Set up AI routing

Configure auto-send, reply timing, and Signal Guard so AI-generated responses reach your customers safely.

What AI routing does

When a routing rule fires a generate_reply action, AI writes a response using your brand voice and message context. AI routing controls what happens next:

Should the reply send automatically, or should a human review it first?
How long should the AI wait before sending (to feel natural)?
What happens when no routing rule matches: should AI still reply?

These settings sit between AI reply generation and the customer. Think of AI routing as the approval gate.

Where to find it

Go to AI & Automation → Routing Rules in the sidebar, then click the AI Routing tab.

Step 1: Enable AI reply generation

The AI reply generation master toggle must be on for any AI replies to generate. When off, generate_reply actions in your routing rules are skipped entirely.

This toggle is separate from your plan's AI capabilities: even on Pro/Enterprise, you can turn off AI replies while keeping AI classification and scoring active.

Step 2: Enable fallback replies

Fallback replies determine what happens when an inbound message doesn't match any routing rule.

On: AI generates a reply even when no rule matches. This ensures every message gets a response.
Off: Messages that don't match any rule get no AI reply. Only rules with explicit generate_reply actions produce responses.

For most businesses, turn fallback replies on. It catches the long tail of messages that don't fit neatly into your routing rules.

Step 3: Configure auto-send vs. human review

This is the core decision:

Auto-send ON:

AI replies that pass Signal Guard guardrails send automatically to the customer
Replies flagged by guardrails are held in the Work Queue for agent review
Best for: high-volume businesses where speed matters and your brand voice is well-tuned

Auto-send OFF:

Every AI reply goes to the Work Queue for human review before sending
Agents can approve, edit, or dismiss each draft
Best for: new setups, regulated industries (healthcare, legal), or businesses still refining their brand voice

Recommendation: Start with auto-send off for the first week. Review the drafts in your Work Queue. Once you're confident the AI is generating appropriate replies, turn auto-send on.

Step 4: Configure reply timing (Reply Wait)

When auto-send is on, Reply Wait adds a random delay before sending the AI reply. This makes responses feel more natural — an instant reply can feel robotic.

Min delay (default: 30 seconds): The shortest possible wait before sending.
Max delay (default: 120 seconds): The longest possible wait. The actual delay is randomly chosen between min and max.

The delay only applies to auto-sent SMS and email replies. Voicemail and missed call auto-replies always send immediately — callers expect a prompt response.

To configure reply timing:

Enable Auto-send (Step 3)
Toggle Reply Wait on
Set your preferred min and max delay (30–600 seconds)

Tip: Start with the defaults (30–120s). If customers mention responses feel slow, reduce the max. If responses feel too fast, increase the min.

Step 5: Understand draft gate reasons

When a reply is held in the Work Queue, the gate reason tells you why:

Gate reason	What it means	What to do
`auto_send_disabled`	Auto-send is turned off, so all drafts are held	Expected behavior when auto-send is off
`guardrail_review`	Signal Guard flagged the reply for human review	Check the reply for prohibited phrases or compliance issues
`judge_failed`	The AI safety judge couldn't evaluate the reply	Review manually — the guardrail system had an error
`pii_detected`	PII was detected in the generated reply	Review carefully; the reply may contain personal information that shouldn't be sent
`no_sender`	No messaging sender is configured for the reply channel	Check your phone or email setup
`landline_phone`	The contact's phone is a landline (can't receive SMS)	Consider calling instead, or dismiss the draft

Gate reasons appear as an info icon on held drafts in the Work Queue. Click or hover to see the reason.

How it fits in the pipeline

Inbound message
  → Classify (intent, urgency, sentiment)
  → Score (lead quality 0–100)
  → Match routing rules
  → Rule fires generate_reply action
  → AI generates a reply using brand voice
  → AI ROUTING decides: auto-send or hold for review
  → If held → Work Queue (agent reviews)
  → If auto-send → Reply Wait delay (30–120s random)
  → Signal Guard check (regex + AI judge)
  → If Signal Guard passes → Send to customer

AI routing sits after reply generation and before Signal Guard guardrails. Even auto-sent replies still pass through Signal Guard before reaching the customer.

Relationship to Signal Guard

AI routing and Signal Guard guardrails are different checkpoints:

AI routing: controls whether replies auto-send or wait for human review
Signal Guard: compliance-based. "Does this reply violate any prohibited phrases, compliance rules, or industry regulations?"

A reply can pass AI routing (auto-send enabled) but still be flagged by Signal Guard guardrails (contains a prohibited claim). Both must pass for a reply to send automatically.

Tips

Start conservative. Auto-send off for the first week. You can always enable it later.
Check your Work Queue daily during the first week. It shows you exactly what the AI is generating and why drafts are held.
Fallback replies are powerful. They handle the messages you didn't anticipate when building routing rules.
Reply Wait makes auto-send feel human. A 30–120 second delay prevents the "instant bot reply" feel.
Changes save automatically. Toggling auto-send, adjusting reply timing, or enabling fallback replies takes effect immediately.

Set up AI routing

On this page