Salesforce Inc. said today at its annual user conference Dreamforce that it has partnered with Nvidia Corp. to train and release Koa, a specialized artificial intelligence model that’s custom built for reasoning about customer relationship management data.
Based on Nvidia’s Nemotron model architecture, Koa is designed specifically to help AI agents reason through complex, multistep workflows within Salesforce’s Agentforce platform, so they can perform specialized tasks across CRM data. It’s an example of an AI model that was post-trained on specialized data, but in a briefing, Salesforce Chief Platform and Engineering Officer Rohan Kumar insisted that it has not tapped any of the vast volumes of customer information stored within its CRM platform. Instead, the company used its customer data to build a massive synthetic dataset crafted out of almost 30 years’ worth of internal CRM deployment experiences.
By using synthetic data that simulates realistic enterprise workflows across 14 industries, including healthcare, finance and manufacturing, Salesforce avoided using any customer data at all, Kumar said. “There is so much knowledge that we’ve sort of gathered in building out the CRM over the last 27 years,” he explained. “And there is this talk around domain-specific models, which take the intelligence the model already has and you then post-train the model. Basically adding more information specific to a certain kind of job. That’s the genesis of Koa.”
According to Kumar, Salesforce combined supervised fine-tuning with reinforcement learning and a technique known as “group relative policy optimization” to teach Koa how to execute multistep tool calls with accuracy. The company went above and beyond to get the training right, and in its benchmark tests, the results suggest that those efforts really paid off, with Koa matching or exceeding the performance of multiple leading general-purpose models on CRM-based tasks with three times fewer errors.
Salesforce said Koa will be useful to enterprise customers that want to try and automate long, multistep processes that usually require human expertise, such as routing support cases, updating sales opportunities, qualifying sales leads and executing multisystem workflows. Because Salesforce hosts and runs Koa within its own trust boundaries within Agentforce via Nvidia’s open framework, customer data never leaves its controlled environment, ensuring it can perform inference safely, Kumar said.
In early pilots, customers highlighted Koa’s ability to streamline complex sales operations. “Accounting requires navigating tax rules, financial data, documents and the unique circumstances of every customer,” said Ryan Teeples, chief strategy officer at 1-800Accountant. “Koa gives our agents the reasoning to work through that complexity step by step and use the right tools along the way. That means we can extend more of our accountants’ expertise across every customer interaction and help people get to the right outcome faster.”
Koa still isn’t generally available yet. Instead, it’s currently being made available to select customers via an expanded pilot program within the Agentforce platform. General availability is slated for winter, Kumar said.
Customers will at least be able to access Nvidia’s Nemotron models via the Salesforce Missionforce platform, a secure cloud platform for high-security operations. The company said it’s integrating Nvidia’s full suite of Nemotron models within Missionforce Operations to provide customers in regulated sectors such as government with “mission-specific intelligence” that taps into their most sensitive data.
This means customers will be able to run the Nemotron models within air-gapped networks and private clouds and automate back-office workflows such as procurement and supply chain management on dedicated infrastructure that doesn’t expose sensitive information to the internet.
Salesforce said it’s going to roll out Nemotron models for Missionforce Operations in October.
