🔍 You can now filter content to find what interests you! Log in to use the filters. New here? Register and finish setting up your account to get started.

background shape
background shape

Salesforce Launches Koa CRM Reasoning Model for Agentforce

Salesforce announced Koa, its first CRM reasoning model for Agentforce, on September 15, 2026, giving selected customers access to a Salesforce-controlled model designed for complex, multi-step CRM workflows. Koa is available to select pilot customers now, with general availability expected in winter 2026 in U.S. regions, according to the official Koa announcement.

Koa Adds CRM-Specific Reasoning to Agentforce

Koa was built by post-training NVIDIA Nemotron 3 Super with synthetic scenarios modeled on nearly three decades of Salesforce CRM deployment knowledge. The model is intended to help Agentforce agents reason through tasks such as updating opportunities, routing cases, scheduling follow-ups, qualifying leads, and resolving service requests.

The training scenarios cover workflows across more than 14 industries, including manufacturing, financial services, healthcare, and travel. Salesforce says the model was trained without customer data. Instead, the scenarios simulate enterprise personas, policies, tool calls, and sequences of actions that an agent must complete. Independent coverage from CIO confirms the model’s CRM-specific training approach and pilot status. (cio.com)

Salesforce Controls Koa Inside Its Trust Boundary

Salesforce applied supervised fine-tuning and reinforcement learning with Group Relative Policy Optimization, using NVIDIA NeMo RL, NeMo Gym, and NeMo AutoModel. Salesforce controls Koa’s model weights and runs post-training and inference within its own infrastructure, giving Agentforce customers a specialized model option without using customer records as training data.

Salesforce reports that Koa matches or exceeds leading model performance on its CRM benchmark with three times fewer errors. The benchmark covers operational tasks such as updating an opportunity, routing a case, and scheduling a follow-up. That result is Salesforce’s own benchmark claim rather than an independent comparison, so teams evaluating a pilot should validate performance against their actual agent actions, data access rules, and workflow failure modes. (salesforce.com)

Koa Remains Pilot-Only Before Winter 2026 Availability

Koa is already used inside Salesforce, including an internal Slack agent, and is moving into customer pilots with 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero. The model is being positioned as an additional Agentforce model option, not as a replacement for every external model available through Salesforce’s broader AI model strategy. (salesforce.com)

Because Koa is not generally available, most Salesforce teams cannot yet treat it as a standard production model choice. The immediate implementation question for pilot organizations is whether its CRM-specific reasoning improves multi-step task completion enough to justify changing model routing, testing procedures, and production governance once the model becomes generally available.

Oh hi there 👋
I have a SSJS skill for you.

Sign up now to get an SSJS skill that can be used with your AI companion

We don’t spam! Read our privacy policy for more info.

Share With Others

The Author
Marcel Szimonisz Platinum

Marcel Szimonisz

MarTech consultant

I specialize in solving problems, automating processes, and driving innovation through major marketing automation platforms, particularly Salesforce Marketing Cloud and Adobe Campaign.

Your email address will not be published. Required fields are marked *

Subscribe

Get exclusive tips, scripts and news

Choose your topics

We don’t spam! Read our privacy policy for more info.

Similar posts
[mautic type='focus' id='1']