87% self-serve, 95% accuracy: how two companies actually got there

The specific architectural choice that separates AI support tools that hold up in production from ones that only look good in a demo

Why Intercom’s own engineers swapped a more powerful model for a cheaper one, and what that reveals about evaluating any AI vendor’s model claims

Why vendor-reported “resolution rates” in this category tend to run well above what independent, production-level testing finds, and how to define the metric for yourself

The scoping and ramp-time pattern behind two documented deployments, Synthesia and Fundrise, that hit real numbers within three to six months

A five-step framework for evaluating or building an AI support agent that won’t fall apart under real ticket volume

Intercom’s Fin AI Agent resolves support conversations using a multi-step pipeline, not a chatbot connected directly to a language model.

Two documented customer deployments show real, specific outcomes: Synthesia resolved over 6,000 conversations and saved over 1,300 hours in six months; Fundrise automated more than half its support volume in three months at 95% accuracy.