Salesforce and NVIDIA on September 15, 2026 announced Koa, a CRM reasoning model for Salesforce’s Agentforce platform built by post-training NVIDIA’s Nemotron 3 Super model on a proprietary synthetic dataset modeled on enterprise knowledge from nearly three decades of CRM deployments.
Salesforce described Koa as its first CRM reasoning model, purpose-built to help agents reason through complex, multistep workflows and use the right tools to complete work. In the companies’ joint announcement, Salesforce Chair and CEO Marc Benioff said: “The most valuable thing Salesforce has built isn’t our platform — it’s the accumulated knowledge of how enterprise business actually works. With Koa, the knowledge is put inside the model itself.” Jensen Huang, NVIDIA’s founder and CEO, said the Nemotron open models gave Salesforce the basis for a CRM model that can reason and act securely.
Salesforce said it controls the Koa model weights and performs post-training and inference entirely within its own trust boundary, so that no customer data crosses that boundary during inference. Koa gives customers a specialized, Salesforce-hosted option for powering Agentforce use cases.
Synthetic Training Data and Post-Training Method
No customer data was used to train Koa, according to Salesforce. The training corpus was built entirely from synthetic scenarios covering the reasoning, tool use, and decision-making skills Agentforce agents apply across CRM workflows such as generating leads, qualifying opportunities, and resolving service cases.
The scenarios were designed to simulate real-world enterprise workflows across more than 14 industries, among them manufacturing, financial services, healthcare, and travel. In each scenario, a persona was paired with a set of tasks, and the sequence of actions and tool calls required to carry them out was mapped in advance.
For post-training, Salesforce applied Supervised Fine-Tuning and reinforcement learning with Group Relative Policy Optimization (GRPO), using NVIDIA’s NeMo RL, NeMo Gym, and NeMo AutoModel tooling. Salesforce said that training on a targeted set of prioritized enterprise tasks gave the model deeper expertise, teaching it to reach a goal through a sequence of correct actions rather than stopping at a correct answer.
Research Paper and Reported Results
In a research paper submitted to arXiv on September 14, 2026, the model’s developers describe Koa as the result of applying GRPO reinforcement learning to the open-weight Nemotron-3-Super-120B foundation model, using public and synthetically generated data and no customer data. The authors identify as the model’s distinctive component a simulation-to-reward pipeline that turns workflow specifications into persona-conditioned, multi-turn tasks, with rewards tied to resolving each task through successful tool use for data-dependent requests. For enterprise domains, those specifications are written in Agent Script, Salesforce’s declarative language for building Agentforce agents; for public tool-use domains, the workflow structure was synthesized directly, with the same simulation and grounded-reward machinery driving GRPO across both.
The authors report that across public tool-use, agentic-reasoning, and enterprise CRM benchmarks, Koa improves on its open-weight base, with its clearest gains on multi-turn tool use, surpasses a strong proprietary baseline, and remains below the strongest frontier models. They conclude that specification-driven reinforcement learning offers a practical route to specializing open-weight foundation models for enterprise agentic work.
Separately, Salesforce said that in its CRM benchmark, a suite of real-world tasks such as updating an opportunity, routing a case, or scheduling a follow-up, Koa matches or exceeds leading model performance on CRM actions with three times fewer errors.
Pilots, Missionforce, and Availability
Koa is already in use inside Salesforce, including in a Slack agent that helps employees find information and complete everyday tasks. Customer pilots are under way with 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero.
Ryan Teeples, chief strategy officer at 1-800Accountant, said Koa lets the firm’s agents reason step by step through tax rules, financial data, and customer documents. John Sahagian, senior vice president and chief data officer at Baxter Credit Union, said the model can help digital agents weigh the information, tools, and policies behind member goals such as buying a home. Elia Wallen, founder and CEO of Engine, said business travel involves countless moving parts and that his teams need a model able to reason precisely through multi-step problems rather than one that simply sounds confident. Andrew Chang, chief marketing officer at UChicago Medicine, said Koa can handle longer, multi-step coordination workflows so teams have more time for patient care.
Alongside Koa, the two companies are bringing Nemotron-based models and accelerated computing into Missionforce for government and regulated organizations. According to the announcement, many of these organizations must control the model and training data and deploy in specialized environments — private clouds, classified networks, and fully air-gapped systems — that never touch public infrastructure.
Post-trained NVIDIA models will power agents for Missionforce Operations, a product that digitizes and automates government workflows including procurement, supplier management, and logistics. The models are trained on each organization’s operational data and terminology so agents can work through back-office processes and take action inside the customer’s own environment, including air-gapped networks. The companies said they are also working to bring post-trained NVIDIA models into additional Missionforce capabilities, which would let customers build and run models on their own infrastructure for specialized workloads.
Koa is available to select pilot customers now in Agentforce, with general availability expected in winter 2026 in U.S. regions. Missionforce Operations is generally available now in U.S. regions, and the post-trained NVIDIA models will be available to select customers in October 2026.
