Last updated: October 26, 2026
Short answer: Salesforce Koa is Salesforce’s first CRM reasoning model for Agentforce, built with NVIDIA on the Nemotron 3 Super model and announced on September 15, 2026. It is trained to work through multi-step CRM tasks such as qualifying leads, routing cases and updating opportunities, and Salesforce reports it matches or beats leading models on CRM actions with about 3x fewer errors. It is in pilot now, with general availability expected in Winter 2026 in US regions. Most orgs should test it on high-volume agent tasks when it reaches their region, but should not rip out their current model until Salesforce publishes pricing.
What is Salesforce Koa?
Salesforce Koa is a reasoning model built for one job: powering Agentforce agents that work inside CRM processes. Announced at Dreamforce on September 15, 2026 as a joint effort between Salesforce and NVIDIA, it is not a general chatbot model. It was built to understand the shape of a deal, the lifecycle of a service case and the step-by-step tool use that real business workflows need.
Two things make it different from the models Agentforce has used so far. First, it is specialized: the training focused entirely on CRM-style reasoning rather than general knowledge. Second, it is Salesforce-hosted. Salesforce controls the model weights and runs post-training and inference inside its own trust boundary, which matters to any company whose blocker on AI agents has been data leaving the vendor perimeter.
Koa vs a general-purpose model: at a glance
| Salesforce Koa | General frontier models in Agentforce | |
|---|---|---|
| Built for | CRM workflows: leads, opportunities, cases, follow-ups | Everything, with no CRM specialization |
| Base | NVIDIA Nemotron 3 Super, post-trained by Salesforce | Vendor frontier models (OpenAI, Anthropic, Google) |
| Hosting | Inside Salesforce infrastructure, weights controlled by Salesforce | Calls out to the model provider |
| Reported strength | Matches or exceeds leading models on CRM actions with about 3x fewer errors (Salesforce CRM Bench) | Stronger on unusual, complex or open-ended tasks |
| Availability | Pilot now; GA expected Winter 2026, US regions first | Available now |
| Pricing | Not published yet | Known through Agentforce consumption |
How Salesforce Koa was built, in plain terms
Salesforce started from NVIDIA Nemotron 3 Super, an open-weight 120B model, and post-trained it using supervised fine-tuning plus reinforcement learning (a method called GRPO), with NVIDIA NeMo tooling. The training corpus is fully synthetic: scenarios modeled on 27 years of CRM deployment knowledge across more than 14 industries, covering the reasoning, tool selection and decision-making an agent needs for tasks like generating leads, qualifying opportunities and resolving service cases. Salesforce states that no customer data was used to train the model.
That last point is worth pausing on. The pattern here, a strong open-weight base plus deep domain knowledge layered on top, is becoming the standard recipe for enterprise AI. The moat is not the biggest model; it is the workflow knowledge the vendor can encode into a smaller one.
Where Salesforce Koa shows up in your org
Once it is available to you, Salesforce Koa appears in three places, and you opt in like any other provider:
- As a managed LLM in the Data Cloud generative models catalogue (Data Cloud is now called Data 360; here is what that rename actually changed)
- Org-wide, as a fourth model provider option in Agentforce Setup
- Per agent and sub-agent, selected at the individual agent level
The per-agent option is the one to care about. It means you can route your high-volume, well-defined agents to Koa while keeping a frontier model behind the agents that handle messy, open-ended work. Model routing, not model loyalty, is where Agentforce economics are heading, and routing decisions feed straight into how your Agentforce credits get consumed.
The claims, and what is still missing
Salesforce reports that on CRM Bench, its benchmark of real CRM actions like updating an opportunity, routing a case and scheduling a follow-up, Koa matches or exceeds leading model performance with about three times fewer errors. Independent reporting places Koa ahead of some general models on public tool-use benchmarks while still below the top frontier models, which is exactly what you would expect from a specialized model. Early pilot customers include 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero.
Three numbers have not been published: price, context window and latency. For a hosted model sold inside Agentforce, price is the one that decides adoption. Until it lands, treat every switch-over plan as provisional.
Should your agents switch to Salesforce Koa?
Our read, as a team that configures Agentforce for US clients:
- Raise your hand for the pilot if you run Agentforce in a US region and your main blocker has been data boundary or model governance. Koa staying inside Salesforce infrastructure is the strongest part of the offer today.
