Salesforce, in a company post, introduced Koa, described as its first AI reasoning model optimized for enterprise work and built to run inside the Salesforce trust boundary. Koa is a post-trained version of NVIDIA's Nemotron 3 Super, a 120-billion-parameter open model; Salesforce says the open weights and published training data let it verify what went into the model rather than take a vendor's word for it.
Salesforce's argument is that a general frontier model applies the same broad reasoning to a refund policy, a physics problem and a travel plan, working each from first principles and sometimes landing on a different answer to the same repeated task. Koa was instead trained through reinforcement learning on simulated enterprise workflows spanning 14+ industries — manufacturing, financial services, healthcare, travel — with synthetic customers exhibiting different moods and a judge scoring whether the issue was actually resolved. None of the training used real customer data.
One trained behavior singled out: when the correct tool isn't available for a request, Koa is meant to stop, tell the customer what it can't do, ask for what it needs, and hand off to a human rather than call a similar-sounding tool or confirm an action that never happened.
Koa sits alongside other Salesforce-built specialized models — HyperClassifier, TextEval, Moirai — with Koa handling multi-step reasoning while those handle narrower jobs like intent classification and search reranking.
Koa is already running in Salesforce's internal employee workflows, and the company says it is now moving into customer pilots in service, sales and commerce, including regulated industries such as financial services, healthcare, travel and accounting. Salesforce reports early benchmarking on CRM-specific tasks — updating an opportunity, routing a case, scheduling a follow-up — outperforming general-purpose models, though this is Salesforce's own early benchmarking and not yet independently verified. The company also states that customer data and interaction traces never leave the customer's control and are not used to further train the model.
