Salesforce Koa Explained: Agentforce Access, Pilot Status and Model Selection

What is Salesforce Koa?

Koa is Salesforce’s first CRM reasoning model for Agentforce. Instead of positioning it as another general-purpose chatbot model, Salesforce says Koa is designed to reason through the kinds of multistep work that happen inside customer relationship management systems: qualifying opportunities, routing service cases, scheduling follow-ups, updating records, and choosing the right tools across a workflow.

Salesforce developed Koa with NVIDIA and built it on NVIDIA Nemotron 3 Super. The important part is not merely the base architecture. Salesforce post-trained that model on enterprise scenarios intended to encode the structure of CRM work into the reasoning process itself.

That makes Koa a domain-specific model. A broad frontier model is trained to perform well across many kinds of language and reasoning tasks. Koa is intended to specialize more deeply in enterprise CRM operations where understanding process state, business rules, tool calls and the next correct action matters as much as generating fluent text.

Official source: Salesforce and NVIDIA announce Koa.

Why Salesforce built a CRM-specific reasoning model

Agentforce agents do not only answer questions. They can be asked to complete work that spans records, policies, tools and multiple decisions. A service agent may need to understand why a case exists, check account context, apply a policy, decide whether to escalate, update the case and then choose the next action. A sales agent may need to evaluate a lead, inspect opportunity context, qualify it and trigger a follow-up.

Salesforce’s argument is that a model trained specifically around these patterns can become better at the sequence of reasoning and tool use required to finish enterprise work, rather than simply knowing more generic facts about CRM.

In Salesforce’s own CRM benchmark, the company says Koa matches or exceeds leading model performance on CRM actions with three times fewer errors. That is a Salesforce-reported benchmark result, not an independent guarantee of performance in every production deployment. The useful takeaway is the type of problem Salesforce is optimizing for: correct action sequences in CRM workflows.

Salesforce describes benchmark tasks such as updating an opportunity, routing a case and scheduling a follow-up. Those are concrete business actions where an incorrect tool choice or skipped step can matter more than the quality of the prose around the answer.

How was Koa trained?

Salesforce says Koa was built by post-training NVIDIA Nemotron 3 Super with a proprietary synthetic dataset modeled on enterprise knowledge accumulated from nearly three decades of CRM deployments.

The training scenarios were designed to simulate enterprise workflows across more than 14 industries, including manufacturing, financial services, healthcare and travel. Salesforce says each scenario paired a persona with particular tasks and mapped the sequence of actions and tool calls required to complete them.

Salesforce says it used Supervised Fine-Tuning and reinforcement learning with Group Relative Policy Optimization, using NVIDIA NeMo RL, NeMo Gym and NeMo AutoModel. In practical terms, the goal was not simply to teach Koa how a CRM task should be described, but to improve its ability to select and execute the steps required to reach an outcome.

Training details come from Salesforce’s September 15, 2026 release.

Was Salesforce customer data used to train Koa?

Salesforce says no customer data was used to train Koa. According to the company, the training corpus was built entirely from synthetic scenarios that reflect the kinds of reasoning, tool use and decision-making Agentforce agents perform across CRM workflows.

This point matters because Salesforce has an enormous amount of customer CRM data, and “trained on decades of CRM intelligence” can easily be misread as “trained on customers’ CRM records.” Salesforce explicitly separates those ideas. The company says it used accumulated enterprise process knowledge to design synthetic training scenarios rather than feeding customer records into the model’s training corpus.

That does not mean customer data is irrelevant when Koa is used in production. An Agentforce agent can still reason over authorized business context at inference time. The training claim is narrower: Salesforce says the model itself was not trained on customer data.

Who can access Salesforce Koa right now?

As of September 15, 2026, Koa is available only to select pilot customers in Agentforce. Salesforce has not announced general self-service access for every Agentforce customer.

This is the most important access detail for an admin who reads about Koa, opens Salesforce Setup and cannot find it. The absence of a Koa option does not mean the org is misconfigured. The model is still in a controlled customer pilot.

Salesforce has not published general pricing for Koa in the launch announcement. Until the model reaches broader availability, organizations should not assume it is included automatically in every Agentforce entitlement or infer a separate price from the announcement alone.

Which companies are already piloting Koa?

Salesforce names six organizations as customer pilots in its announcement:

• 1-800Accountant
• Baxter Credit Union (BCU)
• Engine
• Formula 1
• UChicago Medicine
• Xero

Salesforce also says Koa is already being used internally, including in an agent in Slack that helps Salesforce employees find information and complete everyday tasks.

The pilot examples reinforce the intended workload. The quoted customer scenarios focus on navigating complex accounting context, member financial goals, business-travel service operations and healthcare workflows rather than generic chat or content creation.

