The $3.6B headline on the Fin acquisition is the boring number. The interesting line is buried two paragraphs down: Fin’s agent runs on Apex, a proprietary model “purpose-built for customer support” that they claim beats top frontier models on resolution rate.

That’s the actual bet, and it cuts against the prevailing wisdom. For two years the enterprise-AI default has been “wrap a frontier model, the base model keeps getting better, don’t waste capital training your own.” Buying a company whose moat is a domain-specialized model says the opposite: on a narrow, high-volume task like support deflection, a smaller model trained on the right transcripts and tools can out-resolve a general model that knows everything and nothing.

I find that credible, with one caveat — resolution rate is a slippery metric. Resolved-by-whose-judgment? End-to-end across chat, email, WhatsApp, SMS, and phone (as the release lists) is a much harder claim than resolving a tidy single-turn FAQ. Anyone who has run conversational agents in enterprise CSM deployments knows the gap between “the model returned a confident answer” and “the customer’s problem is actually gone, no human touched it, no callback next week.” That second number is the only one worth paying $3.6B for.

What this really signals: the model layer for vertical agents is now an acquisition target, not just a feature. CRM vendors are deciding the support agent is too strategic to rent.

The HN thread is mostly arguing valuation. The question I’d ask instead: if a purpose-built support model genuinely beats frontier models here, how many other enterprise workflows are about to grow their own specialized models — and how soon does “just call the biggest model” stop being the default architecture?