- Plan a comparison test, not a migration, if you already have agents in production. Pick one high-volume agent, run the same scenarios on your current model and on Koa when it reaches GA, and compare error rates and credit burn. Switch only on evidence.
- Fix your data first if your agents are not live yet. A specialized reasoning model still reasons over whatever your CRM contains, and duplicates or stale records produce confidently wrong answers from any model. Data readiness is step one of any Agentforce rollout we run.
- Wait if you are outside US regions or price-sensitive. GA is US-first and the pricing is unknown, so there is nothing to act on yet beyond awareness.
Free Salesforce implementation checklist
Thinking about Agentforce and Koa? The groundwork is the same: scoping, data readiness and adoption. These are the steps we check on every Salesforce project.
How Salesforce Koa compares on benchmarks
Benchmark numbers deserve some skepticism, especially when the benchmark belongs to the vendor, so here is the picture across both kinds of tests.
On CRM Bench, which is Salesforce’s own suite of real CRM actions, Koa matches or exceeds the leading general models while making roughly a third of the errors. On public benchmarks reported independently, Salesforce Koa lands ahead of several well-known general models on multi-turn tool use while sitting clearly below the top frontier models on raw capability. Both results can be true at the same time, and together they describe a specialist: strong inside its domain, ordinary outside it.
The practical question for a business is not which model is smarter in the abstract. It is which model makes fewer mistakes on the twenty tasks your agents repeat hundreds of times a week, and at what cost per task. A model that mis-routes one case in two hundred instead of one in seventy changes your escalation volume, and that shows up in staffing before it shows up in any benchmark chart.
It also matters where the model runs. Salesforce Koa is hosted inside Salesforce’s infrastructure with a serving layer that adds trust and safety controls on top of the model itself. For regulated industries, Salesforce and NVIDIA are extending the same approach to Missionforce and environments like private clouds, where the organization keeps control over the model, data and deployment. If your security team has been the reason your agent project is stuck, these are the paragraphs of the announcement to forward to them.
What to do this quarter if you run Agentforce
Salesforce Koa does not change what makes agent projects succeed. It changes one input, the model, and leaves the rest of the work where it always was. Between now and general availability, three moves make sense.
First, inventory your agents by task type. Mark the high-volume, well-defined workflows (case routing, follow-up scheduling, record updates) separately from the open-ended ones (research, drafting, judgment calls). The first group is Koa’s territory; the second will probably stay on frontier models. Having this split written down turns the GA announcement into a one-day decision instead of a quarter-long debate.
Second, baseline your current error rates and credit consumption per agent. You cannot tell whether a new model is an improvement if you never measured the old one. Even a simple weekly count of wrong answers, escalations and credits burned gives you the comparison you will need.
Third, clean the data your agents read. Every reasoning model, Koa included, answers from the records in front of it. Duplicate accounts, stale pricing in knowledge articles and orphaned contacts produce confident nonsense from the best model on the market. This is unglamorous work, and it is the single highest-leverage preparation for any model you choose.
Frequently asked questions
What is Salesforce Koa?
Koa is Salesforce’s first CRM reasoning model for Agentforce, built with NVIDIA on the Nemotron 3 Super model and trained on synthetic scenarios modeled on 27 years of CRM workflows. It is designed to reason through multi-step CRM tasks and pick the right tools along the way.
Is Koa better than GPT or other frontier models for Agentforce?
It depends on the task. Salesforce reports Koa matches or exceeds leading models on CRM-specific actions with about 3x fewer errors, while independent benchmarks place it below top frontier models on general tasks. Specialized work favors Koa; unusual or open-ended work still favors frontier models.
How much does Salesforce Koa cost?
Salesforce has not published Koa pricing yet, and the context window and latency figures are also unannounced. Pricing will likely sit inside Agentforce consumption, and it is the number to wait for before planning a switch.
Do I need Data 360 to use Koa?
Koa can be selected as a managed LLM in the Data 360 (formerly Data Cloud) generative models catalogue, but it is also available org-wide as a model provider in Agentforce Setup and at the individual agent level. You do not need to adopt the whole data platform just to try the model.
When will Koa be generally available?
Salesforce expects general availability in Winter 2026, starting in US regions. It is currently with select pilot customers, including 1-800Accountant, Formula 1, UChicago Medicine and Xero.
Was customer data used to train Koa?
No. Salesforce states the training corpus was entirely synthetic, built to reflect CRM reasoning and tool use, and that no customer data was used. The model also runs inside Salesforce’s own infrastructure rather than calling an external provider.
