Introduction to Salesforce Koa: The CRM Reasoning Model Powering Agentforce

What happens when an AI agent knows how to answer a CRM question but cannot work through the business process needed to complete it? A sales agent may need to review an opportunity, check account history, interpret pricing rules, update several records, schedule a follow-up, and then notify the right person.

A service agent may need to understand a customer’s history, apply company policy, route a case, and trigger the next action. These are not single-prompt tasks. They require context, reasoning, tool use, and a sequence of decisions.

That is the problem Salesforce is targeting Koa, its first CRM reasoning model for Agentforce, built on NVIDIA Nemotron. Salesforce says Koa is trained using 27 years of CRM intelligence and designed specifically to help agents reason through complex, multi-step enterprise work.

Salesforce’s latest research shows how quickly AI agents are moving into enterprise workflows, while the company is shifting Agentforce from simple conversational assistance towards agents that can take actions across business processes.

Rather than being another general-purpose language model, it is purpose-built for CRM work, including tasks such as updating opportunities, routing service cases, scheduling follow-ups, and working through customer lifecycle processes.

So, what exactly is Salesforce Koa and how does the Salesforce Koa model differ from a general AI model? Where does it fit inside Agentforce? Let’s find out.

What Is Salesforce Koa?

Salesforce Koa is Salesforce’s first CRM reasoning model for Agentforce, developed with NVIDIA and built on NVIDIA Nemotron 3 Super. The important word here is reasoning.

A traditional generative AI model may produce a useful response to a customer question. A CRM reasoning model needs to understand what should happen next and which tools or actions are required to complete the task.

Completing that request can involve several steps:

1. Retrieve the opportunity.
2. Review recent emails, activities, and customer interactions.
3. Understand the opportunity stage and business context.
4. Determine the appropriate next action.
5. Update the opportunity if required.
6. Check the user’s availability.
7. Schedule the follow-up.
8. Confirm what was completed.

Koa is designed for this type of multi-step CRM work rather than simply generating text. Salesforce describes Koa as a model grounded in enterprise processes, workflows, and operational policies, with the goal of helping Agentforce agents reason through tasks and use the right tools to complete them.

Why Salesforce Built a CRM-Specific Reasoning Model

General-purpose AI models are trained to handle an enormous range of tasks. That flexibility is valuable, but enterprise CRM work has its own requirements.

A CRM agent operates within defined processes. A sales opportunity has a lifecycle. A service case has routing rules. A customer account has permissions and policies. A discount may require approval. A follow-up may need to be assigned to a specific employee.

The AI therefore needs to understand more than language, and it needs to understand how work gets done.

Salesforce says, “Koa was developed by post-training NVIDIA Nemotron using a proprietary synthetic dataset based on nearly three decades of CRM deployments.” The training scenarios model business processes, workflows, operational policies, reasoning, tool use, and decision-making across CRM tasks.

This creates an important distinction:

That distinction becomes particularly important when Agentforce is expected to move from answering questions to performing work.

How Salesforce Koa Is Trained for Enterprise CRM Work

Koa’s training approach is one of its defining characteristics. Salesforce says its training corpus is built entirely from synthetic scenarios, rather than customer data. These scenarios simulate the reasoning, tool use, and decision-making required by Agentforce agents across the customer lifecycle.

The scenarios cover 14+ industries, including:

• Manufacturing
• Financial services
• Healthcare
• Travel

Each scenario combines a persona with a specific business task and maps the sequence of actions required to complete it.

Consider a travel company. A customer may ask an agent to change a booking. The agent may need to identify the booking, check the fare rules, verify availability, calculate any applicable difference, update the reservation, and communicate the new itinerary.

That is very different from answering, “What is your cancellation policy?”

The second is primarily an information request. The first requires reasoning across a process. Koa is designed for the second category.

Koa and Agentforce: Where Does It Fit?

The Salesforce Agentforce Koa relationship is straightforward: Koa is a reasoning model option designed to power Agentforce agents.

Agentforce provides the agent layer, including the instructions, actions, tools, data, and business context required to complete a task. Koa provides the reasoning capability that helps determine how the agent should work through that task.

For example, a customer service agent could receive:

“The customer says their replacement order has not arrived. Find the order, check its current status, review previous cases, and determine what should happen next.”

An Agentforce agent needs access to the relevant records and actions. Koa is designed to reason through the sequence required to reach the appropriate outcome.

Salesforce says Koa can be selected at multiple levels. It can be used in Agentforce Builder for an agent or sub-agent, selected as an organization-wide model provider, or accessed as a managed LLM through the Data Cloud generative models catalogue.

What Makes Koa Different From a General AI Model?

