On September 15 (US time), Salesforce used its annual Dreamforce 2026 conference in San Francisco to announce Koa, the company's first CRM reasoning model, developed jointly with NVIDIA. Koa is built on NVIDIA's open model Nemotron 3 Super and post-trained on synthetic data that mimics sales and customer-support work, and it runs inside Salesforce's AI agent platform, Agentforce. Pilot deployments with select customers are already under way, and general availability in US regions is planned for winter 2026.
What Koa Is Designed to Do
Koa handles the kind of multistep chores that fill a CRM workday: updating an opportunity, routing a support case, scheduling a follow-up. Rather than simply returning an answer, it is specialized in agent-style reasoning, meaning it lays out the steps needed to reach a goal and calls the right tools in the right order to finish the job.
According to Salesforce's internal CRM benchmark, Koa matches or exceeds leading general-purpose models on these actions while making one-third as many errors. That figure comes from an evaluation Salesforce built itself and has not been independently verified, so it deserves some caution. Still, framing the gap in terms of fewer mistakes rather than raw accuracy signals how squarely the model is aimed at business use.
Rohan Kumar, Salesforce's president of platform and engineering, who presented the announcement, stressed that "not a single byte" of customer data was used in training. Chair and CEO Marc Benioff added that Salesforce's most valuable asset is not its platform but its accumulated knowledge of how enterprise business actually works, and that Koa puts that knowledge inside the model itself.
Turning 27 Years of Know-How Into Synthetic Data
How do you teach a model about business operations without touching customer data? Salesforce says the training corpus consists entirely of synthetic scenarios. Drawing on roughly 27 years of CRM deployments and the workflows and internal policies learned along the way, the company built scenarios across more than 14 industries, including manufacturing, financial services, healthcare, and travel, each pairing a persona with specific tasks and mapping which tools are used in what order. The subject matter spans the customer lifecycle: generating leads, qualifying opportunities, and resolving service cases.
Post-training combined supervised fine-tuning (SFT) with reinforcement learning using GRPO (Group Relative Policy Optimization), built on NVIDIA's NeMo RL, NeMo Gym, and NeMo AutoModel. By training on a targeted set of prioritized enterprise tasks, the companies say, the model learned not only to produce the right answer but to take the right action, step by step, until the goal is reached.
The fact that Nemotron is an open model with published weights underpins the whole effort. Salesforce holds the model weights itself and performs both post-training and inference entirely on its own infrastructure, inside what it calls its trust boundary. Instead of calling an external AI provider's API, the company guarantees that no customer data leaves during inference by hosting the model itself.
Early Adopters and Timeline
Koa is already in use internally, powering an agent in Slack that helps employees find information and complete everyday tasks. Six customers were named as pilot participants: accounting service 1-800Accountant, credit union Baxter Credit Union, business-travel platform Engine, motorsport organization Formula 1, healthcare provider UChicago Medicine, and accounting software maker Xero. According to NVIDIA, customer pilots will ramp up in October, with Koa offered as a customer-selectable model in Agentforce.
General availability is slated for winter 2026, starting in US regions. No timing has been given for Japan.
Nemotron Also Comes to "Missionforce" for Government
The partnership goes beyond Koa. Nemotron-based models and NVIDIA accelerated computing are also being brought into Missionforce, Salesforce's offering for government and highly regulated industries. The goal is to let organizations run models tuned on their own data and terminology in private clouds and in air-gapped environments that are physically cut off from outside networks.
Specifically, post-trained NVIDIA models will power agents in Missionforce Operations, a product that digitizes and automates government back-office workflows such as procurement, supplier management, and logistics. Missionforce Operations itself is already generally available in US regions, and the post-trained NVIDIA models are due to reach select customers in October 2026.
What Jensen Huang Said on Stage
To mark the announcement, NVIDIA founder and CEO Jensen Huang joined Benioff for the keynote. Huang framed the moment in terms of earlier revolutions: electricity let us power everything, the internet let us find anything, and now AI as a planetary infrastructure layer lets us know everything and do anything.
He also noted that open models have gone from roughly 30 percent of usage at the beginning of last year to close to 70 percent today, and predicted that every software company will become an AI company and every enterprise an agentic one. On safety, Huang called it "job one" but described it fundamentally as an engineering problem: if you are not confident in a product's functionality or safety, do not ship it. Speed and safe products, he argued, are not a trade-off.
Summary
Koa is a CRM-specific reasoning model born from Salesforce converting its own operational knowledge into synthetic data and post-training NVIDIA's open Nemotron 3 Super on it. It is distinguished by keeping everything from weight ownership to inference on Salesforce's own infrastructure and by using no customer data in training, and Salesforce's internal benchmark puts its error rate at one-third that of leading models. Six companies are already piloting it, with general availability starting in the US in winter 2026. Coming just weeks after Salesforce unveiled Claudeforce, which embeds Anthropic's Claude deeply into its products, the move shows the company lining up frontier external models alongside a self-hosted specialist and letting customers choose. How enterprises balance general-purpose intelligence against task-specific reliability looks set to become a central question in business AI.
