Agentforce
Koa questions, answered: availability, pricing, what it runs on, and who should pilot
Is Koa available, is it US-only, what is it built on, did Salesforce train it on your data, and what does three times fewer errors mean. Twenty of the questions people are searching about Salesforce's CRM reasoning model, answered from the press release and the research paper, with the unpublished parts left unpublished.
Koa is the first Dreamforce announcement in years that came with a research paper attached. The press release gives the claims, and the paper on arXiv, published the day before, gives the numbers behind them, including the ones the keynote did not mention.
That makes Koa unusually answerable. Most of the questions people are searching have a documented answer, a few have a hedged one, and price is not among them. I wrote a longer piece on why Koa exists and whether to care; this page is the short answers.
What Koa is
What is Koa?
Koa is Salesforce’s first CRM reasoning model, built for Agentforce. Koa is a language model post-trained to reason through multistep CRM workflows: updating opportunities, routing cases, scheduling follow-ups, and the tool calls that make those happen. Salesforce and NVIDIA announced it on September 15, 2026, and it sits alongside the rest of the Dreamforce announcements.
What is Koa built on?
NVIDIA’s Nemotron 3 Super, an open-weight model. The paper identifies the base as Nemotron-3-Super-120B and describes post-training with reinforcement learning. Salesforce says the training used proprietary synthetic datasets modelled on nearly three decades of CRM deployments across more than 14 industries.
Is Koa open weight?
Nemotron 3 Super, the base, is an open-weight model from NVIDIA. Whether Salesforce will release Koa’s own post-trained weights is not stated in the press release or the paper, and the deployment described is Salesforce-hosted inside Agentforce. Treat Koa as a hosted model unless Salesforce says otherwise.
Did Salesforce train Koa on my data?
No, according to both sources. The press release says no customer data was used, and the paper states that training used only public resources and synthetically generated interactions. The synthetic scenarios were built to reflect reasoning, tool use and decision-making across CRM workflows.
Is Koa a chatbot or an agent?
Neither. Koa is a model, the reasoning component an Agentforce agent uses to decide what to do next. The agent is still built in Agentforce Builder with topics, actions and guardrails. Koa is one option for the model underneath, alongside Claude, Gemini and Salesforce’s default.
Does Koa replace the Atlas reasoning engine?
No. Atlas is the planning loop that classifies a request, picks a topic, plans actions and checks results. The model inside that loop is what Koa changes. Salesforce’s positioning is that Koa makes the loop’s decisions more accurate on CRM tasks, not that it replaces the loop.
Availability
Is Koa available now?
Only in pilot, to select customers. The press release names 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero. Nobody outside a pilot can use it yet.
When is Koa generally available?
Salesforce says general availability is expected in winter 2026, in US regions. That is the only date and the only region it has published.
Is Koa US-only?
At general availability, yes. Salesforce’s wording is US regions, and practitioner coverage has pointed out the consequence: an org whose data residency is in the EU or APAC cannot plan on Koa for 2026. No date for other regions has been given.
How do I join the Koa pilot?
Through your Salesforce account team. The press release describes pilots as available to select customers and does not publish criteria or a sign-up path. If you have a high-volume agent doing CRM actions today, that is the profile the named pilot customers share.
Where do I select Koa in Agentforce?
The press release does not say. Model selection in Agentforce is at the org level today, across Gemini, Claude and the Salesforce default, and coverage suggests Koa joins that list as a Salesforce-hosted option from winter. Until Salesforce documents it, treat the exact location as unpublished. The model selection post covers how the menu works now.
Cost
How much does Koa cost?
Not published. The press release contains no pricing, and Dreamforce added none. Agentforce actions are metered in Flex Credits regardless of model, so the question is whether Koa changes the per-action rate, and Salesforce has not said.
Will Koa be cheaper than Claude or Gemini in Agentforce?
Unknown. A 120-billion-parameter model hosted by Salesforce could be priced below a frontier model, and Salesforce’s own framing leans on accuracy rather than cost. I would not put a number in a budget until the rate card exists.
Is Koa included in an edition?
