Begin with the decision
An AI review should start with the decision a system supports and the consequences of getting it wrong. Identify who is affected, who uses the output, and who has authority to challenge it. A model’s technical performance is only one part of that context.
Describe the intended use and the uses that are outside scope. This helps prevent a model developed for one setting from quietly becoming a decision tool in another setting where its evaluation no longer applies.
Record where the evidence comes from
Document the data sources, relevant permissions, preparation steps, and known gaps. Check whether evaluation data resembles the setting in which the model will operate. Look for leakage between training and testing, and consider whether important groups or conditions are under-represented.
Keep the evaluation reproducible. Record model versions, configuration, metrics, and the reasons for choosing them. A later reviewer should be able to understand what was tested without relying on the memory of the original developer.
Test the workflow as well as the model
Assess how users interpret the output, when they are expected to override it, and what information supports that choice. A nominal human approval step offers little protection if the reviewer lacks time, authority, or a workable alternative.
Examine error patterns and performance across relevant groups and conditions. Avoid collapsing conflicting outcomes into one headline score. The review should make trade-offs explicit and identify who is responsible for accepting them.
Plan for change after launch
Define monitoring, escalation triggers, and conditions for pausing or retiring the system. Changes in input data, user behaviour, or the surrounding service can undermine an earlier evaluation even when the model itself has not changed.
Maintain an assurance record that grows with the system: purpose, data lineage, tests, approvals, incidents, and changes. The practical outcome is a decision process that can be inspected and improved, with responsibility assigned throughout the model’s life.