Salesforce has introduced Koa, an enterprise language model built by post-training Nvidia's open-weight Nemotron-3-Super-120B for tool use and CRM workflows. The launch lands inside a business that is already monetising: Salesforce said fiscal second-quarter Agentforce ARR exceeded $1.5 billion, up more than 240% year on year under the company's broadened definition, while Agentforce and Data 360 ARR reached nearly $3.9 billion.
That makes Koa an input to a paid execution layer rather than a standalone model bet. Salesforce is trying to move enterprise AI from answering questions to acting across customer records, applications and business processes while it controls the permissions, workflow context and billing surface around those actions.
The training perimeter is important. Salesforce's research paper says Koa was trained on public and synthetically generated data, with no customer data. Its CRM specialisation comes from workflow specifications, simulated tasks and reinforcement learning around successful tool use, not from ingesting clients' proprietary CRM records into the training set.
Salesforce also reported 7 billion Agentic Work Units delivered to date, including 3.2 billion in the second quarter. The usage figure does not map one-for-one to revenue, but it gives the company a consumption surface that can sit alongside its established seat-based software model.
Usage is becoming a billing surface
Agentic Work Units are an activity measure, not revenue. Salesforce's pricing model shows how activity can become revenue: customers can buy Flex Credits consumed by agent actions, pay by conversation, or use per-user and bundled editions. In the US, Salesforce lists Flex Credits at $500 per 100,000 credits, with standard Agentforce actions drawing 20 credits.
That means the 7 billion AWU headline should not be multiplied by a public list price - the units, contracts and buying models are not interchangeable. It does, however, show that Salesforce is building a metered execution layer alongside its established seat-based software business.
Koa gives the company more control over the model layer without forcing customers into a single closed foundation model. Salesforce says its benchmark gains are strongest on multi-turn tool use, while the paper also acknowledges that Koa remains below the strongest frontier models. That is an important competitive boundary: specialisation can improve enterprise task economics without eliminating dependence on external frontier models for the hardest reasoning.
Agentforce has already crossed into disclosed ARR at meaningful scale. Koa earns its place if it lowers model cost, improves task reliability or increases paid agent activity inside that revenue stream. The strategic asset is Salesforce's attempt to make model choice subordinate to the workflow, permission and billing system it controls.