What happened
Intercom shipped Fin Apex 1.0, a custom-trained model that now handles approximately 100% of English chat and email customer conversations on their platform. Built by a 60-person AI team on an undisclosed open-weight base and trained on billions of proprietary customer service interactions, Apex achieves a 73.1% resolution rate compared to 71.1% for both GPT-5.4 and Claude Opus 4.5. One large gaming customer saw resolution rates jump from 68% to 75% overnight — a 22% reduction in unresolved conversations. Fin is approaching $100M ARR, resolving nearly 2M customer issues per week.
Why it matters
This is a concrete proof point for the "vertical model" thesis: a domain-specific model trained on proprietary data can beat frontier general-purpose models at a specific task while being faster and cheaper. Intercom's CEO frames this as the beginning of "the age of vertical models" where application companies with enough domain data can build their own specialized models rather than depending on frontier lab APIs. If this pattern holds, it has implications for how AI companies capture value.
Who should pay attention
- SaaS companies with large proprietary datasets considering custom model training
- AI product leaders evaluating build-vs-buy decisions for domain-specific AI
- Frontier model providers tracking vertical competition
- Customer service teams benchmarking AI agent performance