Salesforce launches Koa, its first CRM reasoning model, claiming 3x fewer errors

LiJunnan0409 · x · 2026-09-16

Salesforce Research announced Koa, its first CRM reasoning model for Agentforce, built on NVIDIA Nemotron. It reportedly matches or exceeds leading models on CRM actions with 3x fewer errors, trained without customer data, now in pilot.

The core idea is model right-sizing: targeted jobs like intent classification, toxicity screening, and search reranking are routed to domain-specialized models, while Koa handles multi-step enterprise reasoning. A companion post explains that Agentforce previously ran on a single rented model with a token bill growing linearly with traffic (citing Uber burning its AI budget in four months and GitHub's usage-pricing backlash), arguing a single rented model is also a chokepoint for latency, roadmap, and safety logic — the real problem is architectural agency, not just token spend.

Related event: Salesforce Unveils Koa, Its First CRM Reasoning Model Built with Nvidia(2 posts)→

Original post →

More from coding & agent

coding & agent channel →