How Intercom went from nearly dead to a $3.6B sale as Fin

I’ve never met anyone who enjoys having to reach out to customer service. Sometimes you spend 30 minutes stuck in the phone tree, clicking different numbers to get to an actual human. Sometimes that actual human can’t help you or doesn’t want to help. And on rare occasions, you actually get an 11/10 rep who not only helps you but makes you feel like you can do anything.

Although what’s easier than any of that is if you can just chat with a chatbot that resolves your issue for you. Unless you’re me and accidentally had a ClassPass account without knowing, and the AI chatbot did nothing to help, so you were forced to escalate to a very helpful human.

Intercom is one of the OG companies that made it easier to run your customer support team until they decided to pivot entirely to a customer support AI agent company and rebrand themselves as Fin. Today, I thought it would be fun to learn more about them, especially since they just signed a deal to get acquired by Salesforce.

TL;DR

The thirty-second version of what Fin did.

You've probably used Intercom even if you've never heard the name. It's that little chat bubble in the bottom-right corner of a thousand websites, the one you poke when your order's late and you've started typing in all caps. For 14 years, Intercom sold that bubble (plus a helpdesk) to businesses, and it charged the way software always has: by the seat. Every support rep who logged in was another line on the invoice.

You can probably already spot the problem AI creates here. Intercom's customers are companies with support teams. If an AI can answer their customers' questions for them, those companies need fewer reps, fewer seats, and Intercom's business gets a little smaller every time the technology gets better. When ChatGPT showed up in late 2022, that went from a someday-problem to a right-now one. McCabe, the CEO, has since admitted they were basically dead in the water: five straight quarters of shrinking sales, growth sliding toward zero, the whole thing running out of road.

Here's what they did about it. Instead of selling tools that help reps answer tickets, Intercom built an AI agent that answers them itself. It's called Fin, and it talks directly to your customers, sorts out their problem, and takes the action a rep used to take, issuing the refund, changing the subscription, tracking down the order. And the economics are kind of staggering: a support ticket that costs a company around $22 when a human handles it, Fin resolves for 99 cents. Then Intercom went all the way and renamed the 14-year-old company from Intercom to Fin, the change-your-name-at-the-courthouse kind of commitment (which, fun fact, I have personal experience with, and can confirm is a miserable amount of paperwork even when it's just your wife and not a billion-dollar company).

The bet: when AI is going to replace your business anyway, you want to be the one who does it. And it worked. Really worked. Fin is now growing more than 300% a year, and last month Salesforce agreed to buy the company for $3.6 billion.

Under The Hood

The three strategies that make it work.

1. Bet the whole company, not a side project

Six weeks after ChatGPT came out, Intercom had a rough, barely-working version of Fin. McCabe looked at that janky demo and decided it was the entire company's future.

He moved about 80% of engineering onto Fin while it was still a blip on revenue, and put $100M of cash behind it back when "we're an AI company now" was something every struggling startup was posting on LinkedIn. He even repriced the old business on purpose, cutting what customers paid for everything outside the core and giving up about $60M in revenue Intercom was already collecting, all to push the whole customer base toward the new model. The harder changes were to the company itself. He rewrote the values and turned over about 40% of the staff (there was a near-revolt he openly calls a "soft coup"). He cleared out the experienced, play-it-safe board members and brought in startup founders. He pointed every marketing dollar at Fin. And, as you know, he changed the name.

Why does all the messy org stuff belong in a strategy piece? Because it's what makes the bet real. A chatbot bolted onto the side of your existing product is something you can shut down in a quarter, the second the board gets nervous. A renamed company with a new team, a new board, and most of engineering on the new thing is a point of no return. Most pivots die because the company keeps one foot in the old world, ready to run back the moment things get scary. Intercom made running back impossible.

To be fair, they had a head start: they were already desperate, McCabe was a brand-new CEO with permission to break things, and they had an AI team in-house before most companies did. Easier to bet the house when it's already on fire.

2. Charge for the outcome, not the activity

Intercom's old pricing was the standard software setup: pay per seat. You paid the same whether the tool solved a thousand problems or none of them. That worked fine when the software just helped humans. It stopped making sense the second the AI was doing the work, because now Intercom was charging for human seats on a product built to need fewer humans. The incentives were pointing in opposite directions.

So they tore it up and built what McCabe claims they invented: outcome-based pricing. Fin costs 99 cents when it resolves a customer's problem, and nothing when it doesn't. Outcome-based pricing ties what Intercom earns to the exact thing the customer came for: resolutions.

"Resolves" is the word that really matters here, so let me explain how Fin actually counts one. A ticket counts as resolved one of two ways: the customer says some version of "great, that fixed it," or the customer moves on and never reopens the chat, at which point Fin waits a full 24 hours before assuming it got handled. If the customer gets annoyed and asks for a human, that's an escalation, and Intercom doesn't charge for it.

