Building a customer support automation in Relay.app

A solid support automation usually follows this pattern. You can get this set up in minutes with the starter prompt template here.

1. Trigger on a new ticket or email

Start the workflow when an email hits your support inbox, a ticket is created in Zendesk or Intercom, or a customer submits a form. However you typically get reports from customers.

2. Classify and summarize with AI

Use an AI step to detect intent (bug, refund, billing, feature request), sentiment, and urgency. Pull out the customer's name, account, and a one-line summary.

3. Look up context about the customer

Query your CRM (HubSpot, Salesforce), billing tool (Stripe), or internal database to enrich the ticket with the customer's plan, lifetime value, and recent activity.

4. Route the ticket based on rules

Here's where your workflow branches: Bug reports get logged in Linear or Jira, simple FAQs get auto-drafted replies, and VIP customers go straight to a senior rep in Slack.

5. Draft a reply with AI

Use an AI step to generate a personalized response. You can reference your help center documentation, past tickets, and optionally the customer's account information for context to write a better email.

6. Get human-in-the-loop approval from a teammate before the response goes out

Send the text of the draft to an agent in Slack or email. They can approve, edit, or reject before anything reaches the customer.

7. Send the response and log the communication

Once approved, send the reply through your help desk and update the ticket status, tags, and CRM record.

Scenarios where this pays off

• Answering common questions. Auto-answer common questions like password resets, billing dates, or shipping status using your knowledge base, with a human review step for new question types.

• Urgent escalations for VIP customers. Detect angry or churn-risk language with AI sentiment analysis and ping a manager in Slack within seconds.

• Frontline bug triage. Extract reproduction steps from a ticket, create a Linear issue, and reply to the customer with a tracking link automatically.

• Speed up refund requests. Pull the customer's order from Stripe, check eligibility against your policy, and route to finance for one-click approval.

Build on the basics

Once you've got your basic customer success process set up, there is so much you can do to build on it! And you don't have to work too hard to add onto your workflow; just describe your idea to the Relay.app AI agent, and it will work with you to make it happen.

• Add knowledge and context lookups relevant to each ticket category. Since your workflow branches in a few paths, you can optimize each path for the specific query. You might find it useful to look up relevant documentation for "how do I…" questions, or search for related bugs to tag for new bug reports.

• Learn where a human-in-the-loop isn't needed. As your process matures and you optimize your prompts, you may find that select AI-generated communications can be sent automatically, without a person always checking the response. You can send a pre-formatted response when AI can confidently choose the right template, or go all in and allow AI to reply when it's highly confident in its response. (Tip: Be transparent and tell the customer the response is AI generated and to reply if it made a mistake!) Have a team member periodically review the responses in bulk (with AI to help) for a quality check.

• Use AI to decode feedback patterns. Your customer support tickets contain a real wealth of knowledge. Summarize themes across last week's tickets and post a digest to your product team in Notion or Slack every Monday.

• Rate your AI's responses to improve them over time. There are several ways you can approach this, such as: 1) maintaining a record of good responses the AI can reference, 2) asking teammates to submit feedback on the AI answer during the approval step and logging the feedback to give to the Relay.app AI agent, or 3) giving the Relay.app AI agent feedback directly when a response isn't quite right.