Bad AI support is worse than no AI support. We've all hit the bot that can't answer anything real, loops you back to the FAQ, and makes you hunt for the 'talk to a human' button. Done wrong, it destroys trust. Done right, it handles 70-80% of tickets in seconds and routes the rest to humans who can actually help.
Here's how we build the version that works.
The architecture that actually works
The system has three parts: classification, resolution, and escalation. Every ticket gets classified first - what type of query is this, how urgent is it, can it be resolved from existing knowledge? If yes, the AI drafts a response using your actual product data (not generic training). If no, it routes to the right human with context already written up. The human never starts from scratch.
RAG: why your AI needs your actual data
A general AI model doesn't know your return policy, your specific product specs, or the status of order #4521. RAG (Retrieval-Augmented Generation) fixes this by storing your docs in a vector database and pulling the relevant ones at query time. When a customer asks 'what's your return policy?', the system finds the policy doc, passes it to Claude, and Claude answers based on your actual policy - not a guess.
What to put in your knowledge base:
- FAQ documents (the 50 questions you answer every week)
- Product documentation and specs
- Return/refund policy
- Shipping information
- Troubleshooting guides
- Pricing and plan details
- Past resolved tickets (anonymized) for edge cases
The handoff to humans
The classification step determines which tickets escalate. High-urgency tickets (angry customer, billing dispute, legal language) skip the AI draft entirely and go straight to a human with an AI-generated summary. Complex technical issues that the knowledge base doesn't cover escalate with context: 'Customer reports X, searched knowledge base, found no relevant articles, suggesting escalation to [team name].' The human who gets it knows immediately what they're dealing with.
The draft review workflow
For mid-complexity tickets, the AI drafts but a human reviews before sending. The interface we typically build: agent sees the ticket on the left, AI draft on the right. They read it (10 seconds), edit if needed, click send. This is how you get 70% automation without 70% of your customers getting a wrong answer. The 30% human review catches edge cases.
What the build looks like
We typically build this with: Claude for classification and response generation, Pinecone or pgvector for the knowledge base, Make for the workflow orchestration (connecting your helpdesk to the AI pipeline), and your existing support tool (Intercom, Freshdesk, Zendesk) as the interface agents use. Build time is typically 2-3 weeks. Cost: $2,500-5,000 depending on the complexity of your product and the size of your knowledge base.
Ongoing maintenance
The knowledge base needs updating when your product changes. We build a simple flow: when a ticket is resolved with an answer that isn't in the knowledge base, it gets flagged for review and potential addition. The system improves over time without requiring constant manual work.
Running a support team and drowning in tickets? Book a 30-min call to map your specific support workflow and see what's automatable.
Free 30-min audit. We tell you what to automate first and what it would cost.