Salesforce pays $3.6bn for the agent it couldn't build for SMBs

Salesforce announced on Sunday that it will acquire Fin, the AI customer service agent Intercom spun out last year, for $3.6 billion in cash. The press release frames the deal as accelerating Agentforce, Salesforce's existing agent platform, which Marc Benioff says has reached $1.2bn ARR (annual recurring revenue, the run-rate of subscription contracts) growing 205% year-over-year. If Agentforce is growing that fast, the interesting question is why Salesforce is paying nearly 3x its run-rate for a competing product in the same category.

What was actually filed. The Salesforce release is short on numbers and long on adjectives. Fin runs on a proprietary model called Apex, trained specifically on customer support resolutions, and the release claims a 76% autonomous resolution rate across Fin's customer base — meaning roughly three in four support tickets close without a human agent touching them. The deal closes in Salesforce's FY27 Q4, is all-cash, and does not change FY27 guidance. The release does not disclose Fin's standalone revenue, its customer count, or how the Apex model's training data pipeline transfers post-close.

That number is the asset. Everything else in the release is wrapping.

The Agentforce gap, stated plainly. Agentforce is a platform: customers configure their own agents on top of Salesforce's stack. That motion works for large enterprises with implementation budgets and internal AI teams. It does not work for the mid-market and SMB tier, where the buyer wants a product that resolves tickets on day one, not a toolkit. TechCrunch's read is that Salesforce is filling a coverage gap, and the press release confirms it in softer language: Fin brings "packaged, fast-deploy" capabilities to "SMB and commercial segments". That is the concession. A platform growing 205% can still have a hole in its addressable market, and $3.6bn is what the hole costs to fill.

What Salesforce actually bought. Read the deal as a model acquisition and it becomes legible. Apex is a vertical model — trained on a narrow, high-signal domain (years of support conversations across Intercom's customer base) rather than a general-purpose LLM pointed at the same task. Vertical models of this shape are difficult to replicate even with frontier compute, because the data is the moat and the data sits inside the incumbent. Salesforce is buying weights, and the weights encode something that Einstein, Salesforce's in-house AI layer, has not been able to reproduce on Agentforce's own data.

This is the pattern I'd flag for anyone watching enterprise AI M&A: the durable IP is moving from product features to model weights. Patents do not protect a support model; the trained artefact does. When Salesforce writes the cheque, it is paying for a file on a server, plus the data-access rights to keep that file current.

Which raises the obvious question. The press release does not address what happens to Apex's freshness when the training pipeline, Intercom's live conversation flow across its remaining customer base, is no longer owned by the same entity that owns the model. If Intercom retains the customer relationships and Salesforce takes the weights, Apex degrades from the day it closes. If the deal includes a perpetual data-licence from Intercom to Salesforce, the release does not say. This is the kind of detail that will surface in the proxy filing and is worth waiting for.

The seat-licence problem, and Salesforce's awkward position. Fin's pricing is resolution-based: customers pay per resolved ticket, not per seat. If 76% of tickets close autonomously, the economic case for staffing a human support team, at Salesforce's own Service Cloud seat-licence rates, gets harder to defend every quarter. Salesforce now owns the product that compresses its own legacy revenue line. The charitable read is that this is what disruption-from-within looks like; the less charitable read is that Service Cloud's seat-based ARR is the next thing to come under pressure, and Benioff would rather cannibalise himself than let Intercom keep doing it.

Either way, the integration question becomes: does Fin's outcome-based pricing get preserved inside Agentforce, or does Salesforce repackage it back into something seat-shaped to protect the existing revenue mix? FY27 disclosures will show the answer. I'd watch the Service Cloud ARR line specifically.

Vendor-reported numbers, handled as such. The 76% figure is Fin's. Intercom's earlier resolution-rate claims were revised downward under scrutiny, and "resolution" in support is a methodology choice — tickets auto-closed without user confirmation can be counted or not, depending on how you draw the line. The number is directionally meaningful and load-bearing for the deal logic; it is not an independently audited benchmark. Treat it as the vendor figure it is.

The standalone agent vendor, exiting quietly. Fin sold at $3.6bn. Intercom's last private valuation, in 2020, was $1.275bn for the whole business. The premium is real, and the outcome is good for Eoghan McCabe and the Intercom-Fin team. It is also a sale, not an IPO, and that distinction matters for anyone holding the thesis that vertical AI agent vendors can build durable independent businesses. The economics of enterprise distribution, long sales cycles, integration burden, security review, procurement, favour the platforms. Fin reached a credible scale and a credible model, and the exit was still into Salesforce rather than past it. The independent vertical-agent playbook in customer support now has a ceiling, and the ceiling is roughly $3.6bn.

What to watch. Three things. First, the proxy filing for the Apex data-licence terms — that determines whether Salesforce bought a model or a depreciating asset. Second, Service Cloud's seat-licence ARR in Salesforce's FY27 quarterlies — whether outcome-based Fin pricing cannibalises it, or whether Salesforce repackages Fin to protect it. Third, the next two acquisitions in this category. ServiceNow and Microsoft both have horizontal agent platforms with the same SMB-coverage problem Agentforce just admitted to. If one of them buys a vertical support-agent vendor in the next two quarters, the pattern is established and the standalone thesis is finished.

Footnotes

¶