Machine learning and AI models are often labelled as "black box" approaches and given the requirement to not leave discretion, how do we interrogate the product and allow transparency?