Salesforce and Nvidia launch Koa, a CRM reasoning model
Koa post-trains Nvidia's Nemotron 3 Super on synthetic CRM workflows and runs inside Salesforce's own infrastructure, with customer pilots already under way at Formula 1, Xero and UChicago Medicine.
Salesforce has built its first reasoning model, and it is not aimed at the frontier. Koa, announced with Nvidia on 15 September, exists to run the multistep chores of enterprise software — updating an opportunity, routing a support case, booking the follow-up — inside Agentforce, the company's agent platform. It is the clearest signal yet that the enterprise AI market is splitting away from the general-purpose models the big labs keep racing to build.
What Koa actually is
Koa is built by post-training Nvidia's open-weight Nemotron 3 Super model rather than training from scratch. Salesforce fine-tuned it on a proprietary synthetic dataset modelled on nearly three decades of CRM deployments, covering more than 14 industries including manufacturing, financial services, healthcare and travel. Each scenario pairs a persona with a task and maps the exact sequence of actions and tool calls an agent needs to finish it. Post-training used supervised fine-tuning and reinforcement learning with Group Relative Policy Optimisation, run through Nvidia's NeMo RL, NeMo Gym and NeMo AutoModel tooling.
Salesforce is explicit that no customer data went into the model. Because Nemotron is open-weight, the company controls the weights and runs both post-training and inference inside its own trust boundary, so nothing crosses out to a third-party model provider. Nvidia's Jensen Huang framed the deal as the point of open models: “Every company needs useful AI, tailored to its knowledge, expertise, and work.”
The numbers Salesforce is quoting
Salesforce says Koa already “matches or exceeds leading model performance on CRM actions with three times fewer errors” on its own CRM Bench, a suite of real tasks drawn from sales and service work. Its product materials put the gains at 11% better action precision, about 2.1 times more reliable recall of customer context, and 15% better context retention across long conversations. Those are vendor-run benchmarks on a vendor-designed test, so they are a starting point rather than independent proof, but they describe a narrower and more measurable target than the general capability claims frontier labs usually publish.
Where you can use it
Koa is available three ways: as a managed model in Salesforce's generative AI model catalogue, as a selectable model provider across an entire Agentforce org, and inside Agentforce Builder at the agent and sub-agent level. Salesforce is already using a Koa-powered agent in Slack for internal staff queries, and says customer pilots are under way with 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero. The two companies are also extending Nemotron-based models and accelerated computing into Missionforce for government and regulated customers, including private clouds and air-gapped networks.
Our opinion
The interesting part of Koa is not the model, it is the sourcing. By post-training an open Nemotron rather than renting a frontier API, Salesforce gets a model it owns, hosts and can point at a benchmark it designed — which is exactly why enterprises that spent the last two years worrying about data leaving their perimeter should pay attention. The catch is that the performance claims only exist inside Salesforce's house benchmark, and a model tuned on CRM workflows will look sharp on CRM workflows. The real test arrives when those pilots go up against a customer's messiest case queue, and when Salesforce has to price Koa against the frontier models it is quietly arguing are the wrong shape for the job.