Overview
- Salesforce announced Koa on Tuesday as its first CRM reasoning model, post‑trained from NVIDIA’s open‑weight Nemotron to handle multi‑turn, tool‑using CRM tasks.
- The company says Koa was trained only on public and synthetically generated scenarios using supervised fine‑tuning and reinforcement learning with Group Relative Policy Optimization and a simulation‑to‑reward pipeline.
- Koa will be delivered inside Salesforce’s Agentforce/AIforce stack and runs alongside integrations such as Claudeforce, a public beta in which about 40 customers are testing 37 prebuilt sales skills.
- Salesforce reports Koa is already running internally, will enter customer pilots in October, and is targeted for general availability in U.S. regions in Winter 2026.
- The launch pairs model work with an Enterprise AI Harness that enforces identity, policy, data access and observability for agents, and Salesforce’s performance claims remain company‑reported pending independent validation as pilots scale.