What Is Salesforce Koa, and Is It GA for Agentforce?

Salesforce announced a model. Within a day the name was being used three different ways: a new Agentforce edition, a replacement for AIforce, a credit type. It is none of those. Primary sources for everything below: the Koa press release, Sep 15, 2026, the engineering story, Sep 16, 2026, and the Koa product page.

This page is a decoder for buyers. No prices, no GA claim, no recommendation to buy.

01. What Koa is

Salesforce describes Koa as its first CRM reasoning model for Agentforce, built on NVIDIA Nemotron. The engineering story is more specific: Salesforce co-engineered Koa with NVIDIA by post-training the open Nemotron 3 Super model at 120B parameters, using a proprietary synthetic dataset.

Three claims matter to a buyer, and all three are Salesforce's own words. No customer data was used to train the model. Salesforce controls the weights. Post-training and inference both run inside the Salesforce trust boundary. The engineering story extends that to usage: Salesforce says traces and data on Koa do not train the model, and contrasts that with sending your traces to a rented frontier model.

Koa is not the only model Salesforce post-trains. The engineering story names HyperClassifier, TextEval, and Moirai for specialized jobs. Koa is the one aimed at multi-step enterprise reasoning, the kind of work an agent does when it has to chain several CRM actions together.

02. Where Koa sits, and where it does not

Use three names and add nothing.

1. AIforce: the interface layer. Salesforce data, workflows, logic, and permissions brought to any AI interface. Decoder: https://www.therevenueaireport.com/blog/aiforce-vs-agentforce-claudeforce-slackforce
2. Agentforce: the digital workforce.
3. Koa: a CRM reasoning model option that can power Agentforce processes.

Do not flatten Koa into AIforce. They are different layers doing different jobs. AIforce decides where the work shows up. Koa is one of the engines that can do the reasoning behind it. A deck that says Koa is the new AIforce is wrong against both primary sources.

Koa is also not a usage type. Nothing in the press release, the engineering story, or the product page publishes a Koa rate, a Koa credit line, or a Koa edition. If a meter exists, it has not been published. See tokens versus Flex Credits for how Agentforce consumption is actually reported today.

03. Availability: pilot, not GA

The Sep 15 press release availability line is the one to quote: Koa is available to select pilot customers now in Agentforce, and general availability is expected winter 2026 in U.S. regions. Expected is the word. Not committed, not dated.

The product page carries the same pilot line and adds that Salesforce plans an open beta shortly after GA. The engineering story describes a small group of customer pilots running in service, sales, and commerce. Three Salesforce sources, one consistent status: pilot.

The press release names the pilot customers: 1-800Accountant, Baxter Credit Union (BCU), Engine, Formula 1, UChicago Medicine, and Xero. A named pilot list is useful evidence that the pilot is real. It is not evidence that the model is available to you.

One adjacent line so nobody confuses the tracks. The press release lists Missionforce Operations as GA now in the U.S., with post-trained NVIDIA models for Missionforce select customers in October 2026. That is a separate track from Koa in Agentforce. Do not use a Missionforce GA date to argue Koa is shipping.

04. The benchmark numbers, and who is making them

Salesforce publishes two sets of numbers. Both are vendor benchmarks. Attribute them every time you put them on a slide.

• Press release (vendor): Koa matches or exceeds leading model performance on CRM actions with three times fewer errors.
• Product page, CRM Bench (vendor): +11% precision calling the right action, 2.1x reliability recalling context, 15% better context in long conversations.

None of these are Report research and none of them are your numbers. A vendor benchmark run on a vendor-defined CRM benchmark tells you the model is competitive on the tasks the vendor chose. It tells you nothing about error rates on your objects, your permissions model, or your messy pipeline data. The only benchmark that settles the question is the one you run in your own org against your current model, side by side.

05. How you would actually turn it on

The product page describes three deploy paths, and they matter because they tell you how granular the decision is.

1. As a managed LLM in the Data Cloud generative models catalogue.
2. As the org-wide Agentforce model provider in Setup.
3. Selected at the Agent, Sub-Agent, or router level in the builder.

That third path is the one to plan around. If the model can be selected per agent or per router, the rollout is not an org-wide switch, and the rollback is not one either. You can put Koa behind one sub-agent, hold your current model everywhere else, and compare. Build the rollback into the plan before the pilot starts: rollback plan before production.

The product page FAQ also states that no customer data is used in training, describes a temperature 0 serving configuration, and describes a serving harness. Temperature 0 is a determinism setting: the same input is meant to produce the same output rather than a different phrasing each run. If that claim is load-bearing for your audit or compliance review, get it confirmed for your org in writing.

06. The four questions for your AE

1. Are we on the pilot list, and if not, what is the path onto it.
2. Which agents, sub-agents, or routers can select Koa in our org today.
3. Does Koa consume Flex Credits or tokens differently from the default model, and where does that appear in Digital Wallet.
4. Confirm the temperature-0 serving and trust-boundary claims for our org, in writing.

Get the answers on the Order Form or in release notes, not in a call recap. Winter 2026 in U.S. regions is a stated expectation from a press release, and expectations move. Budget against the contract, not the announcement.

What this page is not

• Not a price list. Salesforce has published no Koa list price, credit rate, or edition gate.
• Not a GA claim. Pilot is the word in all three Salesforce sources.
• Not a recommendation to buy or to wait.
• Not a claim about Koa's real-world error rate. Every number above is a vendor benchmark.

More: AIforce, Agentforce, Claudeforce, Slackforce · Tokens versus Flex Credits · AI vendor diligence questions · Buying AI software · AI pricing and contracts

Every Reality Check lands in the weekly issue. Subscribe at therevenueaireport.com/subscribe.