The Pulse
Salesforce Unveils Koa CRM Model for Agentforce
Salesforce introduces Koa, a CRM reasoning model built on NVIDIA Nemotron for Agentforce agents. The company says Koa improves action selection, context retention and reliability, but its published results come from Salesforce’s own CRM Ben
AI.info Team ·
Salesforce says its new Koa model performs better than general-purpose models on CRM tasks, but the evidence comes from a benchmark the company built and runs itself. The model, unveiled September 15, is designed to power Agentforce agents through multi-step work such as updating opportunities, routing service cases and scheduling follow-ups.
Salesforce describes Koa as its first CRM reasoning model. Built by post-training NVIDIA Nemotron open models, Koa uses synthetic business scenarios rather than customer data and runs inside Salesforce’s own infrastructure. The company says the design gives it more control over model weights, data handling and the behavior of agents operating on Salesforce records.
Salesforce’s benchmark makes the case for specialization
Salesforce reports that Koa matches or exceeds leading model performance on CRM actions in its CRM Bench, while producing three times fewer errors. The benchmark includes tasks drawn from common CRM workflows, including opportunity updates, case routing and follow-up scheduling.
Against what Salesforce calls its current default general-intelligence models, the company reports that Koa is 11% more precise when selecting the right action, 2.1 times more reliable at recalling customer context and 15% better at retaining context through long conversations. Those figures are Salesforce measurements; the company’s announcement does not identify the competing models or provide an independent evaluation.
That distinction matters because Koa is not presented as a general-purpose model. Salesforce has trained it for the instructions, tool calls and business policies that Agentforce agents use across customer-service, sales and operational workflows. The published results therefore make a narrower claim: a model tuned for Salesforce work may perform better on Salesforce work.
Twenty-seven years of CRM workflows, without customer data
Koa’s training corpus consists entirely of synthetic scenarios representing the reasoning and tool use required across the customer lifecycle. Salesforce says those scenarios cover more than 14 industries, including manufacturing, financial services, healthcare and travel.
Each scenario pairs a persona with a task and maps the sequence of actions required to complete it. Examples include generating leads, qualifying opportunities and resolving service cases. Salesforce says no customer data is used to train Koa.
The model also operates under a set of restrictions designed for enterprise deployments. Salesforce says Koa runs at “temperature 0” to produce more consistent outputs, follows customer instructions and guardrails, and uses a separate serving harness for additional trust and safety controls.
Koa joins Agentforce’s model choices
Customers can select Koa as a managed large language model in the Data Cloud generative models catalogue, choose it as an organization-wide Agentforce model provider or assign it to individual agents and sub-agents. Salesforce also says an agent router can use Koa, giving administrators a way to apply the model to selected parts of a broader workflow.
The company positions that flexibility as a response to the limits of a single model for every enterprise task. A model may be selected across an organization, used only by a particular agent or applied to a sub-agent handling a specific class of work. Koa is therefore being added as an option inside Agentforce rather than replacing the platform’s existing model arrangements.
Customer pilots start before general availability
Koa is already used inside Salesforce’s Employee Agent, which helps staff find information and complete routine tasks in Slack. Salesforce says the model is moving into customer pilots with 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero.
“At Engine, we're building the future of business travel, and that starts with equipping our teams to deliver exceptional service. Business travel has countless moving parts, and what we need from AI isn't a model that sounds confident — it's one that can reason precisely through complex, multi-step problems. Collaborating with Salesforce and NVIDIA on a reasoning model purpose-built for Agentforce lets us push that vision further, faster.”
Elia Wallen, Founder and CEO, Engine
Salesforce says Koa is available now to select pilot customers. General availability in U.S. regions is planned for Winter 2026, with an open beta expected shortly afterward.
The immediate test for Koa will not be whether it can produce convincing answers. Agentforce customers will judge it on narrower measures: whether it selects the correct action, preserves the right customer context and completes a workflow without sending an employee back through the CRM to fix the result.