Routing and handoffs

Start by deciding how conversations should move between assistants and humans across channels.

• List your channels (SMS, chat, voice, email) and where each should start (for example, general support assistant).
• Identify specialized intents that deserve their own assistant or queue: Sales, Support, Billing, Technical, General.
• Define when AI escalates to humans (for example, refunds over a certain amount) and when humans can hand a conversation back to AI.
• Sketch at least:
  • One default route per channel (where new conversations land).
  • One or two escalation paths (for example, Billing → Human billing queue).

You should end this step with a diagram or list of which assistants and teams own which intents and channels.

Before you create routing rules, make sure the destinations are ready.

• In Assistants, configure:
  • A general assistant for broad questions.
  • Specialized assistants (for example, Billing, Technical, Sales) with focused training context and playbooks.
• Align each assistant to its supported channels (SMS, chat, voice, email) so you do not route traffic to an unsupported channel.
• Confirm each assistant behaves as expected using test conversations before sending routed traffic.

A rule is only as good as the assistant or queue it sends traffic to.

Open the Routing Rules tab and create a small set of high-value rules before you attempt full coverage.

Focus on routes that have clear patterns and high impact:

• Billing phrases such as invoice, refund, overcharged on chat or email.
• Sales intent on voice calls for a sales-focused assistant.
• Technical intent on SMS for a concise troubleshooting assistant.

Create the rules in this order:

1. High-priority exceptions (for example, VIP billing customers).
2. Main escalation rules (for example, Billing intent → Billing assistant).
3. Fallback default rule (for example, everything else stays with the general assistant).

Keep them inactive until you have tested them in the next step.

Switch to the Handoff Settings tab to decide how transfers behave and how often they are allowed.

Work through these in order:

• Turn on agent handoff when you are ready for live transfers.
• Pick a message style preset that matches your brand and channel tone.
• Edit the handoff message template to:
  • Introduce the new agent or assistant.
  • Summarize why the handoff is happening.
  • Set expectations using {agent_name}, {customer_intent}, and {task_description} placeholders.
• Set guardrails:
  • Start with max 3–5 handoffs per conversation.
  • Add a cooldown by messages (for example, 3–5 messages) and/or minutes (for example, 5–10 minutes).
  • Enable loop prevention to avoid back‑and‑forth like Billing → Support → Billing.

Start strict, then relax limits once you see stable behavior.

Still in Handoff Settings, configure when human queues should accept routed conversations.

• Choose the correct timezone for your primary operations team.
• For each weekday:
  • Enable the day.
  • Set start and end times for business hours.
• Decide after-hours behavior, for example:
  • Keep AI in control and show a message like Our team is offline, but I can still help with....
  • Allow handoffs only to specific 24/7 queues.
• Add holiday overrides for days your team is closed or operating on special hours.

This ensures the system does not route to humans when no one is available, and that customers get clear expectations.

Test before turning rules live on production traffic.

• Use internal test numbers, email aliases, or staging channels to trigger each routing rule.
• For each test:
  • Confirm the expected rule matches (intent or keywords).
  • Verify guardrails behave as expected (no rapid ping‑pong, loops blocked).
  • Check that the handoff message fills in {agent_name}, {customer_intent}, and {task_description} correctly.
• Ask a few agents to review the incoming handoff messages and confirm they have enough context to act quickly.

Only activate rules for production channels after they pass this test checklist.

Turn routing into an operational control surface, not a one‑time project.

• Activate a small set of rules first (for example, only Billing on chat and email).
• Monitor:
  • Handoff volume by rule and channel.
  • Blocked handoffs due to guardrails.
  • Resolution rates and agent feedback.
• Adjust:
  • Rule keywords and intent if they are too broad or too narrow.
  • Priorities when multiple rules match the same conversations.
  • Guardrail limits if agents see too much noise or too many escalations are blocked.

Expand coverage to more intents and channels once results are predictable.

On the Routing Rules tab, add a rule and start with:

• A descriptive name like Billing chat escalations or Sales voice warm transfers.
• The channels this rule should affect (for example, Chat and Email, but not Voice).
• A priority that reflects importance:
  • Lower number or higher value, depending on UI, for more specific or sensitive rules.
  • Ensure exception rules run before broad defaults.

Ask yourself: If two rules matched, which one should win? Set the priority accordingly.

Use triggers to match the right conversations without catching everything.

• For keywords, enter a comma‑separated list of phrases customers actually use:
  • Good: refund, wrong charge, invoice issue
  • Risky: billing alone (may be too broad with unrelated mentions)
• For intent, choose from the dropdown:
  • Use Billing, Sales, Support, Technical, or General when the classifier is accurate for that category.
  • Use None if you want the rule to rely only on keywords.

Combine intent and keywords for high‑risk paths. For example:

• Intent: Billing
• Keywords: refund, dispute, chargeback
• Destination: Billing assistant on chat and email

This catches serious billing cases without over‑routing generic questions.

Choose the target assistant that should take over if the rule matches and guardrails allow.

• Send Billing conversations to a billing assistant tuned for invoices and payments.
• Send Sales conversations on voice to a sales assistant or human sales queue.
• Keep low‑stakes or general questions with your general support assistant.

Verify the target assistant supports the rule's channels; do not route SMS to an email‑only assistant, for example.

When you activate a rule:

• Review other rules for the same channels and intents.
• Ensure priorities reflect your intent:
  • Enterprise billing rule should outrank General billing.
  • After-hours billing rule (if implemented separately) should outrank general billing during that time window when applicable.
• Run a few manual tests to confirm:
  • The right rule fires.
  • No unexpected rules match the same conversation.

Keep a short internal list of “golden test cases” (sample messages) that you reuse when you tweak rules.

In Handoff Settings, first:

• Choose the timezone for the primary team handling routed conversations.
• For each active day (for example, Monday–Friday), enable the day and set:
  • Start time when humans begin accepting routed conversations.
  • End time when handoffs to humans should stop.

If you have multiple regions, pick one timezone as the “source of truth” and document how other regions map to it.

Decide what should happen when a routing rule matches outside business hours.

Common patterns:

• AI-only after hours
  Keep AI in control and show a message like Our team is offline right now, but I can still help with common questions and log issues for follow-up.

• Limited after-hours queue
  Route only critical intents (for example, Billing for outages) to a 24/7 queue; keep everything else with AI.

Align this behavior with your SLAs so customers are never promised a human response you cannot deliver.

Use holiday overrides for days when your regular schedule does not apply:

• Company-wide holidays.
• Partial days (for example, closing at 3pm for an all‑hands meeting).
• Regional holidays if they affect key teams.

For each override, specify:

• The date.
• Whether you are closed or operating on special hours.
• What should happen to handoffs (often same as after-hours behavior).