The investigative questionnaires that are required to underwrite derive inspiration from the AI risk management frameworks like NIST AI RMF.
- The investigative questionnaire for initial review focuses on outlining the AI use case, business function, stakeholders of the AI system, type of algorithm and some questions on AI governance practices of the organization.
- The investigative questionnaire for the detailed analysis of the risks associated with an AI system focuses on practices/controls adopted during data collection, annotation process, data processing, training of the AI and production setup. These are easy to extract if the organization has an algorithmic audit report available.
It is also important to collect metrics on data quality, annotation quality, training quality and production quality. For example, while underwriting AI performance guarantee, it is important to take the error distribution characteristics like variance and mean into consideration. The AI models are most sensitive to the data. Hence, small changes in the distribution of the data during production can plummet the model accuracy. Insufficient controls to ensure data correctness in production can badly impact the model performance down the pipeline. The annotation of data specifies the ground truth and in some scenarios (e.g. healthcare) it’s hard to get the annotations within an acceptable confidence interval.