Technology and Science · updated 1h ago · 3 min read
Salesforce And NVIDIA Unveil Koa CRM Reasoning Model Built On Nemotron
Salesforce and Nvidia unveiled Koa, a CRM reasoning model built on Nemotron at Dreamforce 2026. Koa is trained on 27 years of Salesforce CRM knowledge for multi-step tasks.
Whether Koa mainly signals frontier-lab disruption or CRM marketing moat.
6 of 8 outlets skipped it: koa is already deployed via an agent in Slack.
13outlets compared
Same story, two versions
tap a side to read it in full
TechCrunch
“a shining example of how the enterprise world’s needs for AI are diverging from what the frontier labs are offering.”Read the original ↗
MarTech
“Salesforce hopes its new tailored-for-CRM AI will be a boon for marketers and a moat against the general-purpose models of companies like OpenAI, Anthropic, and Google.”Read the original ↗
TechCrunch fears frontier labs being outflanked, while MarTech stresses a CRM-specific marketing moat.
Koa for CRM agents
Salesforce and NVIDIA unveiled Koa, Salesforce’s first CRM reasoning model for its Agentforce platform, built on NVIDIA Nemotron and designed to help agents reason through complex, multistep workflows.
“No customer data was used to train the Koa reasoning model.”
Salesforce said Koa was built by post-training NVIDIA Nemotron 3 Super with a proprietary synthetic dataset modeled on enterprise knowledge from nearly three decades of CRM deployments, and that “No customer data was used to train the Koa reasoning model.”
In Salesforce’s CRM benchmark, the company said Koa “already matches or exceeds leading model performance on CRM actions with three times fewer errors,” including tasks like updating an opportunity, routing a case, or scheduling a follow-up.
The company also said Salesforce controls the model weights and performs post-training and inference entirely within its own trust boundary, so customer data doesn’t cross that boundary during inference.
At Dreamforce, Salesforce framed Koa as part of a broader shift where specialized reasoning models can be paired with other AI interfaces, while Koa handles CRM process knowledge and tool sequences rather than just generating content.
Partnerships and governance
Salesforce positioned Koa alongside other model options inside Agentforce, including Gemini models through Salesforce’s Google Cloud partnership and models available through Amazon Bedrock via an expanded AWS integration.
Computer Weekly described AIforce releases that let users work through Anthropic’s Claude or Slack instead of logging on to Salesforce, with Claudeforce embedding Salesforce data and actions inside Claude and Slackforce pairing Slackbot with a capability to build small applications inside Slack.
In that same coverage, Salesforce’s Patrick Stokes said, “The UI is not actually our product,” adding that the Salesforce product is the platform where customers store their data, metadata, workflows, and permissions.
Anthropic’s Scott White said the partnership fills “the two things an agent needs to be useful at work,” and he emphasized access control by saying, “Claude only sees what the person is allowed to see.”
Omdia’s Lian Jye Su argued that “Agents with skills generally work well in specific enterprise workflows,” and said skills can make agents more predictable and cost-efficient, with vendors able to upgrade and update skills over time.
Pilots and Missionforce
Salesforce said Koa is already in use inside Salesforce, including “an agent in Slack that helps employees find information and complete everyday tasks in Slack,” and it is moving into customer pilots with 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero.
“an agent in Slack that helps employees find information and complete everyday tasks in Slack”
In the pilots, Ryan Teeples of 1-800Accountant said, “Accounting requires navigating tax rules, financial data, documents, and the unique circumstances of every customer,” while John Sahagian of Baxter Credit Union said Koa can help digital agents “reason across the information, tools, and policies needed to move them forward.”
Salesforce and NVIDIA also said they are bringing Nemotron-based models and accelerated computing into Missionforce for government and regulated organizations, extending control over model, data, and deployment environment to private clouds and air-gapped networks.
MarketScreener España reported that Salesforce applied supervised fine-tuning (SFT) and GRPO using NVIDIA NeMo RL, NeMo Gym, and NeMo AutoModel, and that Salesforce controls weighting parameters and runs Koa within its own infrastructure to ensure “no customer data crosses trust boundaries during the inference process.”
The same announcement framed Missionforce Operations agents as digitizing and automating government workflows, including procurement, supplier management, and logistics, while positioning Koa and Nemotron-based models as mission-specific AI for secure environments.