So why 99 cents? A support ticket handled by a human costs a company around $22. Fin does the same job for a buck. That's a 95% discount on every ticket. A mid-sized support team handles 100,000 tickets a month easily. At $22 per human-handled ticket, that's $2.2M/month in support costs. With Fin at 99 cents, the same volume costs $99,000. That's $2.1 million per month in savings. Or roughly $25 million a year. Per customer. The kind of number that finally gets your CFO to approve the holiday party, open bar included.

But here’s one of the most important parts of outcome pricing: you have to actually make money on each resolution (or at least most of them). When Fin launched, it cost about $1.20 to resolve a ticket while they were charging 99 cents. They lost money on every single one. McCabe's bet was that the cost of running AI would keep dropping, and it did (which, btw, is a similar bet that Duolingo made). And that cost isn't even fixed: one chat is a ten-second FAQ lookup, the next chews through a dozen back-and-forths and tool calls and costs way more. So you're always managing an average and pricing accordingly.

There's a catch, and it loops right back to that word "resolved." Intercom is the one who decides what counts. That 24-hour-silence rule means you can end up paying for customers who never actually got helped; they just gave up and closed the tab. That's the core of what critics like Siena argue about paying per outcome: whoever defines the outcome holds the pen. Intercom's hedge is to keep seat pricing on the helpdesk too, so nobody's forced all the way onto one model.

3. Beat the frontier labs by going narrow

For its first few years, Fin ran on OpenAI's models and Anthropic's Claude. Intercom rented the intelligence, wrapped its own product and support data around them, and shipped. That's how a barely-working 2023 demo became a real product fast.

That changed once Fin grew from an experiment into the entire company. Intercom invested in building its own AI models. They broke customer service down into its actual steps, figuring out the customer's language, summarizing the problem, finding the right help doc, deciding whether a human is needed, writing the final answer, and building a small, specialized model for each one. Seven little models, each great at one tiny job, working together as a pipeline. While everyone else is reaching for one model to rule them all, Intercom built a relay team.

And the anchor of that lineup, the model that writes the actual answer (they call it Apex), beats frontier models like GPT-5.4 and Claude at support, with a higher resolution rate, 65% fewer made-up answers, and a fraction of the cost. A 14-year-old helpdesk company, beating the Frontier Labs at their own game.

How is that even possible? A frontier model has an impossible job; it has to be good at everything at once, like reciting poetry, coding, being my personal therapist, and answering support tickets. Intercom only has to be good at one thing, and they've got what OpenAI doesn't: years of real support conversations and people who understand support in their bones. In a narrow domain, your data and expertise beat raw scale. The general model is a mile wide. You get to be a mile deep. And that's the most hopeful thing in this whole story: you'll never out-spend OpenAI on compute, and you don't have to.

One practical note on timing, so you don't take the wrong lesson: you don't need to train your own model on day one. Rent the best one, ship, and figure out what "good" even means for your problem. Build your own later, once the AI is core enough that owning it changes your economics and you've collected enough data to make yours better.

What I’d Ship

1. LinkedIn Recruiter → pay per candidate who actually replies.

Recruiters pay LinkedIn $2,000 to $13,000 a seat per year for the right to fire off InMails, most of which get ignored. You're paying to send messages, and whether anyone writes back is your problem. So flip it: an agent finds the right people, writes the outreach, and sends it, and you only pay when a candidate actually replies and shows interest. The reply lands right in the LinkedIn inbox, so it's trivial to verify; it works the same for in-house or agency recruiters, and it's a better deal for LinkedIn, too, since the more replies they can generate, the more they can charge.

2. DocuSign → pay per signed contract, not per seat.

DocuSign charges companies per user to do something pretty mundane: shuttle a document around and collect signatures. That's a high-volume, repetitive job an agent can take over end to end, drafting from your template, redlining against your standard terms, chasing the other side until they actually sign. It’d be funny to imagine an agent who continually bothers you until you sign. So charge per executed contract instead of per seat.

Extra Credit

Pricing your AI product (YouTube)

A Lenny’s Podcast deep dive on all things related to AI monetization and pricing strategy. He defines a 2x2 quadrant for determining how you can price your AI product with the two axes:

• Attribution: This refers to the ability to measure, monitor, and clearly prove that the value a business receives is directly caused by the AI product (ex. resolved tickets)

• Autonomy: This refers to the degree to which the AI can function independently without humans in the loop

Worth the listen if you’re considering pricing for one of your products.

The history and psychology of hold music

Hold music was created by accident when a factory’s telephone system had a loose wire that picked up neighboring radio broadcasts when the caller was on hold. There’s a fun fact for you if you ever need one.

Maybe AI customer service agents will kill the need for hold music entirely, but I somehow doubt it. Music helps ease the passage of time, so as agents need to run more complex queries, I think that we might actually see agentic companies create their own versions of hold music.

This article surprised me. I thought it would be fun to write about a company that completely went AI-native (a la Allbirds, sorry “Smartbird”), but I actually found it really inspiring that they created their own models that outperform ChatGPT & Claude. It feels like the rebels in Star Wars fight against the Empire. But if they got bought by Salesforce, are they also now the Empire?

If you also enjoyed this article, share it with a colleague, friend, or even a frenemy. I’ll see you next week.

— Amaraj (aka waiting for the next available agent)

The Meme