insight · 4 min read
Koa: why Salesforce built its own reasoning model
What Koa is, how it was trained, and who can use it today.
The Product Wave ·
Salesforce introduced Koa on September 15, 2026 as its first reasoning model built for CRM work inside Agentforce (press release).
How it was made
Instead of training from scratch, Salesforce post-trained NVIDIA's Nemotron model on synthetic data designed to look like enterprise workflows. Salesforce says no customer data was used. The goal is an agent that can plan multi-step work and choose the right tools, not just answer questions.
The benchmark caveat
Salesforce reports strong results on its own CRM benchmark (TechCrunch coverage). It is a vendor benchmark, so treat it as a claim until independent tests appear.
Who can use it
Koa is in pilot with selected customers. Salesforce expects general availability in winter 2026, in U.S. regions only — teams with data-residency needs elsewhere should not plan around it this year.
Why it matters
Koa signals that Salesforce wants a model it controls inside its reasoning engine, alongside partner models. For buyers, the questions to ask are how model choice will be exposed in Agentforce and whether it changes cost. No pricing has been published.