When will Salesforce Koa become generally available?

Salesforce’s current target is winter 2026 in U.S. regions. The company does not give a more precise public GA date in the September 15 announcement.

That wording should be treated as a target, not as a promise that every Agentforce org will receive Koa on the first day of winter. Salesforce can still change release timing, supported editions, region coverage, administration controls or commercial terms before general availability.

There is also no announced global rollout schedule yet in the source material. If you operate outside U.S. regions, the launch statement does not support assuming that your region will receive Koa at the same time as the first GA release.

Availability statement: Koa is available to select pilot customers now; GA is expected winter 2026 in U.S. regions.

Does Koa appear in the normal Agentforce model selector?

Not in Salesforce’s current public model-selection documentation. The official Agentforce help page currently lists three standard model-option paths: Salesforce Default, AWS-Hosted and Google Gemini. Koa is not listed there as a normal option that every admin can select.

That does not mean Koa is disconnected from Agentforce. Salesforce explicitly says the pilot is running in Agentforce. It means the pilot experience is separate from the public, generally documented self-service model-selection menu.

So if you navigate to Setup → Audit, Analytics, and Monitoring → Einstein Audit, Analytics, and Monitoring Setup and do not see “Koa,” that is currently expected for a non-pilot org.

Current selector documentation: Salesforce Help: Select Agentforce Model Option.

What models can Agentforce admins currently select?

Salesforce’s current help documentation gives admins a managed set of model options rather than unrestricted access to every model available through Model Builder or the Models API.

The selected option applies to the Agentforce reasoning engine across agents, although Salesforce notes that particular tasks such as subagent classification or citations can still use Salesforce-owned models.

Salesforce also says custom actions that use prompt templates, Apex or the Models API can reference other Salesforce-managed or bring-your-own models. That broader model flexibility is different from the limited standard Agentforce reasoning-engine selector.

Official source: Select Agentforce Model Option. Google Gemini availability is also documented in Salesforce release notes.

Salesforce Default vs Koa: what is the difference?

Salesforce Default is a routing/configuration option; Koa is a specific Salesforce-hosted reasoning model. Those are not the same kind of thing.

Salesforce Default lets Salesforce choose the model setup it considers appropriate for Agentforce accuracy, trust and performance. The public help page currently says new Builder agents use GPT-4.1 and legacy Builder agents use GPT-4o under that option.

Koa, by contrast, is a model whose weights Salesforce says it controls and whose post-training was specifically targeted at CRM workflows. Its value proposition is specialization: internalized knowledge of CRM process structure and multistep enterprise tool use.

Salesforce has not yet documented whether GA Koa will become its own selectable model option, be incorporated into Salesforce Default routing, or be exposed through a different control. Until Salesforce publishes that configuration, any stronger claim would be speculation.

How Salesforce describes Koa’s trust boundary

Salesforce says it controls Koa’s model weights and performs both post-training and inference entirely within its own trust boundary. The company says no customer data crosses that trust boundary during Koa inference.

That is different from the standard external-model options described in Agentforce. For example, Salesforce’s AWS-hosted Anthropic option uses Claude Haiku 4.5 on a Salesforce-managed Amazon Bedrock instance, while the Gemini option uses Vertex AI. Salesforce says it applies zero-data-retention agreements with external model providers and has additional network and data protections for those managed deployments.

Koa’s positioning is more vertically integrated: Salesforce controls the weights, hosts the model within its own infrastructure and specializes it for its CRM domain.

This does not remove normal enterprise responsibilities around Salesforce permissions, record access, action authorization and agent design. A model running inside the trust boundary can still produce the wrong action if an agent is poorly configured or given inappropriate permissions.

What should Agentforce admins do before Koa reaches GA?

For most organizations, there is no Koa switch to turn on today. The useful work is preparation rather than trying to force access that Salesforce has not made generally available.

1. Confirm whether your organization is actually in the Koa pilot. If not, do not expect a Koa selector in standard Setup.
2. Document your current Agentforce model option. Know whether your org uses Salesforce Default, AWS-hosted Claude or Google Gemini.
3. Identify CRM workflows where domain reasoning matters most. Long, multistep cases with record updates, policy checks and several tool calls are more relevant evaluation targets than simple FAQ responses.
4. Build a regression set now. Save representative tasks and expected outcomes so you can compare Koa with the model configuration you already use if access becomes available.
5. Do not assume the benchmark equals your production result. Test your own data model, actions, permissions and workflow complexity.
6. Watch Salesforce’s U.S. winter 2026 release guidance. Edition support, pricing, region availability and final administration controls have not all been published yet.