Two Philosophies

Intercom's Fin and intent classifiers solve the same problem (automate support) with completely different approaches. Understanding the difference helps you pick the right one — or avoid paying for capabilities you don't need.

How Fin Works

Fin is Intercom's AI agent. When a customer asks a question, Fin:

1. Reads the customer's message
2. Searches your help center (knowledge base) for relevant articles
3. Uses a large language model to generate a natural-language answer based on those articles
4. Delivers the answer in a conversational format
5. Handles follow-up questions in a multi-turn conversation

If Fin can't find an answer or isn't confident, it hands off to a human agent.

This is called retrieval-augmented generation (RAG). Fin retrieves information from your docs and uses an LLM to generate a response.

How Intent Classification Works

An intent classifier takes a different approach:

1. Customer sends a message
2. A trained ML model classifies the message into one of 315 pre-defined intents (like "refund_request," "password_reset," "billing_inquiry")
3. Routing rules fire based on the intent: send a template response, create a ticket, notify Slack, process an action, or escalate
4. The response is a pre-written template you control, not AI-generated text

No knowledge base needed. No LLM generating answers. Just fast, accurate classification and rule-based actions.

Where Fin Wins

Complex, varied questions. If your customers ask questions with lots of variations and your help center has the answers, Fin shines. It can synthesize information across multiple articles and present a natural-language answer. A classifier would need a specific intent for each variation.

Knowledge-heavy products. Developer tools, complex SaaS platforms, products with hundreds of features — these benefit from Fin's ability to pull from extensive documentation and explain things conversationally.

Multi-turn conversations. When the answer depends on follow-up questions ("Which plan are you on?" "When did you first see this?"), Fin handles that naturally. Classifiers work best on single-message interactions.

Where Classification Wins

Predictable question types. If 70% of your support is billing, passwords, refunds, and order tracking, classification handles these faster and cheaper. You don't need an LLM to tell someone their tracking link.

Response control. With classification, you write every response. Legal can review them. You know exactly what customers see. With Fin, the LLM generates responses, and while Fin is generally reliable, it can occasionally rephrase your policies in ways that are inaccurate or misleading.

Speed. Classification takes 100 to 200 milliseconds. Fin takes 2 to 5 seconds (LLM inference + knowledge base search). For high-volume operations where response speed affects customer satisfaction, that gap matters.

Cost. Fin costs $0.99 per resolution. Classification costs $0.20 to $0.30. At 500 resolutions/month, that's $495 vs $125. At 1,000, it's $990 vs $250.

No knowledge base dependency. Fin only works well if your help center is comprehensive and up-to-date. If your docs are sparse, Fin gives bad answers or escalates everything. You're only charged $0.99 when Fin counts it as a resolution, but poor answers that technically "resolve" without a human handoff still cost you. Classification doesn't need docs at all — it routes based on intent, not content.

The Knowledge Base Tax

Here's something nobody talks about: maintaining a knowledge base good enough for Fin to use costs real time and money.

You need:

• 50 to 200+ well-written help articles covering every common question
• Regular updates when your product changes
• Consistent formatting so the search works well
• Gap analysis to find what's missing

For a team that already has great docs, this is free. For everyone else, it's 20 to 40 hours of initial work plus ongoing maintenance. And if your docs fall behind your product, Fin starts giving outdated answers.

Classification doesn't have this dependency. The model is pre-trained on 315 support intents. You set up routing rules and it works day one.

Accuracy Comparison

Fin's accuracy depends entirely on your knowledge base. With excellent docs: 80 to 90% resolution rate. With mediocre docs: 40 to 60%. With sparse docs: barely better than a search bar.

A purpose-built classifier hits 92% accuracy out of the box because it's trained specifically on support intents. It doesn't depend on your content — it depends on its own training data.

The caveat: Fin handles long-tail questions better. If a customer asks something truly unusual, Fin might find the answer in a corner of your docs. A classifier would route it to a human because it doesn't match a known intent.

Can You Use Both?

Yes, and some teams do. Use classification as the first pass — fast, cheap, handles the 70% of predictable questions. For the 30% that need more nuance, escalate to a Fin-style system or a human.

This gives you the cost efficiency of classification for the bulk of volume and the flexibility of LLM-powered answers for the edge cases.

Whether that complexity is worth it depends on your volume. Under 500 messages/month, just pick one approach. Over 1,000, the hybrid model starts making financial sense.

Decision Framework

Choose Fin if:

• You have a comprehensive, well-maintained knowledge base (100+ articles)
• Your questions are varied and hard to predict
• You need multi-turn conversations handled by AI
• Per-resolution cost isn't your primary concern

Choose intent classification if:

• Your support questions are predictable (60%+ fall into known categories)
• You want control over every response
• You don't have a knowledge base (or don't want to maintain one)
• Cost matters — especially at volume

Choose both if:

• You have very high volume (1,000+/month)
• You want to optimize cost while maintaining quality for edge cases
• You have the engineering resources to manage two systems