A model for the job, not the whole world
Salesforce unveiled Koa at Dreamforce with Nvidia. Built by post-training Nemotron 3 Super, it is designed for multistep CRM work such as updating opportunities, routing service cases and scheduling follow-ups. Salesforce says the synthetic training set reflects 27 years of product knowledge across more than 14 industries and does not contain customer data.
Why This Matters
Enterprises do not pay for eloquence; they pay for reliable actions inside permissions, records and processes. Koa challenges the assumption that the largest general model should sit behind every agent. If a smaller domain model makes fewer workflow errors, keeps inference inside a trust boundary and costs less, it can displace frontier APIs without matching their broad knowledge.
Reading the benchmark
Salesforce says Koa matches or exceeds leading models on its CRM benchmark with three times fewer errors. That is an important claim, not an independent verdict. Buyers need task definitions, baseline models, failure severity and real deployment results. One wrong refund or permission change can outweigh dozens of correctly drafted follow-ups.
Winners, losers and competitive pressure
Salesforce gains control over weights, economics and product differentiation. Nvidia extends Nemotron into a major application stack. Customers gain model choice and potentially better governance. OpenAI and Anthropic face a market where platforms use them for general work but substitute proprietary models for high-volume tasks. Consultants may gain integration work; routine CRM labor faces pressure.
What happens next
Koa is entering pilots, with broader U.S. availability expected in winter 2026. The decisive test is not a demo but audited performance over long-running workflows, including escalation when confidence falls. A successful rollout makes domain models a standard enterprise layer. A weak one reinforces the value of frontier models plus retrieval and strict tooling.
Sources: CIO. Facts and figures are a fixed September 19, 2026 reporting snapshot and do not update live.