As businesses put more sensitive customer information into artificial intelligence systems, how much company data should have to leave an enterprise platform for an outside AI model to process it?
Salesforce is addressing that concern with Koa, its first reasoning model built specifically for customer relationship management work. The model gives companies a new way to run complex AI tasks without sending the CRM data involved in those requests to an external model provider.
Announced at Dreamforce 2026, Koa was developed with Nvidia and built by further training its open-weight Nemotron 3 Super model. Salesforce controls Koa’s model weights and runs both its post-training and inference inside its own infrastructure, meaning customer information processed by Koa does not cross Salesforce’s trust boundary.
Before Koa, Agentforce tasks requiring longer or multi-step reasoning could be routed through its AI gateway to models such as ChatGPT or Claude. But now, it aims to provide Agentforce customers an alternative to relying on frontier models from companies such as OpenAI and Anthropic for complex reasoning.
How Salesforce Built Koa
Salesforce says no customer data was used to train Koa. Instead, the company created synthetic scenarios based on nearly three decades of CRM experience, covering business workflows across more than 14 industries including healthcare, manufacturing, financial services, and travel.
The simulations were designed to teach Koa how to handle tasks such as generating leads, qualifying opportunities, routing service cases, and choosing the correct tools while completing longer workflows.
Salesforce CEO Marc Benioff said the model “runs entirely inside your trust boundary,” while Nvidia CEO Jensen Huang said Nemotron allowed Salesforce to turn its enterprise knowledge into specialised AI that can reason and securely take action.
Salesforce Is Still Working With OpenAI
Salesforce is not ending its relationship with OpenAI or claiming that OpenAI improperly uses customer data.
Salesforce already operates an Einstein Trust Layer that can mask sensitive information before prompts are sent to external models. The company also has zero data retention agreements with providers including OpenAI, which means data sent through Salesforce is not retained by those providers after a response is returned.
In fact, Koa goes one step further by giving customers the option to keep the entire reasoning process inside Salesforce for supported CRM tasks, and this could be particularly useful for regulated organisations and businesses that want greater control over where sensitive information is processed.
Can Koa Compete With OpenAI and Anthropic?
Salesforce is positioning Koa as a specialist for CRM work rather than a replacement for every general-purpose AI model.
In Salesforce’s technical paper, Koa scored 0.86 on CRM Bench, compared with 0.81 for OpenAI’s GPT-4.1. However, it remained slightly behind Claude Opus 4.8 at 0.87 and OpenAI’s GPT-5.5 at 0.90. The researchers concluded that Koa improved on its Nemotron base model, particularly for multi-turn tool use, while still trailing the strongest frontier models overall.
But Koa is already being used inside Salesforce and is available to select Agentforce pilot customers.
