The Chatbot Problem

You have probably interacted with a support chatbot that went something like this:

You: "I was charged twice for my subscription."

Bot: "I'd be happy to help! Can you tell me more about your issue?"

You: "I was charged twice. I need a refund for the extra charge."

Bot: "I understand you're having a billing issue. Let me transfer you to a human agent."

Three messages, zero progress. The chatbot was a speed bump, not a solution. The customer is now more frustrated than when they started.

This is the typical experience with generic AI chatbots, and it is why 73% of customers say they prefer dealing with a human over a chatbot (Salesforce State of Service report).

Why Generic Chatbots Fail

Problem 1: They try to have a conversation. Most chatbots are designed as conversational agents. They ask follow-up questions, try to keep the dialogue going, and attempt to understand context through back-and-forth. But customers do not want a conversation with a bot. They want an answer.

Problem 2: They are general-purpose. A chatbot built on GPT-4 or a similar large language model knows a lot about everything and not enough about your specific product. It can generate plausible-sounding responses, but it does not know your refund policy, your shipping times, or how your billing system works.

Problem 3: They hallucinate. General LLMs sometimes make up answers that sound confident but are wrong. In a support context, this is dangerous. A chatbot that tells a customer "your refund will be processed in 24 hours" when your actual policy is 5 to 7 business days creates a support problem worse than the original one.

Problem 4: They are slow. Generating a conversational response with a large language model takes 2 to 5 seconds. For simple questions, that is slower than a pre-written response delivered instantly.

What Works Instead: Intent Classification

Instead of trying to have a conversation, classify the customer's intent and take the right action immediately.

When a customer says "I was charged twice for my subscription," a purpose-built classifier does not ask follow-up questions. It identifies the intent as billing_dispute with 94% confidence, then:

1. Sends an immediate acknowledgment: "We received your billing concern and are looking into it."
2. Creates a ticket for your billing team with the message details.
3. Notifies the right person on Slack.

Total time: 2 to 3 seconds. No back-and-forth. No hallucinated policies. No frustration.

The Technical Difference

A purpose-built intent classifier is a smaller, specialized model trained specifically on support messages. It does not generate text. It classifies.

Think of it like this: a general LLM is a doctor who knows a little about every specialty. A purpose-built classifier is a lab test that gives you a specific result. You do not need the doctor for every blood sample. You need the lab test to be accurate and fast.

Key differences:

• Training data: Hundreds of thousands of support messages, categorized into 315 specific intents
• Output: A classification label + confidence score, not generated text
• Speed: 50 to 200ms per classification vs 2 to 5 seconds for LLM generation
• Accuracy: 92% out of the box on support messages. General LLMs vary wildly.
• Hallucination risk: Zero. The model picks from a fixed set of intents. It cannot invent new ones.

When to Use Each Approach

Use intent classification when:

• You want to route or auto-respond to common questions
• Speed and accuracy matter more than conversational flow
• You want predictable, controllable behavior
• You are handling text-based support (chat, email, widget)

Use a conversational chatbot when:

• You need multi-turn dialogue (booking appointments, guided troubleshooting)
• The customer needs to provide several pieces of information
• You have a curated knowledge base the bot can search
• You are willing to invest in training and maintaining the bot

Use both together when:

• Classify intent first, then decide the action
• Simple intents get auto-resolved, complex ones get routed to humans
• The classification layer acts as a fast filter before any slower processing

The Bottom Line

The reason most AI chatbots fail is that they are solving the wrong problem. Customers do not want a chat partner. They want their question answered. Intent classification focuses on understanding what the customer wants and acting on it immediately, which is what good support actually looks like.