Intercom Fin Tutorial

Step 1: Activate Fin and Connect Your Knowledge

First, log into your Intercom workspace. Navigate to 'Settings' (the gear icon), then find 'AI Agent' or 'Fin' in the products menu. Click 'Get Started'. This is where the real work begins. You'll be prompted to connect your knowledge sources. I tested this with a messy Google Drive folder and a public help center URL. What surprised me was how well Fin parsed the disparate documents. You'll see options to add URLs (your public help center), upload files (PDFs, .docx), or connect via API. My recommendation? Start with your top 5-10 most critical help articles. Don't dump everything in at once. After connecting sources, you must define Fin's 'Resolution Confidence' threshold. This is crucial. I set mine to 'High' initially, meaning Fin only auto-resolves tickets where it's over 90% confident. This prevents embarrassing wrong answers while it learns.

Step 2: Navigate and Master the Fin Dashboard

Once activated, you'll spend most of your time in the 'Fin' section of your Intercom inbox. The key areas are: 1) The 'Fin Overview' dashboard, showing resolutions, deflection rate, and cost savings. 2) The 'Conversations' view filtered to 'Handled by Fin'. This is your audit log. Click any conversation to see exactly what Fin said. 3) The 'Knowledge' tab, where you manage connected sources and see which articles are most used. 4) The 'Settings' area for fine-tuning. In my experience, the 'Conversations' log is your most powerful tool. I review it daily for the first week. You can see where Fin succeeded and, more importantly, where it failed or handed off to a human. This direct feedback loop is what makes Fin improve. Don't just glance at the high-level metrics; dig into the actual dialogues.

Step 3: Configure Your First AI Resolution Workflow

Fin doesn't just answer; it can fully resolve and close tickets. To set this up, go to Fin Settings > Resolution Settings. Here, you'll configure the core automation. First, decide if Fin can close tickets automatically or just suggest a resolution for human approval. I was initially skeptical, but after testing, I let it auto-close for high-confidence answers. You must define the closing message. I use something like: "I've found the answer in our help center here: [Link]. I've gone ahead and resolved this ticket for you. If you need more help, just reply to reopen." This transparency is key. Next, set up the handoff rules. This is non-negotiable. Configure triggers for handoff to a human agent: if a customer says "speak to a person," if Fin's confidence is low, or if the topic is tagged as 'billing' or 'complaint'. My stance: be overly cautious here at first.

Step 4: Train and Correct Fin in Real-Time

Fin's launch is just the beginning. Your job is to train it. When you see a conversation in the 'Handled by Fin' view where the answer was wrong or suboptimal, click into it. You'll see a button to 'Provide feedback'. Use this relentlessly. You can thumbs down an answer, and more importantly, you can provide the correct answer. This feedback is fed directly back into the model. What surprised me was how quickly it learned. I corrected Fin on a specific pricing nuance on Monday, and by Wednesday, it was answering that question perfectly. Also, regularly check the 'Knowledge' tab's 'Performance' section. It shows which articles are being used and which aren't. If a key article isn't being referenced, its title or content might not be clear to Fin's language model. Rewrite it with simpler, direct question-and-answer formatting.

Step 5: Analyze Performance and Calculate ROI

The 'Overview' dashboard is your report card. Key metrics: 'Resolution Rate' (what % of conversations Fin fully resolved), 'Deflection Rate' (what % never touched a human), and 'Median First Response Time' (aim for under 10 seconds). My honest opinion? Don't obsess over 100% resolution. A 40-50% auto-resolution rate in the first month is fantastic. That's 40-50% of tickets your team never saw. Click on the 'Savings' widget. This translates Fin's activity into estimated agent hours saved and cost savings based on the $0.99/resolution fee. In my testing, the cost per resolution was consistently lower than the cost of human agent time, but you must verify this for your volume. Export this data weekly to share with leadership. This tangible ROI is what secures budget for continued use.

Step 6: Scale and Integrate with Advanced Features

Once Fin is stable, explore advanced features. First, set up 'Article Suggestions' in the Fin Settings. This allows Fin to recommend specific help articles to human agents during live chats, making your whole team faster. Second, integrate Fin with your product using the 'Product Tours' feature. Fin can trigger interactive guides when a user asks how to use a feature. Third, use the Resolution API to pull Fin's data into your internal systems for custom reporting. My final recommendation: experiment cautiously with lowering the 'Resolution Confidence' threshold by 5% increments. This allows Fin to tackle more ambiguous questions. I was surprised that with good foundational knowledge, a lower threshold (like 80%) didn't increase errors significantly, but it boosted auto-resolution by 15%. Always A/B test these changes.