ourselves why are they failing? There are of course countless reasons for failing, ranging from culture to more technical aspects. But there are two challenges that have captured our attention:

• Improper practices for deployment of machine learning solutions
• Lack of trust for “black box” models

Firstly, we have learnt a lot of how to employ service design and rapid prototyping to quickly formulate and explore the solutions and evaluate whether it is worthwhile to pursue. Secondly, if the solution proves to be worthwile to pursue, we believe that a large share of the effort should be reserved for the deployment phase with all its considerations, including how to handle the moving target of ever changing data.