The biggest difference is specialization. A general-purpose model needs to reason about many types of work. Koa has been specifically post-trained for enterprise CRM scenarios. Salesforce says Koa is built around:

• CRM workflows
• Enterprise processes
• Operational policies
• Tool use
• Multi-step reasoning
• Customer lifecycle tasks
• Agent actions

This is particularly relevant for organizations where an AI agent needs to take controlled actions rather than simply generate a response.

Imagine a financial services customer asking, “I want to change my repayment arrangement. What are my options?”

A useful agent needs to understand the customer’s context, identify applicable policies, retrieve the right information, determine what actions are available, and potentially route the request to the appropriate team.

The value of a CRM reasoning model is its ability to work through that process within the boundaries established by the organization.

Salesforce Koa Performance: What the Early Results Show

Salesforce is using itself as “customer zero” to test Koa across CRM use cases, including help agents, employee agents, event agents, and web agents. Its CRM Bench includes real-world tasks such as updating opportunities, routing cases, and scheduling follow-ups.

According to Salesforce’s published results, Koa:

• Is 11% more precise at selecting the appropriate action.
• Provides 2.1x greater reliability in recalling customer context.
• Performs 15% better at retaining context during long back-and-forth conversations.
• Matches or exceeds leading model performance on CRM actions with three times fewer errors in Salesforce’s CRM benchmark.

These figures are Salesforce’s own benchmark and comparison results, so organizations evaluating Salesforce Koa benefits should consider the specific workloads, evaluation methodology, and production conditions relevant to their environment.

Still, the direction is important. For enterprise AI, fewer wrong actions can matter more than simply producing a fluent answer. If an agent incorrectly routes a case, updates the wrong opportunity, or misses an important step in a workflow, the business impact can be much greater than a poorly worded response.

Real-World Use Cases And Benefits of Salesforce Koa

1. Sales Opportunity Management

Sales teams spend significant time updating CRM records and coordinating follow-ups.

A Koa-powered Agentforce agent could help work through tasks such as reviewing an opportunity, identifying missing information, preparing the next action, or scheduling follow-ups.

For example, after a customer meeting, a seller could ask an employee agent to review the opportunity history, identify outstanding actions, update relevant information, and create a follow-up task.

The objective is not simply to summarize the meeting. It is to reduce the manual steps required to keep the CRM current.

2. Customer Service Case Resolution

Service cases often involve several connected records and policies. A customer might contact support about a delayed shipment.

The agent could need to review the order, previous interactions, shipping status, account information, and applicable service policies before determining the appropriate next step. Koa’s focus on multi-step CRM reasoning is designed for precisely this type of workflow.

3. Healthcare Operations

Healthcare provides a strong example because operational workflows can involve many connected steps.

UChicago Medicine is among the organizations moving into customer pilots with Koa. Salesforce highlights the potential for Koa to help with longer, multi-step workflows where information coordination and getting the right next action at the right time are important.

Importantly, this does not mean replacing clinical decision-making with an AI model. The example is more relevant to operational coordination, where agents can assist with structured, non-clinical workflows.

4. Financial Services

Financial services workflows often combine customer information, policies, eligibility requirements, documentation, and multiple operational systems.

Baxter Credit Union is among Koa’s customer pilots. Salesforce says its digital agents can use Koa to understand the context behind member goals and reason across information, tools, and policies.

For example, a member asking about buying a home may require an agent to connect the conversation with account context, relevant financial information, applicable processes, and the next step.

5. Accounting and Tax Services

Accounting workflows can involve tax rules, financial information, documents, and individual customer circumstances.

1-800 Accountant is another organization piloting Koa. Salesforce says the model can help agents work through these complexities step by step and use the appropriate tools along the way.

This illustrates where specialized reasoning can become useful: the agent needs to understand the workflow rather than simply produce a generic accounting response.

6. Business Travel

Business travel involves multiple moving parts, including itineraries, bookings, traveller preferences, changes, policies, and service requests.

Engine, a business travel company, is also among Koa’s customer pilots. Salesforce cites its interest in using reasoning that can work through complex, multi-step travel problems rather than simply generate confident-sounding answers.

A practical example could involve changing a traveler’s flight while considering booking rules, availability, timing, and related itinerary details.

Koa and Data Security: Keeping the Model Inside the Trust Boundary

For enterprise decision-makers, model performance is only one consideration. Data governance is equally important. Salesforce says no customer data was used to train Koa. Instead, its training corpus was built from synthetic scenarios representing CRM workflows.

Salesforce also controls Koa’s model weights and runs post-training and inference within its own infrastructure. The company says this keeps customer data inside the Salesforce trust boundary during inference.

Koa is also designed to work with the context an organization chooses to provide. Salesforce says the model only sees the data, records, grounding information, and instructions made available to it.

An enterprise evaluating an AI reasoning model should ask:

• Where does the model run?
• What data is used for training?
• What customer data is exposed during inference?
• How are permissions enforced?
• What controls govern agent actions?
• Can the organization control which context the agent receives?

