Salesforce Announces Koa: The First CRM Reasoning Model Built on NVIDIA Nemotron
Salesforce and NVIDIA have announced Salesforce Koa, described as Salesforce’s first CRM reasoning model for Agentforce. Rather than being a general-purpose AI model retrofitted for business use, Koa was built specifically to reason through the kind of multi-step work that fills a CRM: qualifying a lead, updating an opportunity, routing a service case.
This matters because most AI models are trained to sound convincing, not to reliably follow a business process from start to finish. Koa is Salesforce’s attempt to close that gap. This article explains what Koa is, how it was built, what Salesforce and independent research say about its performance, and what to realistically expect at this stage.
What Salesforce Actually Announced
Salesforce says Koa was developed through collaboration with NVIDIA, built by post-training NVIDIA’s open-weight Nemotron 3 Super model. According to Salesforce’s official announcement, Koa is a specialized reasoning engine for Agentforce, Salesforce’s platform for building AI agents.
The distinction Salesforce draws is between a model that generates plausible text and one that can reason through a task, decide which tool to use, and carry out a sequence of steps correctly. Salesforce says Koa is trained and evaluated specifically on enterprise CRM workflows rather than general internet content.
Why Build a CRM-Specific Reasoning Model at All
General AI applications are trained using publicly available datasets. Such systems usually do well in conversational tasks and summarization tasks; however, a customer relationship management flow is not a conversation. It is rather a process of checking something, applying some policy, deciding about escalation, taking action, and confirming the outcome.
Salesforce’s argument is that CRM usage patterns, how sales, service, and operations teams actually structure their work, are a distinct kind of knowledge that a general model doesn’t necessarily absorb well. Salesforce says Koa’s training data was built from synthetic scenarios modeled on nearly three decades of CRM deployments across more than 14 industries, including manufacturing, financial services, healthcare, and travel.
What NVIDIA Nemotron Contributes
Unlike most LLMs, Koa does not use a clean slate model. It takes NVIDIA Nemotron 3 Super, which is an open-weight foundation model, and uses it as the base that Salesforce customizes for its CRM application. The use of the open-weight foundation allows Salesforce to control the weights and run the training and inference processes within its infrastructure.
The difference is important. Salesforce sees the data isolation provided by Koa as one of the direct advantages of using this approach, since it makes it possible to post-train the model with synthetic scenarios and run it in production without sharing customer data with the model provider.
How Koa Was Trained, and Why Synthetic Data Matters
Salesforce says no customer data was used to train Koa. Instead, Salesforce built a synthetic dataset of enterprise scenarios, simulated CRM situations paired with the sequence of actions and tool calls a Koa reasoning model would need to complete them correctly.
The independent technical paper on arXiv is more specific. It explains that Koa was post-trained using supervised fine-tuning followed by reinforcement learning with Group Relative Policy Optimization, commonly abbreviated as GRPO. In plain terms, GRPO trains a model by generating multiple attempts at a task and rewarding the ones that actually complete it correctly, rather than rewarding text that merely sounds right.
The paper also describes a “simulation-to-reward pipeline,” where workflow specifications, written in Salesforce’s Agent Script language for enterprise domains, get expanded into multi-turn tasks tied to a specific persona. The model is scored on whether it successfully used the right tools to resolve the task, not just whether its answer looked reasonable.
Where Koa Fits Inside Agentforce
Salesforce Koa isn’t a standalone product that customers interact with directly. It’s positioned as a reasoning layer that can sit underneath Agentforce agents, helping them decide what to do next in a workflow and which tool or action to invoke.
In order to illustrate the above idea, one may refer to several examples of how CRM logic might be applied in practice:
- Service agent faces complicated customer request, takes into account customer history and current entitlements, and then makes a decision about resolving the issue personally or escalating it to human
- Sales agent is looking at a new lead, compares it against qualification rules, and changes the stage of the deal
- Operations agent is required to schedule follow-up activity based on customer availability and internal capacity
These are illustrative examples of the kind of multi-step reasoning Salesforce says Koa is designed for, not documented customer outcomes.
