Salesforce and NVIDIA announced Koa on September 15, a reasoning model the companies describe as the first built specifically for customer relationship management work. Rather than training a foundation model from scratch, Salesforce post-trained NVIDIA's open-weight Nemotron 3 Super model — a 120B-parameter base — on a proprietary synthetic dataset modeled on nearly three decades of CRM deployment knowledge. The model now powers specialized agents on Salesforce's Agentforce platform, and pilot customers including Formula 1, Xero and UChicago Medicine are already trying it.
The announcement, made during Salesforce's Dreamforce conference in San Francisco, marks a notable shift in how enterprise software vendors are approaching AI. Instead of renting intelligence from frontier labs, Salesforce built its own specialized model — and kept control of the weights. For more context on this story, see our ongoing AI industry coverage.
How Koa Was Built — Without Customer Data
Salesforce is explicit about one point: no customer data was used to train Koa. The training corpus consists entirely of synthetic scenarios covering the reasoning, tool use and decision-making that Agentforce agents perform across CRM workflows — generating leads, qualifying opportunities, resolving service cases.
According to the joint announcement, the scenarios simulate real enterprise workflows across more than 14 industries, including manufacturing, financial services, healthcare and travel. Each scenario pairs a persona with a set of tasks, and the sequence of actions and tool calls required to complete them is mapped in advance.
For post-training, Salesforce applied supervised fine-tuning followed by reinforcement learning using Group Relative Policy Optimization (GRPO), run on NVIDIA's NeMo RL, NeMo Gym and NeMo AutoModel tooling. The company says training on a targeted set of prioritized enterprise tasks taught the model to reach goals through sequences of correct actions, rather than simply producing a correct answer.
The Research Behind the Model
A research paper submitted to arXiv on September 14 describes Koa's distinctive component as a simulation-to-reward pipeline: workflow specifications are turned into persona-conditioned, multi-turn tasks, with rewards tied to resolving each task through successful tool use. For enterprise domains, those specifications are written in Agent Script, Salesforce's declarative language for building Agentforce agents. The same simulation and grounded-reward machinery drove GRPO across both enterprise and public tool-use domains.
The authors report that Koa improves on its open-weight base across public tool-use, agentic-reasoning and enterprise CRM benchmarks, with the clearest gains in multi-turn tool use. The model surpasses a strong proprietary baseline while remaining below the strongest frontier models — a gap Salesforce is betting does not matter for CRM work.
Benchmarks: Three Times Fewer Errors on CRM Tasks
On Salesforce's own CRM benchmark — a suite of real-world tasks such as updating an opportunity, routing a case or scheduling a follow-up — the company says Koa matches or exceeds leading model performance on CRM actions with three times fewer errors.
That claim matters less for its headline number than for its premise: a smaller model specialized in one domain can beat general-purpose frontier models at that domain's tasks. It is the same argument open-weight proponents have made all year, now with an enterprise testcase attached.
Trust Boundary and Government Push
Salesforce says it controls the Koa model weights and performs post-training and inference entirely within its own trust boundary, so no customer data crosses that boundary during inference. That positioning is aimed squarely at regulated industries.
Alongside Koa, the two companies are bringing Nemotron-based models into Missionforce, Salesforce's government-focused offering. Post-trained NVIDIA models will power Missionforce Operations, which digitizes government workflows including procurement, supplier management and logistics — including deployments on air-gapped networks that never touch public infrastructure. Missionforce Operations is generally available now in US regions, with the post-trained models reaching select customers in October 2026.
Pilots Underway, General Availability This Winter
Koa is already used inside Salesforce itself, including a Slack agent that helps employees find information and complete everyday tasks. Customer pilots are underway with 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero.
Ryan Teeples, chief strategy officer at 1-800Accountant, said Koa lets the firm's agents reason step by step through tax rules, financial data and customer documents. At UChicago Medicine, chief marketing officer Andrew Chang said the model can handle longer multi-step coordination workflows, freeing teams for patient care.
General availability is expected in winter 2026 in US regions.
Why It Matters
The strategic subtext is hard to miss. As Techzine noted, Salesforce announced its own model just weeks after expanding its partnership with Anthropic to bring Claude into its products. CEO Marc Benioff framed Koa as the product of that accumulated experience: "The most valuable thing Salesforce has built isn't our platform — it's the accumulated knowledge of how enterprise business actually works. With Koa, the knowledge is put inside the model itself."
NVIDIA's Jensen Huang argued that Nemotron's open models gave Salesforce the basis for a CRM model that can reason and act securely. For NVIDIA, every enterprise that post-trains Nemotron is another distribution channel for its computing stack.
Koa is also a data point in a broader trend for 2026: enterprises specializing open-weight foundation models for narrow, high-value domains rather than waiting for frontier labs to ship vertical products. Specification-driven reinforcement learning, the arXiv paper's authors conclude, offers a practical route to exactly that.
If the three-times-fewer-errors claim holds in production, expect other enterprise software vendors to follow the same path — and expect frontier labs to argue, with some justification, that their next models will close the specialized gap anyway.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