Koa’s Salesforce-hosted approach is designed to address these enterprise considerations.

How Can Organizations Deploy Salesforce Koa?

Salesforce provides several ways to use Koa within its AI environment.

Agent and Sub-Agent Level

Koa can be selected within Agentforce Builder for individual agents and sub-agents. This is useful when a company wants different reasoning models for different workloads.

For example, a service agent could use Koa for complex case workflows, while another application uses a different model for a more general content-generation task.

Organization-Wide Model Provider

Koa can also be selected as a model provider at the organization level. This gives Salesforce administrators a broader way to establish which model powers Agentforce across the organization, subject to the available configuration and rollout requirements.

Managed LLM

Koa is also available as a managed LLM through the Data Cloud generative models catalogue. That means organizations can use it within Salesforce’s managed AI environment rather than treating it as an entirely separate external model deployment.

Salesforce Koa vs Claude: What Should Enterprises Consider?

Decision makers asking for Salesforce vs. Claude will naturally raise an important question: which model should an enterprise use?

The answer depends on the workload rather than simply choosing a model based on a headline benchmark.

Claude and Koa serve different strategic purposes within Salesforce’s ecosystem. Salesforce’s current Agentforce documentation supports multiple model providers, including Anthropic models hosted through Amazon Bedrock as well as Salesforce-managed models.

Koa’s differentiator is its CRM-specific specialization. It has been post-trained around Salesforce workflows, enterprise CRM tasks, tool use, and operational processes.

For a CRM-heavy workflow such as opportunity updates, case routing, or follow-up scheduling, that specialisation may be particularly relevant. For broader workloads, an organization may have different requirements.

A sensible evaluation should therefore compare:

The important point is that model selection should follow the business workload.

What Koa Means for the Future of Agentforce

Koa signals a broader shift in enterprise AI. For CIOs, CISOs, and enterprise architects, it introduces important considerations beyond model performance, including where reasoning models fit within the AI architecture and how agent actions should be governed.

• More specialized AI agents: Agentforce can use CRM-focused reasoning for complex sales, service, and customer lifecycle workflows.
• Greater task automation: Agents can move beyond answering questions to completing multi-step CRM actions.
• More contextual decisions: Agents can reason across CRM data, business processes, policies, and available tools.
• Stronger enterprise governance: Organizations will need clear permissions, approval rules, action limits, and human escalation paths.
• AI-ready CRM architecture: Clean data, connected systems, APIs, and well-defined workflows will become increasingly important for reliable agent performance.
• Outcome-focused AI adoption: Businesses can measure AI by completed tasks, reduced manual effort, faster service, improved productivity, and other measurable results.

Final Thoughts

Salesforce Koa represents a move towards purpose-built enterprise reasoning, rather than relying exclusively on general-purpose models for every CRM task.

Its value comes from the combination of Salesforce CRM intelligence, NVIDIA Nemotron, synthetic enterprise scenarios, Agentforce tools, and Salesforce’s trust boundary. For decision-makers, Koa should therefore be evaluated as part of an Agentforce architecture, not simply as another LLM.

The practical opportunity is to identify high-value, repeatable workflows where reasoning and tool use can reduce manual effort while maintaining governance and human oversight where required.

Introducing a reasoning model is only one part of building an enterprise-ready AI workforce. Organizations also need clean CRM data, well-defined processes, secure integrations, effective agent actions, appropriate governance, and a clear business case.

At Ksolves, we help organizations design and implement Salesforce solutions across Agentforce, CRM, integrations, automation, data, and custom development. Our Salesforce implementation services can help identify suitable agentic use cases, assess your existing architecture, configure Agentforce, and prepare the underlying CRM environment for AI-powered workflows.

FAQs

1. What is Salesforce Koa?

Salesforce Koa is Salesforce’s first CRM reasoning model for Agentforce, built on NVIDIA Nemotron. It is designed to reason through complex, multi-step CRM tasks and use the appropriate tools to complete them.

2. How is Koa different from a general AI model?

Koa is purpose-built for enterprise CRM work. Its post-training uses synthetic scenarios based on CRM processes, workflows, operational policies, tool use, and decision-making across more than 14 industries.

3. Does Salesforce Koa use customer data for training?

No. Salesforce states that no customer data is used to train Koa. Its training corpus was created entirely from synthetic scenarios representing CRM workflows.

4. Where can Salesforce Koa be used?

Koa can be selected in Agentforce at the agent and sub-agent level, configured as an organization-wide model provider, and accessed as a managed LLM through the Data Cloud generative models catalogue.

5. Is Salesforce Koa generally available?

As of September 2026, Koa is available to select pilot customers in Agentforce, with Salesforce stating that general availability is expected in U.S. regions in Winter 2026. Availability can change by region and program.