AI assurance is the set of practices that produce evidence an AI system works as intended. The term comes from safety and audit disciplines, where assurance means independent, documented proof of a claim. In AI it covers testing, evaluation, monitoring, and the records that let a third party check how the system reached a decision.
Assurance differs from a one-time evaluation. An evaluation answers whether the system met a benchmark on a given day. That answer must be reconfirmed as data, models, prompts, and policies change. For example, a bank that retrains a credit model each quarter needs proof that the data behind each version was fit for that task, and a record that ties each production run to the version it used.
In practice, assurance rests on operating evidence: which data state a run used, what changed since it was qualified, and whether the team rechecked the affected scope. Without that record, a model that passed in testing can drift in production with no one able to show why.