General Reasoning Versus CRM-Specific Reasoning
| Area | General AI Approach | CRM Reasoning Approach |
| Task understanding | Broad, general-purpose reasoning | Reasoning shaped around CRM workflow patterns |
| Tool use | Wide but shallow tool capability | Actions tied to specific business processes |
| Context | General world knowledge | CRM records, policies, and workflow state |
| Decision path | Flexible, less predictable | More structured around enterprise task boundaries |
| Human handoff | Depends entirely on implementation | Can be designed around defined workflow checkpoints |
This table simplifies a genuinely complex comparison, but it captures the core idea: a CRM reasoning model is being optimized for a narrower, more predictable set of tasks rather than open-ended conversation.
What Salesforce Says About Performance, and What That Claim Actually Means
According to Salesforce, when tested using its own benchmark, Koa achieves at least comparable and better model performance than the leading models, while reducing errors threefold for the same types of tasks such as updating the status of an opportunity or routing the case. This claim should definitely be read closely, since it was produced through Salesforce’s own testing process of Koa.
The arXiv paper gives a more accurate portrayal of Koa. It describes how the new Koa model outperforms its base, the open-weight Nemotron, significantly and most notably in multi-turn tool use and that Koa performs better than the best proprietary baseline across the benchmarks used. Most importantly, it also mentions how Koa does not reach the level of frontier models, even though they perform better than all the others across the benchmarks used.
Current Availability and What to Expect
Salesforce says Koa is already running internally, including inside a Slack-based agent used by employees, and is now moving into pilot programs with a small number of named customers, including 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero.
For the general market, Salesforce believes that the general availability will be in winter 2026 in the U.S. regions. Most Salesforce clients do not have access to Koa at present. It makes sense to consider any performance claims as results from a small-scale pilot program.
What Salesforce Teams Should Consider Before Adopting Koa
Before planning around Koa, a few practical questions are worth working through internally:
- Which specific workflows would actually benefit from structured, multi-step reasoning versus simple automation rules
- What data and tool access an Agentforce agent would need to reason through a given process end to end
- How human handoff points should be defined for higher-risk or ambiguous decisions
- Whether existing Agentforce implementations are structured in a way that could take advantage of a reasoning layer once it’s generally available
None of this requires immediate action. It’s groundwork that makes adoption smoother once Koa reaches general availability.
Conclusion
The Salesforce Koa is an example of what kind of risk they are taking when they decide to create a model tailored for CRM workflows which have their own unique features that set them apart from generic text generation and require a unique reasoning model. The benchmarks created by Salesforce itself and the research article prove that point.
For now, Koa remains in limited pilot use, with wider availability expected later this year. Salesforce customers don’t need to make immediate decisions, but reviewing current Agentforce workflows, data readiness, and where multi-step reasoning could genuinely help is a reasonable use of the time before general availability arrives.
FAQs
What is Salesforce Koa?
Salesforce Koa is the first-ever CRM-specific reasoning model from Salesforce that works for Agentforce in partnership with NVIDIA.
On what is Koa based?
Koa is trained after training of the NVIDIA’s open-weight Nemotron 3 Super model using synthetic enterprise scenarios instead of customer data.
Does Salesforce use customer data in the training of Koa?
No. According to Salesforce, Koa is trained completely on synthetic scenarios that simulate CRM deployment patterns and not on customer data.
How does Koa function with Agentforce?
Salesforce Koa works as a reasoning model for Agentforce agents and helps them understand the action to be taken and tool to use at every step of the process.
When would Koa become generally available?
As per Salesforce, Koa is currently available for selected pilot customers and will become generally available in Winter 2026 in U.S. Regions.
For more insights, updates, and expert tips, follow us on LinkedIn.
