The short version
- Koa is Salesforce’s first CRM reasoning model, built with NVIDIA by post-training NVIDIA’s Nemotron on synthetic data modeled on roughly 27 years of CRM work across 14-plus industries. No customer data was used to train it.
- Salesforce says Koa "matches or exceeds leading model performance on CRM actions with three times fewer errors" on its own CRM Benchmark, tasks like updating an opportunity, routing a case, or scheduling a follow-up.
- It runs inside [Agentforce](/solutions/agentforce), chosen automatically by an AI gateway, as a cheaper model for high-volume, well-scoped work, with frontier models like [Claude](/blog/claudeforce) reserved for harder judgment.
- The strategic point is cost. Routing routine work to a cheap CRM-tuned model and hard work to frontier models changes the unit economics of running agents at scale.
- The honest caveat: the headline benchmark is Salesforce’s own, there is no published head-to-head against Claude or ChatGPT, and "cheaper tokens" only wins if Koa does not retry or escalate too often.
The Flashiest Announcement Was AIforce. The Smartest One Was Koa.
The flashiest Dreamforce announcement was AIforce. The most strategically interesting one was a model most customers will never choose on purpose. Koa is Salesforce’s first CRM-specific reasoning model, and the reason it matters is not raw capability. It is cost. Koa was one of several launches at the event, covered in full in our Dreamforce 2026 recap.
What Koa Is, and How It Was Built
Salesforce describes Koa as its "first CRM reasoning model, purpose-built to help agents reason through complex, multistep workflows and use the right tools to complete work." It was built with NVIDIA by post-training NVIDIA’s Nemotron model on a proprietary synthetic dataset, using supervised fine-tuning plus reinforcement learning on NVIDIA’s NeMo tooling.
The training data is the interesting part. Salesforce did not use customer records. Instead, it and NVIDIA generated synthetic scenarios modeled on nearly three decades of Salesforce CRM knowledge across more than 14 industries. As Jayesh Govindarajan, EVP of Salesforce AI, put it: "reasoning has always been something that we’ve relied on the frontier model providers for. Until now."
One nuance a lot of coverage blurred: the base Nemotron model has open weights, but Koa as shipped does not. Salesforce controls Koa’s weights and runs it inside its own infrastructure, so no customer data crosses its trust boundary at inference. "Salesforce released Koa’s weights" would be wrong. "Koa is built on an open-weight foundation and run privately" is right.
The Claim, and the Benchmark Behind It
Salesforce’s headline number is that Koa "already matches or exceeds leading model performance on CRM actions with three times fewer errors." That figure is measured on Salesforce’s own CRM Benchmark, a suite of tasks like updating an opportunity, routing a case, or scheduling a follow-up.
It is a real result and a reasonable thing to test. It is also a vendor-defined, vendor-run benchmark, and Salesforce has not published a head-to-head of Koa against Claude or ChatGPT on the same agentic tasks. A model built and measured by the same company, on tasks it also defined, is a starting point for a pilot, not a finish line.
The Real Story: The Gateway
Koa arrived about three weeks after Salesforce made Claude a default model across Agentforce. That timing is the point. Inside Agentforce, an AI gateway routes each task to a model: the cheap, CRM-tuned Koa handles high-volume, well-scoped actions, and frontier models like Claude and ChatGPT are reserved for work that needs genuine judgment.
NVIDIA’s Kari Ann Briski called it "the trifecta: sovereign AI, time to first token, efficient reasoning." Routing high-volume, simple queries to a cheaper model, at Salesforce’s scale, changes the unit economics of running agents. As one analysis put it, the gateway itself may be the actual product, more than any single model, because it shifts value from fixed per-token pricing to smart routing.
The Honest Read
A small, CRM-specific model for routine actions is genuinely smart architecture. Most agent work in a CRM is repetitive and well-scoped, exactly the kind of task where a purpose-built model can be both cheaper and more reliable than a general one. And the data-trust story, no customer data in training, weights run inside Salesforce’s boundary, is real and matters for regulated firms.
The caveats are equally real. The accuracy claim rests on a benchmark Salesforce built. Synthetic training data captures the common shapes of CRM work, not your customers’ strangest edge cases. And "cheaper tokens" only wins if Koa does not quietly retry or escalate to a frontier model so often that the savings evaporate. That failure-and-escalation rate is the number that matters, and it is not one Salesforce has published.
What a Mid-Market Team Should Do
Understand that you will not choose Koa directly. The gateway routes to it. Your job is not to pick a model; it is to judge outcomes.
When it is generally available, pilot it on your own tasks. Koa is in pilot now, with general availability expected winter 2026 in US regions. Run your real workflows and track how often the agent gets it right the first time, and how often it escalates. That is your true cost and quality picture, not the vendor benchmark.
Do not put "three times fewer errors" in your own deck. It is a vendor claim on a vendor benchmark. Cite your own results once you have them. If you want help measuring agent accuracy on your workflows, that is exactly what our Agentforce practice does, and a free scoping call is the place to start.