企業 AI／代理模型
Salesforce Koa Generates Training Environments from Workflow Specifications, Teaching a 120B Model to Operate CRM Tools
Using Nemotron 3 Super as its foundation, Salesforce trained Koa, its first CRM reasoning model, through synthetic enterprise scenarios, SFT, and GRPO. It delivers clear improvements in multi-turn tool use, but the paper acknowledges that its overall performance still trails the strongest frontier models, while the official claim of “three times fewer errors” comes from Salesforce’s own CRM benchmark.
Salesforce and NVIDIA have introduced Koa, an enterprise agent model post-trained from the open-weight Nemotron 3 Super 120B and slated for integration into Agentforce. Rather than simply feeding the model more CRM documentation, the core approach converts business-process specifications into interactive training environments. Starting from Agent Script descriptions, the system generates tasks with personas, data states, and multi-turn conversations. The model must then call the correct tools to perform actions such as updating opportunities, assigning customer-service cases, or scheduling follow-ups, and receives rewards based on whether it actually resolves the task.
Training consists of supervised fine-tuning and reinforcement learning using GRPO, supported by NeMo RL, NeMo Gym, and NeMo AutoModel. Salesforce says the training corpus covers synthetic scenarios across more than 14 industries and uses no customer data. Salesforce controls the model weights, post-training, and inference, all of which remain within its infrastructure. This gives the enterprise model a different trust boundary from systems that send data to a general-purpose external API, but it does not make Koa a downloadable open-weight model.
The paper’s most notable contribution is its “specification-to-reward” method. For requests that depend on data state, rewards are based on whether tool execution achieves the target outcome, rather than solely on another model’s assessment of the response text. Results show that Koa’s largest gains over the base model are concentrated in multi-turn tool use. Salesforce also claims that, on action tasks in CRM Bench, Koa can match or outperform leading models with three times fewer errors. However, the paper explicitly states that although Koa surpasses a strong proprietary baseline, it still falls short of the strongest frontier models. Salesforce also controls CRM Bench, the synthetic-data distribution, and the deployment environment, limiting independent reproducibility.
Koa is already being used in Salesforce’s internal Slack agent, and selected customers will receive early access. General availability in U.S. regions is planned for winter 2026. Engineering teams should next watch whether model versions can be pinned, how tool permissions and rollback mechanisms are exposed, and whether specification-driven rewards remain reliable when dealing with enterprise custom objects, long-tail workflows, and malicious tool outputs.