Most descriptions of an approval gate stop at "a human signs off before anything ships." That's true, but it undersells what the gate actually does, because approval isn't the only outcome — and the other two are where most of the interesting work happens.
Three outcomes, not two
- Approved as-is — the draft goes out unchanged. This is the outcome you want to become more common over time, not the only one you should expect from the start.
- Edited and approved — a reviewer changes the draft before it ships. The edit itself becomes part of the record, not just the fact that a change happened.
- Rejected — the draft doesn't ship, and the reason is captured, not just the verdict.
Each of the three lands in the same audit trail — logged, timestamped, and attributable to the person who made the call — described in more depth in our piece on STIV's security architecture. Nothing about a rejection is quieter than an approval.
Where the correction actually goes
An edit or a rejection isn't just a decision about that one draft — it's a signal about the playbook the agent is working from. If a Finance report keeps getting the same line item edited the same way, or a Legal redline keeps getting the same fallback clause swapped in, that's a playbook gap, not a one-off mistake. Reviewing the pattern in the audit trail, not just individual incidents, is how a playbook actually improves — the same discipline covered in how to measure whether a division is working.
Why a rejection isn't a failure state
Teams sometimes read a string of rejections as evidence the system isn't working. Early on, it's closer to the opposite — a healthy approval process should be rejecting and editing plenty in month one, because that's the correction signal the agent needs before autonomy expands to more of its output. A rejection rate near zero in week two, as covered in the first 30 days with STIV, usually means a reviewer stopped reading closely, not that the drafts got perfect overnight.
The gate isn't there to eventually disappear. It's there so that when it does let more through unreviewed, that decision is backed by a record of exactly how the agent has handled disagreement — not just how often it happened to be right.