Carry forward the work already owed.
An investigation waiting on a partner needs an owner, context, and a next action after replacement. A payment awaiting confirmation cannot become successful because the old application was retired.
Existing systems and people may provide that continuity. During migration, check that they can find and continue the work. Investigate a missing result before repeating an action.
Review what the new participant may do.
Suppose the earlier assistant drafted replies for a person to send. Its replacement can send eligible replies directly. Better drafting performance does not by itself justify the additional authority.
Review what can now reach the recipient, which decisions the person retains, and when intervention remains possible. Changing the model and changing its permissions are separate decisions.
Recheck evidence affected by the change.
An unchanged business rule can stay. Tests of an unchanged access limit may still apply. Retest routing if the model chooses different paths or serves different people.
Source changes also matter. A provider's status label may mean something different or arrive earlier in the process. The same word can describe a different result.
Test what changed. Keep earlier work where the reasons for trusting it still hold.
During the transition, decide what continues, what needs review and who handles uncertain work. Reduce AI's responsibilities where needed while keeping accepted cases moving.
A useful replacement review leaves two clear answers: Who carries the unfinished work? What has changed enough to require a new decision or new evidence?
Consider the next responsibility ยท Read why the work outlasts its participants