Insulin drafts email replies and puts them in an Approvals queue. It does not send them. The only exception is a rule you explicitly authorised to send without review — which means every automatic email that leaves your account is one a person decided should be automatic.
The anxiety about AI on an inbox is not that the drafts will be bad. Bad drafts are visible and fixable.
It is that something will go out that nobody read. A reply to the wrong thread, a confident answer to a question the model misread, a tone that was fine in general and wrong for this customer. Email is the one channel with no undo, and the recipient is outside your company.
So the default is the conservative one, and the design question that remains is narrower and more interesting: which of your rules has actually earned an exception?
What happens by default
Insulin never sends email on its own. Every AI-generated response waits in the Approvals queue for a person, and sending is a separate, explicit step — because sending cannot be undone.
The queue is a working surface rather than a checkpoint. The reply body is inline-editable, and edits save automatically when you click away from the field, so the ordinary flow is read, adjust a sentence, send. Clicking Send raises a confirmation dialog before anything leaves.
That is three deliberate frictions on the path from draft to sent: the queue, the edit, the confirmation. Each exists because the action is irreversible.
What the queue is really for
It is tempting to read the queue as a safety net — a place to catch mistakes. It is more useful as a calibration instrument.
For the first few weeks, the queue tells you how well each rule understands your business. A rule producing drafts you send unedited is a rule that understands its category. A rule producing drafts you rewrite every time has a filter that is too broad or context sources that are too thin.
Read the queue with that question in mind and it stops being overhead. You are not checking the AI’s work; you are finding out which rules are ready for more responsibility.
Which rules deserve auto-send
Auto-send exists for rules explicitly authored to skip review. The bar should be high, and three tests make it concrete.
Is being wrong cheap and reversible? An acknowledgement sent to the wrong person is awkward. A price quoted to the wrong person is a commercial problem. Only the first category is a candidate.
Does the reply contain any commitment? A date, a number, a promise about what happens next — anything a recipient could reasonably hold you to. If yes, it goes to the queue no matter how good the drafts have been.
Has the rule earned it with evidence? Not “the drafts seem good” but “I have approved forty of these unedited.” A rule that has never been wrong across a meaningful sample has demonstrated something. A rule that has been right five times has not.
The shape that passes all three is narrow: acknowledgements, routing confirmations, “we have received this and someone will reply by Tuesday.” Useful, high-volume, and harmless when wrong.
The third option most teams skip
A rule does not have to choose between drafting and sending. A destination can be a Slack notification — a human is told something arrived, and nothing is composed at all.
This suits the category that causes the most trouble in automated inboxes: mail that needs a person, not an answer. An escalation, a legal question, a customer who is unhappy. Drafting a reply to those is the wrong move even with review, because the draft anchors whoever reads it to a response that may be entirely misjudged.
Route them to a human, with no draft attached. It is the least clever option and often the correct one.
How this relates to agents generally
The same boundary runs through the whole workspace: an agent proposes, a person disposes, and the exceptions are ones a human authorised in advance. Human approval for AI agents covers the general pattern; the inbox is its strictest case because the action is irreversible and the audience is external.
If you are configuring the rules themselves, setting up email rules without writing code covers the filter and context decisions that determine draft quality in the first place — and a rule with good context is a rule you will eventually trust.
Frequently asked questions
Does Insulin send emails automatically? No. Insulin never sends email on its own. Drafts wait in the Approvals queue, and sending is a separate confirmed step because it cannot be undone.
Can automatic sending be enabled? Yes, per rule, and only if you explicitly authorise it. Auto-send exists for rules a person has deliberately chosen to exempt from review.
Can I edit a draft before sending? Yes. The reply body is inline-editable in the Approvals queue, and edits save automatically when you click away from the field.
Which rules should auto-send? Ones where being wrong is cheap and reversible, that contain no commitment, and that have produced a meaningful run of drafts you sent unedited.
What if an email needs a person rather than a reply? Route the rule to a Slack notification instead. A human is alerted and no draft is composed, which avoids anchoring them to a misjudged response.
Is the Approvals queue just a safety net? It is more useful as calibration. Which rules produce drafts you send unedited tells you which ones understand their category.
Takeaways
- Insulin never sends email on its own. Auto-send is a per-rule exception a person authorises.
- Three deliberate frictions guard an irreversible action: the queue, the inline edit, the send confirmation.
- Read the queue as calibration, not as checking. It tells you which rules are ready for more.
- Auto-send only where being wrong is cheap, nothing is committed, and a real run of drafts went out unedited.
- Mail that needs a person should notify a person. A draft anchors the reader to a response that may be wrong.
An inbox assistant is worth having only if you can predict what it will do. See how the Inbox App keeps every send behind a person.
Sources
Primary sources for the platform rules cited above. Last verified August 16, 2026. Cloud providers change fees, eligibility, and program terms without notice — check the source before relying on a figure.
- Suger Insulin docs: Inbox — That Insulin never sends email on its own, the Approvals queue and inline-editable draft body, the separate send confirmation because sending cannot be undone, and that auto-send exists only for rules a user explicitly authorises
Stay Updated
Get the latest Cloud GTM insights, product updates, and marketplace strategies delivered to your inbox.