Salesforce and Nvidia Unveil Koa, an Open-Weight Reasoning Model Built for Enterprise Work

Salesforce and Nvidia have introduced Koa, a new reasoning model designed to handle enterprise tasks such as sales, marketing and customer support, marking a direct challenge to closed AI systems from OpenAI and Anthropic. The model is notable not only because it is Salesforce’s first reasoning model, but because it was built on Nvidia’s open-weight Nemotron base and trained without using actual customer data, a move aimed at lowering risk, reducing costs and giving businesses more control over how AI is deployed.

The launch, announced during Salesforce’s Dreamforce conference on September 15, 2026, underscores a broader shift in enterprise AI: companies increasingly want models that are specialized, secure, token-efficient and easier to govern than general-purpose frontier systems. Salesforce says Koa will sit inside its Agentforce platform as an alternative for customers who need deeper reasoning in multi-step workflows, while still preserving the company’s data safeguards.

What Salesforce is really trying to change with Koa

Koa is Salesforce’s attempt to own a part of enterprise AI that has so far been dominated by frontier model providers. Until now, when an agent built on Salesforce’s platform needed to reason through a complex request, the system could hand the task off to models such as Claude or ChatGPT through Salesforce’s AI gateway.

The new model changes that arrangement. Rather than relying entirely on outside labs for advanced reasoning, Salesforce now has its own enterprise-focused option built for the kinds of repetitive but high-value work that dominate customer operations.

That makes Koa less about abstract benchmarks and more about practical business utility. In Salesforce’s telling, the model is intended to help agents respond to support cases, move sales conversations forward and handle common operational tasks with fewer tokens, less latency and tighter compliance controls.

Why the enterprise AI market is pulling away from frontier labs

The launch points to an increasingly visible divide between what the biggest AI labs are optimizing for and what corporate buyers say they need. Frontier companies often showcase models that can handle math, coding, broad knowledge and open-ended reasoning. Enterprises, by contrast, are usually focused on narrower goals: reliability, data security, cost predictability, auditability and domain performance.

Salesforce is betting that the second group will increasingly want models shaped around specific business workflows rather than one-size-fits-all general intelligence. The company’s message is that enterprise customers do not necessarily need the most dazzling general model; they need the model that fits inside their systems, respects their data boundaries and completes work efficiently.

That is also why Koa is positioned as an “open-weight” alternative to closed systems. Open-weight models give vendors and customers more flexibility in deployment, tuning and governance, while still benefiting from a strong base model underneath.

How Koa was built

Koa was developed with Nvidia’s Nemotron model as its foundation, then post-trained for enterprise use cases. Post-training means taking a more general model and adapting it to excel at specific tasks, in this case work associated with sales, service and marketing.

Salesforce says the model was not trained on any actual customer records. Instead, the companies generated synthetic data designed to imitate the patterns of real business interactions. That included simulated environments with customer-service personas, frustrated callers and sales professionals trying to close deals.

The aim is to make the model familiar with the shape of enterprise work without exposing customer information or creating the kind of data governance concerns that often slow adoption of AI in regulated environments.

Why synthetic data matters

Synthetic data gives Salesforce a way to teach the model useful patterns while avoiding direct ingestion of customer files, support tickets or internal notes. For enterprise buyers, that matters because one of the biggest objections to AI deployment is the fear that sensitive information might be absorbed into a model and later exposed elsewhere.

By keeping customer data out of the training process, Salesforce can claim a cleaner security posture and a clearer path for regulated customers who need strong boundaries around where data lives and how it is used.

Salesforce AI executive Jayesh Govindarajan said the company had long relied on frontier model vendors for reasoning, but that capability is now being brought in-house for the first time through Koa.

Govindarajan also argued that the breakthrough came from Nvidia’s Nemotron family, which he described as a pre-trained base model with the combination of sovereignty, quality and data provenance that Salesforce had been waiting for. He contrasted that with other open-weight models, saying provenance and transparency had been a major limitation in the market.

How Koa fits into Agentforce

Koa is being added as an option inside Agentforce, Salesforce’s platform for building AI agents that automate repetitive work. That includes tasks such as answering customer inquiries, scheduling appointments and helping employees or customers complete structured workflows.

