When a customer sends a message like "I never received my package and I want my money back," there are actually two things happening: a delivery complaint and a refund request. The AI needs to understand both, prioritize the right one, and route the conversation to the correct playbook. This is intent detection -- and it's the backbone of any AI-powered support system.
What Is Intent Detection?
Intent detection is the process of reading a customer's message and classifying what they want to accomplish. It's different from keyword matching. A keyword system sees "refund" and triggers a refund flow. Intent detection understands that "I'd like my money back" and "can you reverse the charge" and "this is unacceptable, I want a full reimbursement" all mean the same thing -- even though they share zero keywords.
Modern intent detection uses large language models (LLMs) like Claude to understand the meaning behind text, not just the words. This matters because customers don't write in the structured format that rules-based systems expect.
How Claude AI Classifies Messages
When a message arrives, SupportPilot sends it to Claude with a structured classification prompt. The AI evaluates the message against a predefined set of intents and returns a structured response:
- Primary intent: The main thing the customer wants (e.g.,
refund). - Secondary intent: An additional request, if present (e.g.,
complaint). - Confidence score: How certain the AI is about its classification (0-100%).
- Urgency level: Low, medium, or high, based on the tone and content.
- Sentiment: Positive, neutral, or negative.
This structured output is what drives every downstream decision. The playbook system reads the intent, the confidence score determines whether the AI acts autonomously or escalates, and the urgency level influences response priority.
The Intent Taxonomy
SupportPilot recognizes the following core intents, tuned specifically for e-commerce support:
question-- General questions about products, policies, shipping, or the store. The most common intent, usually handled by knowledge base search.wismo-- "Where is my order?" Order tracking and delivery status inquiries.refund-- Requests for refunds, partial refunds, or chargebacks.cancellation-- Requests to cancel an order before it ships.modification-- Changes to shipping address, order items, or delivery preferences.complaint-- Expressions of dissatisfaction that may or may not include a specific request.return-- Requests to return an item, often with exchange or store credit preferences.other-- Messages that don't fit any standard category.
Confidence Scores and Escalation
The confidence score is critical. It's not just about accuracy -- it's about knowing when the AI should act and when it should step back.
In SupportPilot, the confidence threshold works like this:
- High confidence (80-100%): The AI is very sure about the intent. If the playbook is in autopilot mode, it handles the conversation end to end.
- Medium confidence (50-79%): The AI has a reasonable guess but isn't certain. It drafts a response for human review (copilot mode kicks in automatically).
- Low confidence (below 50%): The AI can't reliably classify the message. It flags the ticket for human triage and provides its best guess as a suggestion.
This three-tier system prevents the most common problem with AI support: confidently wrong answers. When the AI isn't sure, it says so. That's the difference between helpful automation and a frustrating chatbot.
Multi-Intent Messages
Customers rarely ask one thing at a time. "Where is my order? Also, I need to change the address for my next subscription delivery, and your website was down yesterday." That's three intents in one message: WISMO, modification, and complaint.
The AI detects the primary intent (the most actionable one) and the secondary intent. Each triggers its respective playbook. The response addresses all parts of the message in a natural, unified reply -- not three separate answers stitched together.
How Playbooks Use Intents
Once the intent is classified, the system matches it to the appropriate playbook. The mapping works like this:
- Intent
wismoactivates the WISMO playbook, which calls the Shopify API to look up tracking information. - Intent
refundactivates the Refund playbook, which checks order eligibility and either processes or escalates. - Intent
questiontriggers a knowledge base search, pulling semantically relevant answers from your docs, policies, and FAQ. - Intent
complaintwith high urgency automatically escalates to human review, regardless of playbook mode.
This is the core loop: detect intent, match playbook, execute strategy. It happens in seconds and handles the vast majority of support conversations accurately.
Why This Approach Beats Keywords
Keyword-based systems fail on ambiguity. "I love the product but the delivery was terrible" contains both positive and negative sentiment -- a keyword system would either miss the complaint or misclassify the whole message. LLM-based intent detection understands context, sarcasm, compound requests, and the difference between "I want to cancel" and "can you tell me about your cancellation policy?"
The result is fewer misrouted tickets, fewer wrong answers, and higher customer satisfaction. The AI gets it right the first time, and when it's not sure, it asks for help instead of guessing.
Want to see intent detection in action? Try SupportPilot AI -- the Sandbox feature lets you test how the AI classifies any message you type.