Auto-send vs draft vs suggest
Three reply modes shape how visible the AI is in your inbox. Pick one per channel, or layer them.
Auto-send vs draft vs suggest
Ochre's AI agent has three reply modes. Same drafts, different levels of human involvement. You pick the one that fits each channel and team.
The three modes
Auto-send
The AI sends replies on its own when confidence clears your floor. Internally reply_mode='auto'. The agent sees the conversation already answered when they open it.
Best for:
- High-volume, low-stakes channels (chat widget, FAQ-style email).
- Channels where speed beats nuance.
- Teams that have audited two weeks of AI receipts and trust the output.
Draft
The AI writes the reply directly into the agent's editor. Internally reply_mode='draft'. The agent sees the customer's message, the draft already typed in the composer, and one button to send. They can edit freely first.
Best for:
- Most teams, most of the time.
- Email, where tone matters more.
- The first month of using AI in a workspace.
Suggest
The AI shows its draft beside a blank composer. The agent writes their own reply, with the AI version visible as a reference.
Best for:
- Senior teams who want AI as a sanity check, not a starting point.
- Channels with strong style guides.
- Periods of voice tuning where you want to see what the AI would say without it influencing the agent.
How to set the mode
- Open AI → Drafting.
- Pick the workspace default.
- Optionally override per channel.
Per-channel overrides are common. Most teams run draft on email and auto-send on chat.
How confidence interacts
Each mode is gated by your confidence floor. See Confidence thresholds.
- In auto-send, replies below the floor fall back to draft.
- In draft, replies below the silence floor are not surfaced at all.
- In suggest, replies below the silence floor are not shown.
That cascade means a confidence floor of 85% with auto-send still gets you human review on the unsure 15%.
Bypass labels
Bug, abuse, and spam never get drafted regardless of mode. See Guardrails and bypass labels.
Switching modes
Switching mode is hot. New messages use the new mode immediately. Drafts that were already created stay in their current state.
Mode interactions with Autopilot
Autopilot mode flips reply mode to auto-send for new conversations and changes routing to AI-first. Flipping Autopilot off restores your previous mode setting.
Receipts behave the same in every mode
Whether the AI sent the message or the human did, the reply has a receipt: model, tokens, cost, confidence, sources, and the action that was taken (sent, drafted, suggested, suppressed). See AI receipts.
Common patterns
- Cautious launch. Suggest everywhere for one week, draft for two weeks, then auto-send on chat.
- VIP exception. Auto-send for everyone except a "VIP" customer segment, where you stay on draft.
- Topic split. Auto-send for "how-to", draft for "billing", suggest for "abuse" (or just bypass it).
Editing and learning
When an agent edits a draft before sending, the edit is captured. Patterns across edits feed Quality assurance review so you can see where the AI is consistently off, and the Learned Q&A library promotes the corrected versions for future tickets.
Recommended starting setup
- Workspace default: draft.
- Override on chat widget: auto-send (with confidence floor at 85%).
- Override on VIP segment: suggest.
After two weeks of clean receipts and good CSAT, loosen up.
Related
Was this article helpful?