Reverting AI agent edits in a CRM requires per-record before-images and run-level undo

An AI agent with write access to a CRM can make mistakes like bad merges or mass field overwrites, and an audit log alone does not restore the lost data. A safe undo needs before-and-after snapshots of each record, stored append-only with the actor, run identifier and request ID. Reverts should operate on whole agent runs rather than individual changes, and the agent's own credentials should be unable to undo its work. A reliable revert also skips records a person has edited since, reporting the reason, and keeps the full history visible. Emails that have already been sent cannot be recalled, so outbound workflows benefit from review queues and dry runs.