+1-480-241-8198
+44-7428758945
+61-1300-332-888
+91 9811400594Salesforce has spent the last three years adding AI capabilities to almost every part of its platform. Most of that work has involved connecting existing foundation models, Claude, GPT, Gemini, to Salesforce data and workflows through Agentforce. Salesforce Koa brings a new point of curiosity to this evolving landscape. It is Salesforce’s first CRM reasoning model built from the ground up, trained specifically for CRM reasoning rather than general intelligence. What is it and what makes it worth being looked at in the future? From its capabilities to performance benchmarks and piloting, there are many aspects to consider.
What is Salesforce Koa?
Salesforce and NVIDIA unveiled Koa, Salesforce’s very first reasoning model made to serve the purposes of Agentforce, at Dreamforce on September 15, 2026. It differs from a generic model in that it solves problems step-by-step until the point at which it generates its output. It is significant for CRM specifically because most crucial operations in it are not performed in just one step. Updating an opportunity requires checking pipeline context. Routing a case requires evaluating priority rules, ownership logic, and entitlement history. Scheduling a follow-up requires reading prior interaction data and applying business rules about timing.
The reasoning model was created using the foundation model called Nemotron-3-Super-120B provided by NVIDIA, which is built around a Mixture-of-Experts architecture. In this kind of architecture, a certain set of model’s parameters that is required to perform the operation is activated, while the whole model activation for each request would take much more time. The model includes a context window size of one million tokens, which makes it possible to process the entire account history, conversation thread, and workflow context at once without maintaining everything from the start.
Salesforce trained the model on synthetic CRM data for 27 years in 14 different industries. The post-training used reinforcement learning from Group Relative Policy Optimization — a technique that rewards correct multi-step reasoning rather than just correct final outputs. The result is a model that is not trying to be the best AI in general. It is trying to be the best AI for the specific class of tasks Salesforce customers run every day.
Salesforce Koa Features
CRM-native reasoning
Koa understands Salesforce’s data model structure because it has been trained on the same basis. It recognizes the concepts of object relationships, field dependencies, approval logic, and sequencing of workflows in their true sense.
Stays inside the trust boundary
Salesforce controls the model weights entirely. Post-training and inference happen within Salesforce’s own infrastructure, which means customer data does not leave the Salesforce trust boundary during model calls. For organizations with data residency requirements or sovereignty concerns, this is the material difference between Koa and models accessed through third-party APIs.
Three deployment options
The Salesforce AI model can be selected at three levels inside the platform. Org-wide in Agentforce Setup, making it the default model for every agent in the organization. As a managed large language model in Data Cloud’s generative models catalogue. Or at the per-agent and sub-agent level within the Agent Builder, where it can be implemented on a particular workflow, while other models perform their separate functions. The fact that it is implemented on a per-agent level implies that a mixed modeling architecture is not an exception but a rule.
Works across any agent type
Any Agentforce agent, sub-agent, or agent router can use Koa. It is not restricted to a specific cloud or use case category.
Koa Benchmarks: What the Numbers Actually Say
Koa benchmarks show a model that beats where it is designed to compete and is honest about where it does not. On Salesforce’s own CRM benchmark, which evaluates tasks like updating opportunities, routing cases, and scheduling follow-ups rather than general reasoning, Koa produces roughly two times more reliable results than a general-purpose model. It also outperforms the Nemotron base model it was derived from, particularly on multi-turn tool-use scenarios where the agent needs to make multiple model calls to complete a workflow.
There is a useful distinction to consider here. Salesforce designed Koa for CRM-related tasks, which is why the strongest results are obtained in tasks related to CRM workflow specialization. The published benchmark scores also show that models such as Claude Opus 4.8 and GPT-5.5 perform strongly across the evaluated categories. That does not take away from Koa’s results. General-purpose models bring broad capabilities to the table; Koa has been trained to handle the decisions, tool calls, and workflow steps that Salesforce agents encounter. The benchmark is also vendor-reported and uses Salesforce’s own CRM task set, so independent validation would provide another useful point of comparison.
That is not a weakness in the product. It is the product’s entire point. Koa is not competing with GPT-5.5 on general intelligence. It is competing with what a general model does when asked to handle a CRM workflow, and on that narrower comparison, the two-times-reliability claim holds.
Koa vs Agentforce: Understanding the Relationship
Koa vs Agentforce is not a competition. Koa is a model that runs inside Agentforce, a reasoning engine option, not a separate product. Before Koa, Agentforce used what Salesforce calls the Default mix — currently GPT-4.1 for agents built in the new Agent Builder and GPT-4o for agents in the legacy builder. Claude on Amazon Bedrock and Google Gemini on Vertex AI are also available as model options. Koa becomes the fourth option in Agentforce Setup.
The practical implication is model routing. Salesforce’s AI gateway can direct routine, lower-complexity tasks to Koa, faster and cheaper to run within the trust boundary, while directing harder reasoning tasks to a frontier model when the use case justifies it. That routing logic is what makes Koa economically interesting beyond the accuracy claim: it gives Agentforce customers a cost-efficient, data-sovereign option for the high-volume CRM tasks that currently consume frontier model capacity.
Salesforce Koa Pilot: How to Participate
The Salesforce Koa pilot is currently active with a select group of customers, including 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero. General availability is targeted for US regions in Winter 2026, with an open beta expected shortly after.
Who should go for the pilot test:
Organizations that currently use Agentforce in US regions and have data residency/sovereignty so that customer data cannot be taken out of the Salesforce trust boundary. For such clients, Salesforce Koa remains the only available solution for using Salesforce-hosted reasoning without going through a third-party provider.
How access works:
Koa is not a separate product purchase. There is no published Koa-specific pricing — access is consumed through existing Agentforce billing. When the pilot opens more broadly and reaches general availability, enabling it is an opt-in through Agentforce setup rather than a new contract.
What to prepare before enabling Koa:
The model selection in Agentforce operates at three levels — org-wide, per agent, and per sub-agent. Prior to adopting Koa, identify the agent workflows that have the CRM-specific activities for which Koa’s training is most applicable: generating leads, qualifying opportunities, and resolving service cases across the customer lifecycle. For agents handling tasks outside that scope, keeping the Default mix or a frontier model in place is the right call.
What can you expect in the pilot project:
Pilot customers are currently using a model that has not been made available to all users yet. Feedback during this phase directly influences how Koa develops before the Winter 2026 release. Testing across high-volume, repeatable CRM workflows will produce the most useful signal — particularly for multi-turn tool-use scenarios where the error reduction claim is strongest.
Conclusion
Salesforce Koa is not a general AI model that happens to work in a CRM. It is a CRM reasoning model that happens to be built on one of the most capable foundation architectures available. The distinction shapes everything about how it should be evaluated, deployed, and measured.
The benchmarks are honest about what Koa is good at and where frontier models still lead. The trust boundary is really unique in terms of data sovereignty needs of organizations. The pilot is active, the GA schedule is defined, and the access point through Agentforce billing removes the procurement hurdle that normally hinders enterprise AI adoption. For Agentforce customers in US regions, the question is which workflows to test it on first.
About Author
Akanksha is a technical content writer with 3+ years of experience creating content for IT services, software development, and Salesforce consulting companies. She specializes in SEO-focused content, technical storytelling, and marketing assets that drive visibility, engagement, and business growth.
Share this post on:
