In shadow mode the system sees the same real cases your team handles, but nothing it produces is written to a system of record or sent to a customer. You then compare what it would have done with what your team actually did, which shows how it behaves on live data that a test set may not cover.
Shadow mode sits between offline testing and a pilot. It is a stage with an exit condition, such as agreement with the team holding steady over a defined period, not a permanent state. Its limit is that it proves the output is right, not that the surrounding workflow is ready, so approvals, error handling and a way to switch back still need testing before the system is allowed to act. Evals on a held-back set of past cases usually come first.