Project MonetRequest demo
Home/Blog/Salesforce Koa: CRM Reasoning Model, Benchmarks & Availability

AI · Historical coverage

Salesforce Koa: What the New CRM Reasoning Model Does

Salesforce has introduced Koa, its first reasoning model built specifically for customer relationship management work. Announced with NVIDIA on September 15, 2026, Koa is designed for Agentforce agents that need to reason through multistep sales, service and other CRM workflows rather than simply generate text.

Published 2026-09-17 · Updated 2026-09-17 · By Project Monet Editorial Team

Project Monet editorial illustration for Salesforce Koa: What the New CRM Reasoning Model Does

01

Overview

Salesforce has introduced Koa, its first reasoning model built specifically for customer relationship management work. Announced with NVIDIA on September 15, 2026, Koa is designed for Agentforce agents that need to reason through multistep sales, service and other CRM workflows rather than simply generate text.

02

What is Salesforce Koa?

Koa is a Salesforce-hosted reasoning model for Agentforce. Salesforce built it by post-training NVIDIA Nemotron 3 Super on a proprietary synthetic dataset modeled on knowledge accumulated from nearly three decades of CRM deployments.

Salesforce says the synthetic scenarios cover more than 14 industries, including manufacturing, financial services, healthcare and travel. They simulate jobs such as generating and qualifying leads, updating opportunities, routing service cases and resolving customer issues. Each scenario maps the actions and tool calls needed to complete a task.

The important distinction is that Salesforce says it did not use customer data to train Koa. The company controls the model weights and says post-training and inference run inside Salesforce's own trust boundary.

03

How Koa was trained

Salesforce says Koa uses supervised fine-tuning plus reinforcement learning with Group Relative Policy Optimization. The training stack includes NVIDIA NeMo RL, NeMo Gym and NeMo AutoModel.

Rather than trying to make Koa a general-purpose frontier model, Salesforce has optimized it for CRM reasoning and tool use. That makes Koa an example of a broader trend toward domain-specific reasoning models built on open-model foundations.

04

Koa benchmarks

Salesforce reports that Koa matches or exceeds leading models on its CRM Benchmark while producing three times fewer errors. The benchmark includes real-world-style CRM actions such as updating an opportunity, routing a case and scheduling a follow-up.

On its Koa product page, Salesforce also reports that Koa is 11% more precise at calling the right action, delivers 2.1 times greater reliability and is 15% better at retaining context in long conversations compared with what Salesforce calls today's default general-intelligence models.

These are Salesforce-reported results. They should not be treated as independent proof that Koa is generally better than Claude, GPT, Gemini or other models. Performance outside Salesforce's CRM benchmark may be different, and Salesforce has not published enough launch-day evidence to make a universal model ranking.

05

Where Koa fits in Agentforce

Koa is a model option for Agentforce. Salesforce says customers will be able to select it at the organization level and for individual agents and sub-agents.

The model is intended for workflows where an agent must understand business context, decide what action should happen next and call the correct tools. Examples include lead qualification, opportunity management, service-case routing, follow-ups and longer processes that cross several CRM actions.

Salesforce is already using Koa internally, including in a Slack agent that helps employees find information and complete everyday tasks.

06

Koa availability

Koa is not generally available yet. As of September 16, 2026, Salesforce says it is available to select pilot customers in Agentforce. Named pilot participants include 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero.

Salesforce says general availability is expected in winter 2026 in U.S. regions. That is a target rather than a guaranteed launch date, so buyers should recheck Salesforce documentation before planning a production deployment.

07

Is Salesforce Koa open source?

Koa is built on NVIDIA Nemotron 3 Super, an open-model foundation, but Salesforce's launch materials do not establish Koa itself as a downloadable open-weight release. Salesforce says it controls Koa's weights and hosts the model within its infrastructure.

Do not confuse the openness of the Nemotron foundation with the distribution terms of Salesforce's post-trained Koa model.

08

Does Koa use Salesforce customer data for training?

Salesforce says no customer data was used to train Koa. Its training corpus was built from synthetic enterprise scenarios.

For inference, Salesforce says Koa runs within its trust boundary and only receives context a customer chooses to provide. Those are vendor-described architecture and privacy claims; organizations with regulatory requirements should still evaluate Salesforce's contractual, security and compliance documentation for their own deployment.

09

Koa pricing

Salesforce has not announced a standalone universal price for Koa. Because it is currently a select Agentforce pilot, any article claiming a fixed Koa subscription or per-token price would be premature.

10

Koa vs general-purpose models

Koa's pitch is specialization rather than general intelligence. Salesforce already supports external frontier models in its broader AI stack. Koa gives it a model it controls and can optimize specifically for CRM reasoning, tool selection and enterprise workflows.

That does not mean Koa is automatically better for coding, creative writing, research, mathematics or other general tasks. A useful comparison will need workload-specific evidence after broader availability.

11

Why Koa matters

Koa shows how enterprise software companies may increasingly build specialized reasoning layers on top of open model foundations instead of routing every difficult task to a general-purpose frontier API.

For Salesforce, that could mean tighter control over model behavior, data boundaries and the economics of Agentforce workflows. For users, the practical question will be whether that specialization produces more reliable actions on real CRM tasks once Koa reaches broader production use.

12

FAQs

When was Salesforce Koa announced?

Salesforce and NVIDIA announced Koa on September 15, 2026.

What model is Koa based on?

Salesforce says Koa was created by post-training NVIDIA Nemotron 3 Super.

Can I use Koa now?

Only select Agentforce pilot customers currently have access. Salesforce expects U.S. general availability in winter 2026.

Did Salesforce train Koa on customer data?

Salesforce says no. It says Koa was trained using synthetic scenarios modeled on CRM workflows.

Is there a Koa API?

Salesforce's launch announcement does not establish a standalone public Koa API. Its documented launch surface is Agentforce.

How much does Koa cost?

Salesforce has not published a standalone universal Koa price.

Sources

Primary and supporting sources

Use the linked documentation to check current pricing, access and product behaviour. The dates above describe this article; they do not guarantee that every vendor claim remains current.

Project Monet

Planning a business website?

Our website planning guides explain scope, costs and customer decisions.

Read the website planning guide