Salesforce, Nvidia unveil Koa, a CRM-specific AI model

Salesforce's pitch for Koa rests on a specific engineering choice: rather than training on the vast volumes of customer information stored in its CRM platform, the company built a synthetic dataset from almost 30 years of internal CRM deployment experience, simulating realistic enterprise workflows across 14 industries including healthcare, finance and manufacturing. That approach, combined with supervised fine-tuning, reinforcement learning and group relative policy optimization, is what Salesforce says produced a model that matches or exceeds multiple leading general-purpose models on CRM tasks with three times fewer errors.

The move reflects a growing interest in domain-specific models, which take an existing model's general intelligence and post-train it on data specific to a certain kind of job. Koa is designed to help AI agents work through complex, multistep workflows inside Agentforce — routing support cases, updating sales opportunities, qualifying leads and running multisystem processes. Because Salesforce hosts and runs Koa within its own trust boundaries via Nvidia's open framework, the company says customer data never leaves its controlled environment during inference.

Alongside Koa, Salesforce is integrating Nvidia's full suite of Nemotron models into its Missionforce Operations platform, which serves regulated sectors such as government. Those models will run within air-gapped networks and private clouds and are slated to roll out in October, which may matter for customers that need to automate back-office work like procurement and supply chain management without exposing sensitive data. The near-term test for Koa is whether the expanded pilot translates into the general availability Salesforce has slated for winter.