When do I move an agent from review to auto?
Per action class, after a clean period. Roughly two weeks of approvals with an edit rate under 5% and zero rejections for that class. Move one class at a time, and keep external sends in review the longest.
The short answer
- Move an action class from review to auto only after a clean period. A reasonable starting bar is two weeks of approvals with an edit rate under about 5% and zero rejections for that specific class.
- Move one action class at a time, never the whole agent at once. An agent that has proven reliable at drafting summaries has not proven anything about its judgment on stage changes or sends.
- Keep external sends in review the longest of any action class. The cost of a mistake there is visible outside the company and cannot be quietly corrected the way an internal note can.
- Promote the class, not the agent. "This agent drafts well" is a statement about one action class. It says nothing about how the same agent will handle a different kind of decision, like when to change a deal's stage.
Why 'the agent has been good' is not a promotion criterion
It is tempting to reason about trust the way you would about a person. This agent has run for a month without incident, so it has earned more autonomy. That framing breaks down. An agent's reliability on one kind of task tells you very little about a different kind of task. An agent that reliably drafts accurate summaries is answering a narrow, well-defined question every time. An agent deciding whether to move a deal to a new stage is making a judgment call with more ways to be subtly wrong. A clean record on summaries says nothing about that judgment.
This is why the unit of promotion has to be the action class, not the agent as a whole. "Move the drafting to auto" is a specific, testable claim you can measure with edit and rejection rates. "Trust this agent more" is not testable in the same way. It is how teams end up granting broad autonomy on the strength of narrow evidence.
Action class, test, and how long to hold it in review
| Action class | Suggested clean-period test | Keep in review? |
|---|---|---|
| Summaries, tagging, scoring, internal notes | Two weeks, edit rate under 5%, zero rejections | Can move to auto once the bar is met |
| Classify reply, suppress on unsubscribe/bounce | Two weeks, edit rate under 5%, zero rejections | Can move to auto once the bar is met |
| Create deal from positive reply, book meeting | Longer clean period, since a wrong call here affects the pipeline | Review, then auto after a demonstrably clean stretch |
| External sends (email, LinkedIn, SMS, proposals) | No fixed graduation point; hold in review the longest | Keep in review indefinitely as a default |
| Stage changes, contact merge or delete | Rarely graduates; the cost of a mistake is high and hard to reverse | Keep in review |
What the clean-period numbers are actually protecting against
The two-week window and the roughly 5% edit rate are not arbitrary rituals. They exist to catch a pattern before it becomes routine. A single good week can be luck, especially if volume was low or the contacts involved were unusually easy cases. Two weeks gives enough volume for a real pattern to show. A 5% edit ceiling means the agent's output is close enough to what a person would have written that editing it is a minor adjustment, not a rewrite.
Disclosure: SalesCrew is our product, and its agent policy is built around this same per-class promotion logic. Research, tagging, scoring and internal tasks default to auto. Classification and suppression events default to auto. Creating a deal from a reply or booking a meeting starts in review and can move to auto after a clean week. Stage changes, merges and deletes stay in review. Any spend stays human-only, however long any agent has run well elsewhere.
Promote the class, never the agent
Questions
- What counts as 'zero rejections' if a reviewer just edits everything instead of rejecting?
- Editing instead of rejecting can hide the same signal a rejection would show. Track edit rate closely too. If every item needs a meaningful edit before approval, that class is not clean yet, even with zero outright rejections. The edit rate threshold exists to catch this.
- Can a class go back to review after being moved to auto?
- Yes, and it should if the confidence threshold starts routing more items to review, or if a reviewer spots a pattern of mistakes. Moving to auto is not a one-way door. The same evidence that justified the move justifies reversing it.
- Does 'moving an agent to auto' mean the whole agent stops being reviewed?
- No. Move the action class, not the agent. An agent that drafts well might move its summaries and tags to auto while its stage-change suggestions and sends stay in review indefinitely. Those are different action classes with different risk.