The Pulse
Salesforce puts CRM reasoning inside new Koa model
Salesforce and NVIDIA announced Koa, a CRM reasoning model built for multistep work in Agentforce, with access currently limited to select pilot customers.
AI.info Team ·
Salesforce and NVIDIA are betting that enterprise agents need more than a general-purpose language model. The companies announced Koa on September 15, 2026, a CRM reasoning model designed to handle the multistep work inside Agentforce, from updating opportunities and routing service cases to scheduling follow-ups.
Salesforce says Koa can call the right action 11% more precisely, recall customer context with 2.1 times greater reliability, and retain context in long conversations 15% better than the general-intelligence models it uses as a comparison. Those figures come from Salesforce’s own CRM Bench, not an independent evaluation, and the company describes the model as still moving through customer pilots.
“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. We trained a reasoning engine that understands the structure of a deal, the lifecycle of a service case, and the workflows that vary across industries. That’s a different kind of intelligence, and it runs entirely inside your trust boundary.”
Marc Benioff, Chair and CEO, Salesforce
Salesforce trains Koa for actions, not just answers
Koa is built by post-training NVIDIA Nemotron 3 Super with a proprietary synthetic dataset modeled on enterprise knowledge from Salesforce’s CRM deployments. Salesforce says the dataset contains scenarios covering reasoning, tool use, and decision-making across more than 14 industries, including manufacturing, financial services, healthcare, and travel.
Each scenario pairs a persona with a defined task and maps the sequence of actions and tool calls needed to complete it. That design reflects the job Agentforce is expected to perform: not simply responding to a question, but selecting a workflow, retrieving the permitted context, and taking an action inside a CRM system.
Salesforce says it used supervised fine-tuning and Group Relative Policy Optimization, a reinforcement-learning method, with NVIDIA’s NeMo tooling. The company’s stated goal is to teach Koa both to produce a suitable response and to choose the steps required to reach an outcome.
Customer data stays inside Salesforce’s boundary
Salesforce says no customer data was used to train Koa. The training corpus consists entirely of synthetic scenarios intended to reproduce CRM processes such as generating leads, qualifying opportunities, and resolving service cases.
The company also controls Koa’s model weights and runs post-training and inference within its own infrastructure. Salesforce says customer data does not cross its trust boundary during either training or inference, giving organizations a Salesforce-hosted model option rather than sending CRM context to an external model provider.
Koa is also configured to run at “temperature 0,” according to Salesforce, which is intended to make outputs more consistent and repeatable. A separate serving harness adds trust and safety controls around the model, while customers determine which records, instructions, and grounding data Koa can use.
Koa joins Agentforce’s model menu
Salesforce says Koa can be selected in several places. The company lists it as a managed large language model in the Data Cloud generative models catalogue, as an organization-wide model provider in Agentforce Setup, or at the individual agent, sub-agent, and agent-router level in Agentforce Builder.
The model becomes the fourth provider option in Salesforce Setup, with customers opting in rather than switching every Agentforce deployment at once. Salesforce also says Koa can be used in AI applications, prompts, and agents through the Data Cloud catalogue, although current access is limited to select pilot customers.
That placement gives Salesforce a way to promote a specialized model without removing the model choices already available to Agentforce customers. The company’s developer documentation lists Salesforce-managed models alongside options hosted through other providers, including OpenAI and Anthropic.
Six companies enter Koa pilots
Salesforce says Koa is already used internally, including in an employee agent that helps workers find information and complete everyday tasks in Slack. The model is now moving into pilots with 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero.
Ryan Teeples, chief strategy officer at 1-800Accountant, said the model could help agents work through tax rules, financial data, documents, and each customer’s circumstances step by step. Elia Wallen, founder and CEO of Engine, said the company needed a model that could reason through complex business-travel problems rather than merely sound confident.
Salesforce’s customer examples center on workflows where the value depends on combining business rules, records, and tool calls. UChicago Medicine points to coordination work behind patient care, while Baxter Credit Union describes agents reasoning across member goals, information, tools, and policies.
General availability is planned for winter 2026
Koa is available now to select pilot customers in Agentforce. Salesforce expects general availability in U.S. regions in winter 2026 and plans to open a beta shortly afterward.
The announcement also extends the Salesforce-NVIDIA collaboration beyond commercial CRM. Salesforce says Nemotron-based models and accelerated computing will be brought into Missionforce for government and regulated organizations that need to control model deployment in private clouds, classified networks, or air-gapped environments.
For Agentforce, the immediate test is narrower: whether a model trained around CRM actions can reduce mistakes on the ordinary, consequential steps that make up enterprise work. Salesforce’s own benchmark supplies the initial case for Koa; the customer pilots will show whether those gains hold outside the company’s internal testing.