During the spring of 2018, the Finnish Centre for Pensions made a small-scale proof-of-concept (poc) machine learning test to classify imminent disability pensions among the working age population of Finland. We made the classification to predict disability pensions that will start two years in the future. We used open source tools for building the machine learning models. The project lasted for about four months. The primary object of our poc test was to learn how well the standard machine learning algorithms can predict disability pensions in the near future based on register data available to us. Another goal was to make a write-up of the project and publish it on our website (https://www.etk.fi/en/tiedote/artificial-intelligence-identifies-future-retirement-on-a-disability-pension/). The tools we used were standard machine learning libraries such as scikit-learn and TensorFlow/Keras for Python. We also ran some of the neural network models using Nvidia GPU. Our data sample consisted of the following information of about 500 000 persons: - income data for the past 10 years; - sick leave history for the past 5 years; - socioeconomic status; - education level; - line of work; and - personal information (age, sex and marital status). Half of our data concerned people who had retired on a disability pension in the last 10 years, and the other half formed the control group. The machine learning algorithms predicted disability pensions two years into the future with an accuracy rate of 78 per cent (where 50% is the baseline accuracy), with an 86 per cent AUROC score. We used a couple of models, including random forest, logistic regression and neural networks. The best predictors for retirement on a disability pension were old age, low income and a history of sick leave. Often a low educational level and unemployment predicted retirement on a disability pension. The results of our test were in line with previous studies done at the Finnish Centre for Pensions about retirement on a disability pension in Finland.