Teardown: Intercom Fin AI

How a legacy SaaS company bet the business on AI support — and what every PM can learn from the transformation.

CheatsheetLayout, Section, NumberedItem, Callout, CodeBlock, } from "@/components/CheatsheetLayout";

The setup: a company at $0 net new ARR

In early 2023, Intercom was in trouble. A decade-old customer communication platform with 30,000 paying customers, flat growth, and a product increasingly at odds with where the market was going. Then GPT-3.5 launched. Eoghan McCabe, the founder, realized in 6 weeks what it meant: "We had this giant base of 30,000 paying customers, hundreds of thousands of active users, millions of their users, billions of data points. So we had a lot to play with. The question was whether we'd move fast enough." Six weeks after GPT-3.5, Intercom shipped a beta of Fin — an AI agent that could autonomously resolve customer support tickets. $12M ARR → $80M+ ARR. 300%+ year-over-year growth. #1 on G2 in customer support. On track to cross $100M ARR in under 3 years from launch.

The 3 product decisions that made Fin work

Scope to safe categories first, expand later

Fin didn't launch trying to resolve every ticket. It launched scoped to the high-confidence resolution zone — FAQs, known issues, documentation lookups. Categories where a wrong answer is annoying but not catastrophic. This let Intercom ship fast, measure resolution rate accurately, and build trust with customers before expanding to harder categories. The product didn't ask for trust it hadn't earned.

Treat the AI as a teammate, not a tool

The original Intercom product was anti-bot. Their entire brand was built on "make internet business personal." A support bot was a direct contradiction of their mission. Eoghan's reframe — and the product decision that followed — was crucial: "Providing a customer with a highly engaged, instantly available expert available 24/7 is so much more personal than making them wait 2-3-4 days for a crappy canned response." This reframe changed how they positioned Fin. Not "a bot that handles tickets." A "Fin AI agent" that works alongside human support teams. The distinction changed how customers adopted it.

Use the existing customer base as both distribution and dataset

Intercom didn't start from zero. They had 30,000 customers, billions of support interactions, and years of human-labeled data on what good support looked like. That existing data became Fin's training advantage. New entrants had to build from scratch. Intercom could fine-tune on real customer support patterns from day one. This is the moat that matters in AI products — not the model, but the proprietary feedback loop that improves the model faster than anyone else can.

The transformation story: what the playbook actually looked like

Cultural reset before product reset

Eoghan rewrote Intercom's company values during the pivot — deliberately designed to "cut out the parts of the company that wouldn't be effective." 40% employee turnover followed. This is the part of transformation stories that gets left out. The product change was fast — 6 weeks to beta. The organizational change was harder and more painful.

2x productivity as a measurable goal, not a slogan

Darragh Curran (CTO) set a concrete goal: 2x engineering productivity with AI. He then embedded with individual teams for a week each month to find where AI could actually deliver that — not in aggregate, but in specific workflows. They tracked merged PRs as a proxy metric for AI-assisted code velocity. Not sentiment. Not self-reported time savings. Actual output.

Permission over process

From Intercom's internal AI adoption playbook: "Tell people they can do things. If anything goes wrong, blame me." Engineers didn't need more AI training. They needed permission to experiment without fear of being blamed for failed attempts. The biggest blocker to AI adoption inside organizations is organizational risk aversion, not technical capability.

The agent-friendly SaaS lesson

One of the most underreported stories from Intercom's transformation is what they learned about building SaaS products for an agentic world. From Brian Scanlan (Senior Principal Engineer at Intercom): "Agents will optimize around your UI/UX. If your product's conversion funnel isn't agent-friendly, they'll brute-force it — signing up via web scraping, clicking through flows, burning tokens. The switching cost for agents is literally one keystroke."

What Intercom built to fix this

A CLI that agents can use to: sign up autonomously, verify email addresses programmatically, complete installation without human intervention. No scraping. No flow-clicking. Agent-native pathways.

The implication for every SaaS PM

If agents can't easily use your product, they'll route around it. Your next competitive moat is whether your product is agent-friendly — CLIs, MCPs, ephemeral APIs, programmatic access. This is infrastructure work, not feature work, and it needs to start now. The next "mobile-first" moment is "agent-first." Products that don't offer programmatic, agent-friendly pathways will lose the same way web-only products lost when mobile adoption hit critical mass.

5 things every PM can take from Fin

Your existing data is your moat — if you use it

Intercom won not because they had a better model than OpenAI. They won because they had 30,000 customers' worth of support interactions to train on. What proprietary feedback data does your product already have?

Reframe the contradiction, don't ignore it

Intercom's brand was "anti-bot." They didn't paper over that contradiction — they resolved it with a better story. If AI creates a tension with your existing product promise, the answer is a reframe, not silence.

Scope narrows your risk, not your ambition

Starting with "safe categories only" didn't limit Fin. It gave them a path to earn trust, prove the model, and expand. Don't confuse "scoped MVP" with "small vision."

Post-processing filters matter more than model quality

From Paul Adams (CPO): data privacy, contractual guarantees, and human impact metrics all matter more than raw model capability in enterprise sales. The model is table stakes. The safety layer is the differentiator.

The wartime transformation has a talent cost

40% turnover during a pivot is not a failure. It's an expected consequence of changing what the company values. Build this into your transformation plan. The cost of not changing is higher. Sources: Lenny's Podcast — "How Eoghan McCabe rebuilt Intercom around AI" (Eoghan McCabe), "How Intercom 2x'd their engineering velocity" (Brian Scanlan). Lenny's Newsletter — "Counterintuitive advice for building AI products" (Paul Adams and others), "25 proven tactics to accelerate AI adoption" (multiple operators). );