The $3.6 billion support agent

On June 15, Salesforce agreed to buy the AI customer-service company Fin — the agent formerly known as Intercom — for about $3.6 billion. Fin's agent runs on its own model, Apex, and resolves on average 76% of support volume end to end across chat, email, WhatsApp, SMS, phone, and Slack.

Here's the part that should make you stop. Salesforce already sells a platform for building exactly this kind of agent — Agentforce, which just hit $1.2 billion in annual recurring revenue, up 205% year over year. The company most able to build a support agent looked at the math and paid $3.6 billion to buy one instead. That decision is worth more than the press release. Let me unpack it.

Build versus buy just got a giant data point

Every team shipping AI right now is having the same quiet argument: do we build the agent ourselves on a platform, or buy a finished one that already works? It usually gets settled by gut feel and whoever's loudest in the room.

Salesforce just settled it with $3.6 billion of real money, in the one case where building should have been the obvious call. They own the platform. They have the engineers, the distribution, the data. And they still decided that a packaged agent that already resolves 76% of tickets today was worth more than the version they could build themselves over the next year. When the company best positioned to build chooses to buy, that tells you something about how hard the last mile of a working agent really is.

The model was never the hard part

Notice what Salesforce paid for. Not a smarter model in the abstract — Apex is a model purpose-built for support, not a frontier giant. They paid for the thing wrapped around the model: years of handling real tickets, the connectors into chat and email and phone, the tuning that turns "can answer questions" into "resolves 76% of them without a human." That last 76% is not a model achievement. It's a product achievement, and it took Fin fifteen years as Intercom to get there.

This is the trap in every build-it-yourself plan. The demo — wire a model to your help docs — takes a week. The remaining 90%, the part that makes it reliable enough to trust unwatched, is where the months and the money go. Salesforce just told you, in the bluntest possible terms, that buying past that 90% is worth billions.

What this means if you're not Salesforce

You don't have $3.6 billion, but you have the same decision in miniature, and the same signal applies. Before you commit to building an agent from parts, ask the question Salesforce just answered for you:

• Does a proven, packaged version already exist? If something already resolves your problem at 76% out of the box, building your own to maybe match it is a vanity project, not a strategy.
• Are you building the model, or the boring wrapper? The wrapper — connectors, edge cases, trust — is the expensive part and the part you can't skip.
• What's your real cost of waiting? Salesforce could have built this. They bought it because twelve months of building is twelve months of not having it. Your runway is shorter than theirs.

Buying isn't always right. But "we'll just build it" is a sentence that should now come with a $3.6 billion asterisk.

The bottom line

The headline is a big acquisition. The lesson underneath is about the gap between a model that can answer and a product that actually resolves — and how much that gap is worth.

The company that sells the agent-building platform just paid $3.6 billion to buy a finished agent rather than build one. If they did the build-versus-buy math and landed on buy, it's worth doing that math honestly before you assume your team should build from scratch. The model is the cheap part. The 76% is the part you pay for.