Nothing says so. The September editions, Core, Advanced and Max, bundle Flex Credits and Agentforce access but do not mention Koa. Ask when the pricing is published.
Performance
What does “three times fewer errors” mean?
Salesforce’s claim is that Koa matches or exceeds leading model performance on CRM actions with three times fewer errors, measured on Salesforce’s own CRM benchmark of tasks like updating opportunities, routing cases and scheduling follow-ups. The press release does not publish the methodology behind that specific figure.
Is Koa better than Claude or GPT?
On the paper’s numbers, better than GPT-4.1 and below the strongest frontier models. Koa scores 69.41 on Tau2Bench against GPT-4.1’s mark 14.9 points lower, 66.63% on BFCL against 53.96%, and 0.86 on Salesforce’s CRM Bench against 0.81. On the same CRM Bench, Claude Opus 4.8 scores 0.87, and GPT-5.5 scores 83.99 on Tau2Bench. The paper says frontier models remain superior across benchmarks.
What benchmarks was Koa tested on?
Three named in the paper: Tau2Bench and BFCL, which are public tool-use and agentic-reasoning benchmarks, and Salesforce’s own CRM Bench. Salesforce also publishes CRMArena-Pro, a separate enterprise agent benchmark, and I wrote up how to run your own comparison before accepting a pilot.
How was Koa trained?
The paper describes specification-driven reinforcement learning. Agent Script workflow specifications, the same declarative format that configures an Agentforce agent, are expanded into scenario- and persona-conditioned multi-turn tasks, with rewards tied to whether the task was resolved through correct tool use. The training method is GRPO. Agents written in Agent Script are, in effect, the shape Koa was trained on.
What are Koa’s limitations?
The authors list two. The base model had already been through reinforcement learning, so the paper’s comparison of supervised versus RL training is uncertain. And evaluation covered a limited set of enterprise domains, with broader evaluation still needed. Add the ones the paper implies: no non-US availability, no price, and a headline claim whose methodology is not public.
Who should care
Who should pilot Koa?
Teams with an agent already in production doing many CRM actions per day, where wrong actions are the cost that hurts. A model tuned on tool calls could show a difference there, and it matches the profile of the named pilots. A team still deciding what its first agent should do gains nothing from a model choice yet.
Should I wait for Koa before building an agent?
No. Nothing about building an agent changes with the model: topics, actions, grounding and guardrails are the work, and they carry over. Build on the model available now, keep an evaluation set, and switch if Koa beats your current model on your own tasks when it reaches you.
Understanding the basics
Is Salesforce Koa the same as Agentforce?
No. Agentforce is the platform for building and running agents. Koa is a model that an Agentforce agent can use for its reasoning, expected to be one option among Claude, Gemini and the Salesforce default from winter 2026.
Does Koa work outside the United States?
Not at general availability. Salesforce has said GA is expected in winter 2026 in US regions and has given no date for other regions.
What company built Koa?
Salesforce’s Agentforce and AI Research teams, with NVIDIA. The model is post-trained from NVIDIA’s open-weight Nemotron 3 Super, and the paper lists Salesforce as the author organisation with NVIDIA as partner.
What should I do about Koa today?
Keep building on the model you have, and build the evaluation set you will need to judge Koa when it arrives: fifty or so real CRM tasks from your own agent, with the expected tool calls, run through the Testing Center and scored with Custom Scorers. Without that set, a pilot produces an opinion. With it, the pilot produces a number.
The claim to test, and the one to ignore
Two claims are worth separating. “Three times fewer errors on CRM actions” is testable against your own agent the day you get pilot access, and the paper gives you enough to know what to expect: better than GPT-4.1, close to Opus on CRM tasks, behind the frontier elsewhere. Test it.
“Salesforce built a model” is the claim to ignore. The claim is true, and it changes nothing about the work of building an agent that behaves in production. The teams that get value from Koa will be the ones with an agent worth improving and a test set that can tell.
If you want the evaluation set built before a Koa pilot, or a second opinion on whether your agent is the profile Koa was made for, talk to us, or start with the Agentforce readiness assessment.
