Salesforce Introduces Koa, a Reasoning Model for Targeted Business Work
The planned Agentforce option uses synthetic training data, while Salesforce says it is designed to cut token use on its target tasks.
Listen to this story
The audio brief
Story brief
3 key pointsSalesforce and Nvidia are adding a specialized reasoning option to Agentforce with Koa, jointly post-trained from Nvidia’s open-weight Nemotron model. Rather than replacing frontier models for broad, complex work, Koa is aimed at recurring sales, marketing, and customer-support workflows where Salesforce claims it can use fewer tokens than Claude or ChatGPT. The model was trained on synthetic scenarios—not customer...
- 01
Koa was jointly post-trained on Nvidia’s open-weight Nemotron model for narrower enterprise workflows.
- 02
Training used simulated irate-caller and deal-closing scenarios, not actual Salesforce customer data.
- 03
Agentforce previously routed complex, multi-step requests through its AI gateway to Claude or ChatGPT.
Salesforce and Nvidia have introduced Koa, Salesforce’s first reasoning model, for sales, marketing and customer-support work. Salesforce plans to offer the model in Agentforce as an alternative to other models, saying Koa is designed to use fewer tokens on targeted enterprise tasks than frontier models.
Koa was jointly post-trained on Nvidia’s open-weight Nemotron model. Post-training takes a general-purpose model and tunes it for a particular kind of work; here, that work is sales, marketing and customer support. Before Koa, Agentforce sent complex, multi-step tasks through its AI gateway to frontier models such as Claude or ChatGPT.
A narrower job than frontier reasoning
The distinction is in the assignment. Frontier models previously handled the longer, multi-step requests Agentforce routed to them. Koa, by contrast, is intended as an Agentforce alternative for the business functions Salesforce and Nvidia used in its post-training. The companies are pairing a broadly trained base model with a more focused layer for common enterprise interactions.
Training around simulations, not customer records
Salesforce and Nvidia used synthetic data rather than actual Salesforce customer data for Koa’s post-training. The simulations covered customer-service and sales scenarios, including irate callers and a sales professional trying to close a deal. That gives the model examples of the interactions it is meant to address without using Salesforce customers’ own records for the training.
What the post-training data included
- Simulated customer-service settings, including interactions with irate callers.
- Simulated sales work involving a professional trying to close a deal.
- No actual Salesforce customer data.
Salesforce says Koa is designed to use fewer tokens for its targeted enterprise tasks than routing the same work to Claude or ChatGPT. The efficiency claim is central to the model’s case: Salesforce is proposing that a system tuned for specific business workflows can handle selected reasoning tasks with less model usage. Koa’s introduction establishes that option; its planned Agentforce availability will determine where customers can use it.
Editorial analysis
Our Read
Koa turns model selection into a more visible product choice for Salesforce customers. The company’s argument is not that a specialized model replaces every frontier system; it is that familiar sales, marketing and support tasks may be better served by one trained around those workflows. That is a meaningful bet for a platform that previously sent complex, multi-step requests through an AI gateway to frontier models. The next useful evidence will be how Koa performs on the business tasks it targets, and whether Salesforce’s stated token-efficiency advantage holds in Agentforce.
Sources
- techcrunch.comSalesforce and Nvidia's new reasoning model is everything the AI labs should fear | TechCrunch
Loading discussion...
Reader comments
Newest comments first. Replies stay oldest first.