How Autopilot decides
The exact logic behind whether the AI answers a customer, drafts for review, or escalates to a human.
On this page
Every inbound message runs through the same decision, in this order. The first rule that matches wins.
1. Is Autopilot on?
If the master switch is off, the AI never sends. It still classifies and can still draft, but a human sends everything.
2. Is this a money or judgement request?
Refunds, billing disputes, cancellations, and anything touching a payment are always escalated to a human. This is not affected by your confidence threshold and cannot be switched off.
The reasoning is that the cost of an AI being wrong about a refund is asymmetric — a wrong answer costs money and trust, while a slightly slower human answer costs neither.
3. Did the customer ask for a human?
An explicit request — "let me talk to someone", "is anyone there" — hands off immediately. See Human handoff.
4. Is there an override for this intent?
Each intent can be set to one of three modes:
| Mode | Behaviour |
|---|---|
| Auto | The AI may answer alone, subject to confidence |
| Assist | The AI drafts; a human reviews and sends |
| Escalate | Always goes to a human |
Use Assist for anything you are not yet comfortable automating. It is the setting most teams should start on for most intents.
5. Is the AI confident enough?
The classifier returns a confidence score. Below your reply threshold, the conversation escalates rather than guessing.
This is the setting that matters most. Confidence is not certainty — a threshold of 70 means you accept the AI answering questions it is meaningfully unsure about.
6. Is there quota left?
If your plan's AI reply allowance is exhausted, conversations escalate to humans rather than stopping. Running out of quota degrades the service; it never drops a customer.
See Plans and limits.
Where the config lives
Autopilot settings are stored on your workspace record on the server, not in your browser. This matters: the AI answers customers from the server, so a threshold that only existed in one person's browser would never affect a single real conversation. Change it in one place and it applies everywhere immediately.
Watching it work
Every ticket records what the AI decided and why — the intent, the confidence, and the reason for any handoff. Read these for the first week. It is the fastest way to find out whether your thresholds match your actual risk tolerance.
Related