At Dreamforce on September 16, 2026, Salesforce introduced Koa, which it describes as its first CRM reasoning model. The interesting thing about Koa is not that it exists, but what kind of model it is. Rather than a new general-purpose frontier system, Koa is a deliberately narrow model, built with NVIDIA and tuned for the specific shape of customer-relationship work inside Salesforce's own Agentforce platform. That design choice is the story worth unpacking.
Salesforce
NVIDIAHow Koa was built
Salesforce says Koa was created by post-training NVIDIA's Nemotron 3 Super, an open-weight reasoning model, on a synthetic dataset modeled on 27 years of Salesforce CRM patterns. Those scenarios were designed to mimic real enterprise workflows across more than 14 industries, including financial services, manufacturing, travel, and healthcare. The result is aimed at long-running agents that carry out multi-step tasks: updating opportunities, routing customer cases, and scheduling follow-ups inside Agentforce.
How a general model becomes a CRM specialist
Koa is not a new base model. It is an open model that Salesforce narrowed to the shape of its own product, which is the point of the design.
Source: Salesforce Dreamforce 2026 announcement. Salesforce reports roughly three times fewer errors than a general model on its own CRM benchmarks.
The pattern here is more general than Koa itself. Take a capable open model, then narrow it with proprietary domain data until it behaves reliably on one company's specific tasks. This is a different bet from chasing the single smartest general model. It trades broad capability for depth in one domain, on the theory that a CRM agent does not need to write poetry or solve olympiad math, it needs to update the right record without breaking anything.
Reading the headline claim carefully
Salesforce reports that Koa demonstrates roughly three times fewer errors on its CRM benchmarks than a general model. That is a meaningful claim, and it is also one to read precisely. It is a vendor-reported result, measured on Salesforce's own benchmarks, for the specific class of CRM tasks Koa was tuned to do. None of that makes it wrong. It does mean the comparison is narrow by construction: a model trained on CRM scenarios and then tested on CRM scenarios should be expected to beat a generalist there, and the useful question for a buyer is whether that advantage holds on their data and their workflows, not on the benchmark.
A CRM agent does not need to write poetry. It needs to update the right record without breaking anything.
On the case for a narrow model
The availability details reinforce the caution. Koa is currently in the hands of select Agentforce pilot customers, reported to include Formula 1, Xero, Baxter Credit Union, and UChicago Medicine, with general availability expected in US regions in winter 2026. In other words, the benchmark exists today and the broad production track record does not yet. That is normal for a launch, but it is the difference between a promising internal result and a proven one.
Why this direction matters
Koa is a clear data point in a shift that ran through 2026: the move from one general model doing everything to a portfolio of models chosen for the job. A frontier generalist is still the right tool for open-ended reasoning. A narrow, domain-tuned model can be cheaper to run, easier to make reliable on a fixed set of tasks, and simpler to govern, because its job is bounded. Neither replaces the other. The enterprise question becomes which model to point at which task, and how to keep that choice flexible as new options appear.
That is the same instinct behind model-agnostic platforms more broadly. Whether the best answer for a given task is a frontier model from a major lab or a specialist like Koa, the value comes from being able to route to it rather than being locked to one. Tools built around that idea, Metir among them, treat the model as a swappable component precisely because the right choice keeps changing. Koa is not a threat to general models or a replacement for them. It is evidence that the field is maturing into one where narrow and broad models coexist, and picking between them per task is becoming a core enterprise skill.
The takeaway
The verifiable facts are these: at Dreamforce 2026, Salesforce launched Koa, a CRM reasoning model built with NVIDIA by post-training Nemotron 3 Super on synthetic CRM data across more than 14 industries, aimed at Agentforce agents, with a reported threefold reduction in errors on Salesforce's own CRM benchmarks and general availability expected in US regions by winter 2026. The larger signal is the strategy. Salesforce is betting that a model narrowed to its domain will outperform a generalist on its domain, and the coming pilot is what will show whether the benchmark advantage survives contact with real customer data.
Sources:
- Salesforce's First CRM Reasoning Model 'Koa' Is Revealed at Dreamforce '26 | Salesforce Ben
- Salesforce teams up with Nvidia to launch Koa, a dedicated CRM reasoning model | IT Pro
- Salesforce Unveils AIforce And Koa, Expands Google Cloud, AWS, And Siemens Partnerships | Pulse 2.0
- Dreamforce 2026: The Top Announcements, Incl. AIforce, Koa & New Partnerships | CX Foundation
Image credits
Hero photograph: Salesforce Tower, San Francisco, in fog, by Frank Schulenburg via Wikimedia Commons, licensed under CC BY-SA 4.0. In-body photograph: Salesforce Tower from Salesforce Park, by Dead.rabbit via Wikimedia Commons, licensed under CC BY-SA 4.0.