Auto-send is how a reply reaches a customer with nobody reading it first. It is deliberately hard to trigger by accident: several independent conditions all have to agree, and any one of them can hold a reply back.
A reply sends itself only when all of these are true:
Fail any one and the reply becomes a draft in the inbox instead. Nothing is lost.
Every reply carries a confidence score from 0 to 1. The model proposes one; a gate then caps ungrounded and weakly retrieved factual answers so a leftover sales self-score cannot send. It does not raise a score into the auto-send band. Each agent has its own threshold, set on a slider in its Autopilot sidebar.
| Value | Meaning |
|---|---|
| 90% | The default for a new agent - conservative on purpose. |
| 70% | The floor. Dragging below it asks you to confirm, because that's where replies the agent isn't confident about start going out unread. |
| Anything between | Your call. Loosen it for a support agent answering documented questions. Keep it tight for anything that quotes a price. |
Confidence is about grounding, not grammar
Auto-reply is per channel, so you can let the agent answer website chat instantly while every email waits for a human. Website chat defaults to on for new agents - a live chat widget that takes an hour to answer is worse than no widget - and the rest default to off.
Pausing the agent clears every toggle
Any single conversation can opt out of - or into - automation without touching the agent's settings. The conversation's own setting always wins over the agent's default, in both directions.
| Lead status | Effect on auto-send |
|---|---|
| Active Client / Closed | Auto-send is blocked outright. Existing customers get a human, whatever the score says. |
| Lost | The threshold is raised to at least 90%, so only a very well-grounded reply goes to someone who already said no. |
| Everyone else | Your configured threshold applies unchanged. |
Escalation triggers are evaluated on the customer's message before confidence is even considered. If one fires, the agent hands over regardless of how confident it was:
Every held draft says why in plain language, so you never have to guess which rule caught it:
| What you see | What happened |
|---|---|
| Confidence 62% - below the 90% auto-send threshold. | It scored short. |
| Escalated to you for human review. | A trigger fired. |
| Auto-send is off for this channel. | The score was fine, the toggle wasn't on. |
| Paused - manual / agent off / stage. | Automation is off for this conversation, and why. |
For the mechanics of how confidence is produced, see Auto-Send & Confidence.