siliconangle.com web signal

Salesforce Unveils Koa, a CRM Reasoning Model on Nvidia Nemotron

TL;DR

  • Salesforce's new Koa reasoning model is built on Nvidia's Nemotron architecture and targets multistep CRM workflows inside Agentforce.
  • Salesforce claims Koa matches or beats leading general-purpose models on CRM tasks 'with three times fewer errors,' though it names no benchmark.
  • Training used only synthetic data simulating 'almost 30 years' worth of internal CRM deployment experiences' across 14 industries, no customer data.

Salesforce unveiled Koa, a reasoning model tuned for customer-relationship-management data, and says it matches or exceeds leading general-purpose models on CRM tasks "with three times fewer errors." As SiliconANGLE reports, Koa is built on Nvidia's Nemotron architecture and is designed to power multistep Agentforce workflows — routing support cases, updating sales opportunities, qualifying leads. Salesforce does not name a benchmark or a comparator model behind the three-times figure.

Training used synthetic data rather than customer records. Salesforce simulated "almost 30 years' worth of internal CRM deployment experiences" across 14 industries including healthcare, finance and manufacturing. "There is so much knowledge that we've sort of gathered in building out the CRM over the last 27 years," Chief Platform and Engineering Officer Rohan Kumar told the outlet. The pipeline combined supervised fine-tuning with reinforcement learning and a method called "group relative policy optimization," aimed at multistep tool execution.

The launch customer voice comes from 1-800Accountant. "Koa gives our agents the reasoning to work through that complexity step by step and use the right tools," said Chief Strategy Officer Ryan Teeples.

Koa is expanding through pilot now, with general availability set for winter. It lands inside a crowded stretch of enterprise-AI announcements we've been tracking and continues Nvidia's push to seed Nemotron into partner stacks.