The most common question we get about Signate is some version of: how do you know when it is safe to send without me reviewing it first?
It is the right question. Auto-sending an email you would not have sent yourself is worse than not sending anything at all. So the system has to be genuinely conservative where it matters, not just cautious-sounding in a product description.
Here is how it actually works.
The confidence score
Every incoming email gets classified before a draft is generated. That classification looks at three things: how well the message fits a pattern Signate has seen before, how closely the situation maps to your voice profile, and whether there are any signals in the message that suggest elevated stakes.
A message from a regular contact asking a routine question about a deliverable gets a high confidence classification. Signate has handled similar messages before. The reply is predictable. The cost of a slightly imperfect reply is low.
A message from an email address Signate has never seen, with a question that does not map cleanly to past patterns, gets a lower score. Signate drafts a reply and holds it for your review rather than sending it automatically.
The score is not just based on the content of the message. It incorporates the sender's relationship history, the thread context, and the output of the voice-matching step. A high-confidence draft is one where both the intent classification and the voice match looked strong.
The escalation signals
Beyond the base confidence score, there is a separate escalation layer. Certain signals in an incoming email trigger a review request regardless of how confident the system is about the draft.
The threshold you control
None of this is fixed. You set your own confidence threshold in account settings. Some users start conservative, requiring review for anything that is not clearly routine. Others run with a higher threshold and let Signate handle the majority of their queue automatically, only surfacing the escalated cases.
Most people land somewhere in the middle and adjust over time. As you see how the system performs across different message types, you develop a sense of where it is reliable and where you want to stay in the loop.
The audit log keeps a complete record of every message Signate acted on, what it did, and why. If you ever want to understand a pattern in how it is performing, the log is the place to start.
What happens when you edit a draft
When a draft comes up for review and you change it before sending, Signate notes what was different. Not to immediately update any underlying model, but to inform how similar drafts get generated going forward.
If you consistently shorten the closing paragraph in replies to a specific category of sender, that pattern gets incorporated. If you always remove a particular phrasing the system defaults to, it learns to stop using it in that context.
This is not the same as fine-tuning. We are not retraining anything on your edits. The improvement is in how your voice profile gets applied to new situations, informed by your corrections over time.
Why the conservative default matters
There is a design philosophy behind the default behavior, which is to err toward asking you rather than guessing when anything feels ambiguous.
An email assistant that occasionally asks for review on something routine is a mild inconvenience. An email assistant that occasionally auto-sends something you would not have sent is a trust problem that is hard to recover from.
The goal is a system that earns more autonomy over time, not one that asks you to take it on faith from day one. Starting conservative and loosening as confidence builds is the right order.