Why We Post-Trained Our Own Reasoning Model

For years, Salesforce’s Agentforce relied on a single general model to handle tasks like intent detection and toxicity screening, treating varied problems with uniform reasoning. This approach lacked the precision needed for repeatable, governed enterprise processes. By shifting to specialized models for specific functions, Salesforce aimed to address these limitations, but core multi-step reasoning remained tied to the general model until now.

Koa, developed in collaboration with NVIDIA, is the first AI reasoning model optimized for enterprise work, running within Salesforce’s trust boundary. Unlike frontier models that reason from first principles, Koa was post-trained on NVIDIA’s Nemotron 3 Super model, leveraging direct access to its weights and transparent training datasets. This adaptation ensures Koa applies learned expertise consistently, avoiding the variability of probabilistic reasoning in governed workflows.

To refine Koa’s capabilities, Salesforce created simulated enterprise workflows across 14+ industries, including manufacturing and healthcare, using synthetic customer scenarios. Through reinforcement learning, Koa practiced multi-turn tool use, improving its ability to resolve complex tasks step-by-step. Scenarios deliberately tested Koa’s response to unavailable tools, training it to halt, communicate limitations, and escalate appropriately—mimicking human judgment in enterprise settings.

Koa integrates with Salesforce’s portfolio of specialized models, such as HyperClassifier and Moirai, to handle targeted tasks while focusing on multi-step reasoning. Early benchmarks on CRM-specific tasks like opportunity updates and case routing show Koa outperforming general-purpose models. Running entirely within Salesforce’s infrastructure, Koa ensures customer data remains under the company’s control, with no real customer data used in training simulations.