For most of the last three years, software vendors treated large language models the way they treated cloud compute: rent capacity from a foundation lab, wrap a product interface around it, and compete on integration rather than the model itself. Salesforce and NVIDIA broke that pattern this week with Koa, a CRM-specific reasoning model trained on what Salesforce describes as twenty-seven years of enterprise deployment knowledge. It is the clearest signal yet that the next competitive line in enterprise sales AI runs through who owns the training data, not who has the best prompt.
A model built from CRM history, not a wrapper around one
Koa is not another agent skin layered over a general-purpose model. Salesforce and NVIDIA post-trained NVIDIA’s Nemotron 3 Super using proprietary synthetic data modeled on nearly three decades of CRM deployment patterns spanning fourteen industries, including manufacturing, financial services, healthcare and travel. Engineers applied supervised fine-tuning and reinforcement learning with Group Relative Policy Optimization through NVIDIA’s NeMo RL, NeMo Gym and NeMo AutoModel tooling. Salesforce says no customer data entered the training set and that it controls the resulting model weights, running inference inside its own infrastructure rather than routing customer data through NVIDIA or any outside party.
“The most valuable thing Salesforce has built isn’t our platform, it’s the accumulated knowledge of how enterprise business actually works,” said Marc Benioff, Chair and CEO of Salesforce. “With Koa, the knowledge is put inside the model itself.” Jensen Huang, founder and CEO of NVIDIA, framed the move as a wider shift in where software value will sit: “AI is creating a much larger opportunity for software. Every company needs useful AI, tailored to its knowledge, expertise, and work.”
Advertisement
Why this counts as a bellwether, not a one-vendor story
Salesforce is the largest CRM vendor by installed base, which is exactly why its choice matters beyond its own product line. For two years, the default AI strategy for enterprise software vendors has been to call an outside model’s API and differentiate on workflow. Koa is a bet that the differentiation buyers will actually pay for sits one layer deeper: a model tuned on the specific failure modes of CRM work, such as which fields a rep forgets to update, which stage transitions get logged wrong, and which handoffs between sales and service break down. That is a hard asset to replicate without the underlying transaction history, and it is why rival platforms with less deployment history behind them will find this harder to copy quickly.
Salesforce reports that in its own CRM benchmark, Koa matches or exceeds leading model performance with three times fewer errors on CRM actions such as updating opportunities, routing cases and scheduling follow-ups. That figure comes from Salesforce’s internal benchmark rather than an independent one, which is worth remembering before treating it as a settled comparison. Still, the pilot roster gives the claim some early, real-world weight. Early customers include 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero.
The rest of the CRM field has largely taken the opposite path so far: layer an assistant on top of an existing platform and call whichever foundation model is cheapest or fastest that quarter. That approach is faster to ship but easier for a rival to copy, since the reasoning underneath it belongs to whichever lab licensed the API. A proprietary reasoning model tuned on a vendor’s own multi-decade deployment history is slower to build and far harder to catch up to once it exists, which is exactly the kind of asset a market leader has the scale and history to build first.
What sales leaders are already saying
Ryan Teeples, Chief Strategy Officer at 1-800Accountant, said Koa “gives our agents the reasoning to work through complexity step by step and use the right tools along the way.” John Sahagian, SVP and Chief Data Officer at Baxter Credit Union, said the model “can help our Digital agents understand the full complexity and context behind member goals and reason across the information, tools, and policies.” Elia Wallen, founder and CEO of Engine, put the underlying demand plainly: “What we need from AI isn’t a model that sounds confident, it’s one that can reason precisely through complex, multi-step problems.”
That distinction, between a model that sounds confident and one that reasons correctly through a multistep CRM workflow, is the actual product being sold here. A wave of job-shaped sales agents has already moved AI from a chat assistant into something closer to a named team member. Koa is the layer underneath those agents: the reasoning engine that decides whether an update, a routing decision or a follow-up is actually correct, not just plausible.
What it means for the sales leader
For a revenue organization evaluating AI vendors right now, the practical question shifts. It is no longer just “which foundation model does this agent call,” but “whose deployment history trained the reasoning underneath it.” A sales leader buying an AI SDR, a forecasting agent or a CRM copilot should ask what training data actually sits behind the reasoning layer, whether that data is the vendor’s own transaction history or a generic model with a thin fine-tune on top, and what independent evidence exists beyond the vendor’s own benchmark. Salesforce’s recent move to keep the acquired Fin support agent separate from Agentforce is a reminder that even Salesforce still fills some AI gaps by acquisition rather than in-house training. Koa signals where it wants to build instead of buy.
Koa is available to select pilot customers now, with general availability expected in winter 2026 for U.S. regions. Post-trained NVIDIA models inside Missionforce Operations, aimed at government and regulated organizations that need control over model, data and deployment environment, are generally available now in the U.S., with select customer access in October 2026. Sales and RevOps leaders evaluating any AI agent this quarter should treat the training-data question as a standard part of vendor diligence, not an afterthought, because the vendors are already treating it as the product.
Source: Salesforce
