Build a support loop customers can trust.

Customers should be able to ask for help while they are still inside the product. Keep the launcher visible, the composer focused, and the path back to earlier conversations obvious.

Ask only for information the product cannot provide safely on the customer's behalf.

A support answer improves when the teammate knows who is asking, which application they use, what plan they are on, and where the problem happened. Send that context from your server when it affects identity or permission.

Treat browser-provided page and device details as useful context, not proof of identity.

Every active conversation needs a status, an assignee, and a clear next action. Customers should immediately see the human agent's name and avatar when a handoff happens.

Inside the Inbox, separate replies from internal notes and keep the author attached to every message.

Start AI with repeatable questions that have approved answers. Retrieve only published, authorised Knowledge and require a handoff when the evidence does not cover the question.

A supervised draft is a better first milestone than an autonomous agent with unclear boundaries.

Track conversation volume, first-response time, current status, application source, AI usage, and whether suggestions were accepted, edited, or rejected. The purpose is to find weak content and slow handoffs, not to reward rushed replies.

Review the questions that repeatedly need a human and turn the stable answers into maintained Knowledge.