Customer support automation has a reputation problem. Most people have experienced a chatbot that couldn't understand their question, looped them in circles, and refused to connect them to a human. That's not what good support automation looks like.
Done well, AI handles the questions that have clear answers - order status, returns policy, product specs, FAQs - and routes everything else to a person immediately. Here's how to build that correctly.
Step 1: Categorize your actual support volume
Before building anything, export your last 3 months of support tickets and categorize them. Most businesses find that 60-80% of queries fall into 5-10 categories. These are your automation candidates.
The categories that automate well: order status inquiries, shipping questions, returns and refund requests, product information questions, account access issues, FAQ-type queries. Categories that don't: complaints requiring empathy and negotiation, complex multi-issue cases, anything involving legal or financial sensitivity.
Step 2: Build a knowledge base, not just a bot
The AI needs something to work from. This means structured documents: product catalog with specs and FAQs, return and shipping policies in clear language, common troubleshooting flows with outcomes. The quality of the knowledge base determines 80% of the AI's quality.
This step takes longer than people expect. Budget 2-4 weeks to document everything properly. If your policies are inconsistent or your product info is scattered across ten spreadsheets, fix that first.
Step 3: Define escalation triggers explicitly
Your automation system needs hard rules for when to hand off to a human. Don't leave this to the AI's judgment.
Escalate immediately when:
- The customer expresses frustration or anger
- The query involves a complaint about quality or safety
- The AI isn't confident in the answer (confidence threshold)
- The customer asks for a human directly
- The query doesn't match any known category after 2 attempts
- The order value is above a certain threshold (your call)
Step 4: Choose your channel and tool stack
Where does your support live? Email, live chat, WhatsApp, Instagram DMs? The answer determines the tools. For email: Intercom, Help Scout, or Freshdesk with AI integrations. For live chat: Intercom or Crisp with custom AI layer. For WhatsApp/Instagram: WhatsApp Business API or Meta's Messenger API, with Make or n8n as the automation backbone and Claude or GPT-4o as the AI.
The most flexible approach is to use Claude or GPT-4o via API as the intelligence layer and pipe it into whatever channel system you already use. This gives you full control over the prompts, the escalation logic, and the knowledge base.
Step 5: Measure the right things
Once live, the metrics that matter: containment rate (% of queries fully resolved by AI without human), escalation accuracy (are the right queries getting escalated?), customer satisfaction on AI-handled queries, and average resolution time.
Aim for 60-75% containment in the first month. Pushing for higher containment too fast usually means the AI is handling queries it shouldn't, which kills satisfaction. Increase containment by expanding the knowledge base, not by raising the escalation threshold.
What good looks like
One of our clients - an e-commerce store with a single support person - now has 80% of customer queries handled by AI. The human handles escalations, processes refunds for complex cases, and manages relationships with the top 5% of buyers. Response time dropped from hours to under 3 minutes. The support person now handles twice the ticket volume as before with less stress.
If you want us to look at your support setup and tell you what's worth automating, a free audit takes 30 minutes.
Free 30-min audit. We tell you what to automate first and what it would cost.