Autopilot mode
One toggle that promotes the AI from teammate to first responder. New tickets get answered before a human looks.
Autopilot mode
Autopilot is a single toggle on the AI overview page (/ai). When you turn it on, two things happen at the same time:
- Reply mode for new conversations changes to auto-send (
reply_mode='auto'). 2. Routing default-assignee changes to AI-first, meaning the AI is the first responder and humans only see the ticket if it bounces or escalates.
Autopilot is reversible. Flipping it off restores reply_mode='draft' and your previous default assignee atomically. Nothing in flight gets lost.
Turning it on requires content
Autopilot needs at least 3 published help center articles before the toggle becomes available. The AI cannot retrieve from an empty brain. You will see a "Help center too thin to autopilot" warning on the card if you are below 3.
Who should turn it on
Teams who:
- Have at least two weeks of AI receipts they trust. See AI receipts. - Have voice and length dialed in. See Voice and tone and Reply length. - Have a confidence floor set high enough that auto-send only fires on confident drafts. See Confidence thresholds. - Have bypass labels enabled so the AI never replies to bugs, abuse, or spam. See Guardrails and bypass labels.
If any of those four are not in place, leave Autopilot off and stay in draft mode. See Auto-send vs draft vs suggest.
What Autopilot still respects
Autopilot is not "AI does whatever it wants". It still obeys:
- Confidence floor for auto-send. Below it, replies fall back to draft. - Bypass labels. Bug, abuse, and spam tickets skip drafting entirely. - Per-channel reply mode overrides. If you set chat to auto-send and email to draft, Autopilot keeps that. - Quarantine. Newly-promoted Q&A pairs and recently-edited articles wait in a quarantine state before the AI uses them on live tickets. - Spend caps. When you hit your monthly cap, AI pauses. See Spend caps and alerts.
What changes the moment you flip it
- Turning Autopilot on triggers a re-scan of every published article so the brain is up to date. The card shows live progress: articles processed, chunks indexed, and cost. - New conversations created after the flip route to AI-first. - Existing open conversations are not retroactively reassigned.
What the customer sees
Nothing different. Replies arrive from your support address as usual. There is no "this is an AI" banner unless you choose to add one. Many teams add a small note in the email signature, like "drafted by our AI agent, edited by a human". That is up to you.
Watching Autopilot in production
Three things to watch in the first week:
- Auto-send rate. What percent of new tickets get sent without a human. A healthy number is 40 to 70% for well-tuned setups. Higher is fine if confidence and reviews look good. 2. Bounce rate. Conversations that came back after an AI reply. If this jumps, voice or facts may be off. 3. CSAT. AI replies get scored the same as human replies in CSAT per channel. Watch for drops.
Pausing Autopilot
You can flip the toggle off at any time. New conversations revert to your previous reply mode and routing. In-flight tickets keep their current state.
For a faster brake, the Channel master toggles can disable AI on individual channels in a single click.
Common patterns
- Soft launch. Auto-send for the chat widget. Draft for email. Most teams start here. - Topic gate. Auto-send for "how-to" tickets. Draft for "billing". Set in routing rules. - Confidence gate. Auto-send only above 90% confidence, draft between 70 and 90, silent below 70.
Recommended order of operations
- Run two weeks in draft mode. 2. Audit AI receipts and CSAT. 3. Set confidence floor and bypass labels conservatively. 4. Flip Autopilot on. 5. Loosen confidence floor as you get more data.
Related
Was this article helpful?