Response Behavior

This page explains the core question every team asks:

"Why did Jardine reply here, but escalate there?"

Decision Model

When a message comes in, Jardine checks what the customer is asking, whether the answer can be grounded in your knowledge, whether routing allows AI to handle it, and whether any runtime account context is required. After that, channel controls decide whether the result stays as a draft or is sent to the customer.

Outcomes

A conversation can end in four ways: AI sends a reply, AI generates a draft only, the thread escalates to a human, or a handoff notice is sent if that is configured.

Confidence and Grounding

Confidence alone does not decide everything.

Jardine also evaluates whether the answer is grounded by relevant evidence. Low-confidence, low-grounding, or policy-sensitive intents are candidates for escalation.

Generate vs Send

Jardine separates generation from sending. Generation decides whether AI can produce a response. Sending decides whether that response can be delivered to a customer.

This gives teams a safe rollout path:

1. Enable generation.
2. Review behavior.
3. Enable sending after acceptance criteria pass.

Channel Behavior

Each channel can be controlled independently for generation and sending, so Intercom can run in draft-only while Zendesk or Email is already in auto-send.

Escalation

Escalation usually happens when a tag requires human handling, sensitive verification is needed, runtime account context is missing for an account-specific intent, or policy requires manual review.

Escalation is part of quality and safety control.

Ownership

Every thread has a clear owner: AI or Human.

Ownership makes handoff explicit and auditable in Conversations.

Tuning Order

If behavior looks wrong, tune in this order:

1. Fix source docs.
2. Fix routing rules/conditions.
3. Fix channel toggles.
4. Fix connector identity mapping.

Avoid broad policy changes until you isolate the layer causing drift.