AI and human handoff rules for your first 30 days

Before you start

• Your channels connected to the shared inbox and Shopify connected.
• Your policies and FAQs (shipping, returns, sizes, payment methods) ready to upload.
• A list of your most frequent questions. Your recent conversation history is the best source.
• One person who owns the rollout and has 30 minutes a week for the review.

Set it up, step by step

1. Step 1

   Before day 1: decide what the AI never handles

   Write the list before you switch anything on: refunds and money disputes, order changes, complaints, legal or health questions, anything you'd want a person to own, and any request to talk to a person. These become handoff rules, not suggestions.

   In Kanesh

   • Handoff rules: one line per situation, with the team that should pick it up.

2. Step 2

   Week 1: a narrow scope and your knowledge in place

   Upload your policies and FAQs, and start with the questions that are frequent and low-risk: order status, shipping times, returns policy, product questions. Where your team is unsure, have them work from suggested replies (approve or edit) before the AI answers on its own. Read every handoff once a day.

   In Kanesh

   • Knowledge base: policies and FAQs.
   • Suggested replies your team can approve or edit.
   • Every AI reply is checked against your data and rules before it's sent; if it can't be backed up, it goes to a person.

3. Step 3

   Week 2: fix the gaps the handoffs show

   Group last week's handoffs by reason. "No information" means a gap in your knowledge base: add the missing fact. "Asked for a person" is fine and stays. A handoff that the AI could have answered well means a rule is too broad: narrow it.

   In Kanesh

   • Tags on conversations and saved views to review handoffs by reason.

4. Step 4

   Week 3: make sure someone picks up

   A handoff is only as good as the reply that follows it. Assign each type of handoff to a team, set business hours and an SLA target with alerts, and give your team saved replies for what they repeat.

   In Kanesh

   • Assignment to people or teams, internal notes and mentions.
   • Business hours and SLA targets with alerts.
   • Saved replies and macros.

5. Step 5

   Week 4: review and decide what to widen

   Look at the four numbers below. Widen the AI's scope only where reviewed accuracy is high and customers don't come back with the same question. Keep the "never" list from day 1.

What to measure

Common mistakes

• Switching everything on at once. You can't tell what worked, and early mistakes erode your team's trust.
• Handoffs with no owner. The customer waits twice: once for the AI, once for nobody.
• Fixing single answers instead of the source. If an answer was wrong, correct the knowledge base or the rule.
• Reporting "share of conversations the AI touched". It says nothing about whether customers were helped.
• Hiding that it's an AI. Tell customers they're talking to an AI assistant and how to reach a person.