Most AI demos skip the boring part. The agent reads, decides and acts, all in one smooth motion. That works on stage. It does not work in a finance team, where one wrong entry can take a week to unwind.

Plans before work

In Canopy, every request becomes a plan first: which folders will be read, which rules apply, what will be produced and what will not be touched. A person reads it and approves it, changes it, or rejects it. Approval authorizes that plan only.

If a plan is rejected, nothing has run. That sentence has to be true every time.

Exceptions, not everything

Checking every line of an agent's work would cancel out the time it saved. So the Reviewer re-checks totals in software and brings a person only the decisions that need judgment: a duplicate receipt, a branch total that does not add up.

Completion belongs to a person

The crew can prepare, match and flag. Only a person can mark the work complete. That is a design rule, not a setting.