I've built pilots of these kind of models for my government department (Victoria, Austalia). The prototypes in R using Random Forest models. Pilot version in Python using gradient boosting. We could have used deep learning models, but the increase in accuracy of the predictions is pretty minimal.
We haven't used these models outside of a pilot. There is a long lag to get other departments/services data and de-identify it and link it (we have a team that does this, it has a lot of safeguards to ensure privacy). Although old data is still quite predictive. A simple predictive score didn't help our workers do their job as they needed to justify their decisions with evidence. Giving the workers access to underlying information (they would need to have legal authority to view the data, so that excludes Homelessness support staff) often is missing what they need for their processes.
We haven't used these models outside of a pilot. There is a long lag to get other departments/services data and de-identify it and link it (we have a team that does this, it has a lot of safeguards to ensure privacy). Although old data is still quite predictive. A simple predictive score didn't help our workers do their job as they needed to justify their decisions with evidence. Giving the workers access to underlying information (they would need to have legal authority to view the data, so that excludes Homelessness support staff) often is missing what they need for their processes.