Auto-Send & Confidence

Answering in seconds only counts if the answer is right. Auto-send looks like one decision but is really a sequence of them: whether to reply at all, what to ground the reply in, how sure the model is, and whether anything about this conversation makes sending unwise. This is that sequence.

The pipeline#

  • 1. Filter - is this even a message the agent should answer?
  • 2. Identify - which lead and conversation does it belong to?
  • 3. Retrieve - what does the knowledge base know, and do any custom tools apply?
  • 4. Draft - the model writes a reply and rates its own confidence.
  • 5. Evaluate - escalation triggers, then the confidence threshold, then the channel toggle.
  • 6. Send or hold.

Retrieval and tool calls run in parallel with the rest of the work rather than one after the other, which is why an agent with knowledge and tools isn't slower than one without.

Filters before a reply exists#

Some messages never reach the model at all:

FilterWhat it drops
Relevance classifierNewsletters, marketing blasts, job alerts, OTPs, receipts for unrelated services and general automated noise. It fails open - if the classifier errors, the message is treated as real rather than silently dropped.
Security & bankingMessages classified as security alerts or banking mail are dropped before any agent sees them.
Your privacy exclusionsThe sender domains, addresses and subject keywords you block in Settings, plus the preset banking and security categories.
Self-send guardMail from your own connected address, which would otherwise create a reply loop.

Dropped is not held

A filtered message doesn't appear as a draft awaiting approval - there is no reply to approve. If you expected a conversation and don't see one, it was filtered, not held.

Where confidence comes from#

The model emits a 0-1 confidence score alongside every reply. It scores based on how well grounded the reply is in its knowledge base or conversation context - not how fluent the text sounds. A safety layer then adjusts based on the model's own intent and knowledge classification.

Turn kind (from the LLM)What the gate does
Action (greeting, thanks, scheduling, confirming)Keeps the model score. The prompt asks the model to score true greetings at 0.92+; a mid-band score is not raised into auto-send.
Information with strong knowledge matchKeeps the model score. If grounded, it should already be high.
Information with weak match (<55% similarity)Caps at 55%. A loosely-matching passage cannot push a reply through the threshold.
Information with no knowledge hitCaps at 45%. The model should say let me check and score low; the cap is a safety net.
Mixed (action + info) with no knowledgeCaps at 65%. The action part is fine but the ungrounded info part is risky.

Escalation still overrides everything

Negative sentiment, a request to talk to a human, or a sensitive topic (refund, legal, complaint) holds the reply regardless of confidence score.

The gates#

Once a reply exists, these are checked in order:

#GateIf it fails
1Escalation triggers - negative sentiment, a request for a human, a sensitive topicHeld and marked escalated, with the reason. Confidence is not consulted.
2Lead status - Active Client or ClosedAuto-send is blocked outright. Existing customers get a person.
3Effective threshold - your setting, raised to at least 90% when the lead is LostHeld as a low-confidence draft.
4Channel auto-reply toggleHeld, with a reason saying the channel is off rather than that the reply was poor.
5Conversation-level automation - paused, taken over, or a terminal stageHeld, with the specific reason.

What a held draft tells you#

Held drafts carry a plain-language reason rather than an error code, and each one points at a different fix:

ReasonThe real fix
Confidence NN% - below the NN% thresholdKnowledge, not the slider. Find the question in knowledge gaps and answer it.
Escalated to you for human reviewNothing - this is the system working. Check the trigger was fair.
Auto-send is off for this channelA toggle, not a quality problem. The reply was good enough to send.
Paused - manual / agent off / stageAutomation is off here. Re-enable it if that wasn't deliberate.
Needs review - contradicts callA follow-up disagreed with what was agreed on a call. Read the call notes before sending.

Raising your auto-send rate#

Do these, in order

  • Clear the knowledge gaps list. Every entry is a question that scored badly, with the answer one click from being written back into the knowledge base.
  • Make knowledge specific. Detailed sources with numbers retrieve better and score better than vague summaries.
  • Reduce contradictory rules on the agent. Rules that fight each other produce hedged replies and lower scores.
  • Turn the channel toggles on. A large held queue where every reason says 'channel is off' isn't a quality problem at all.

Don't do this first

Dropping the threshold raises the auto-send rate without improving a single reply. It is the right move only once you've read a batch of held drafts and concluded they were genuinely fine.

For the settings themselves, see Auto-Send Policy.