In practical terms, Agentforce uses an AI gateway that routes a request to the best model for the job. With Koa in the mix, Salesforce can keep more reasoning tasks inside its own environment instead of sending them to an outside model every time a workflow becomes complicated.

This matters because enterprise AI systems do not usually rely on a single model for everything. They often blend smaller task-specific models with larger reasoning engines, choosing among them based on speed, cost, security and task complexity. Koa broadens Salesforce’s menu of options in that architecture.

What changes for customer workflows?

For customers, the most immediate change is choice. If a task is simple, a smaller model may still be enough. If the task requires more nuanced reasoning, Koa can now serve as the in-house alternative before work gets routed to a third-party frontier model.

That gives companies another way to balance performance and expense. Salesforce says Koa should complete certain jobs using fewer tokens than Claude or ChatGPT, which could lower AI operating costs at scale.

Why Nvidia matters here

Nvidia’s role goes beyond supplying infrastructure. The company is increasingly shaping the software stack around enterprise AI, not just the chips that power it. In Koa, Nvidia is contributing a model architecture that Salesforce says is efficient at inference and better suited to token-conscious enterprise deployment.

Kari Ann Briski, Nvidia’s vice president of generative AI software for enterprise, said the appeal lies in a combination of sovereign AI, faster response times and better token economics. In other words, the model is designed not just to work, but to work efficiently inside enterprise systems where every API call and token counts.

That efficiency message is important. As more companies move from experimentation to production, the cost of running models becomes a major issue. A model that can do the same work with fewer tokens can materially change the economics of AI adoption.

What Salesforce is not giving up

Even with Koa, Salesforce is not cutting ties with Anthropic or OpenAI. Instead, it is broadening its options. The company also announced a new partnership with Anthropic called ClaudeForce, which will allow businesses to use Claude as their AI interface while keeping their data inside Salesforce’s system of record.

That dual approach suggests Salesforce sees no single model provider winning enterprise AI outright. Instead, the company appears to be building a platform that can accommodate different models depending on customer needs, regulatory constraints and workflow complexity.

It is a pragmatic strategy. Some customers will prefer the familiarity and power of frontier models. Others will prioritize sovereignty, lower cost or tighter integration with their internal systems. Salesforce wants to serve all of them without forcing a one-model-fits-all decision.

Govindarajan said Salesforce had wanted to build an enterprise-grade frontier model earlier, but lacked a suitable pre-trained foundation that met its standards for sovereignty and provenance.

How the model could affect AI competition

Koa may not be aimed at consumer buzz, but it could influence a far more lucrative segment of the AI market: enterprise deployments. If Salesforce can show that domain-tuned reasoning models are cheaper, safer and “good enough” for many business tasks, it could pressure the largest model labs to adapt their enterprise offers.

The strategic implication is significant. Frontier labs have often assumed that enterprises will accept the cost and complexity of using the biggest models available. Salesforce’s move argues the opposite: enterprises may increasingly prefer specialized systems that are built around their business processes from the start.

If that view spreads, the next phase of AI competition could be less about who has the most capable general model and more about who can deliver the most useful, governable and cost-effective one for specific industries and workflows.

What enterprises are likely to care about most

• Security: whether sensitive information stays inside approved systems.
• Cost: how many tokens are consumed per task.
• Latency: how quickly the model returns useful output.
• Control: whether the company can route and govern model use.
• Data provenance: whether the model’s training history is transparent.

Why the timing matters at Dreamforce

Salesforce used Dreamforce, its annual marquee event, to spotlight Koa because the conference is one of the few places where the company can frame AI as a live product strategy rather than an abstract research direction. The event gives Salesforce a stage to show customers how its platform is changing and why those changes matter now.

Dreamforce announcements also tend to signal the company’s commercial priorities for the next year. By placing Koa near the center of the conference, Salesforce is indicating that enterprise reasoning is no longer a side project. It is becoming a core feature of the company’s AI roadmap.

How Koa compares with general-purpose AI models

Koa is not trying to be everything to everyone. That is one of its main advantages. General-purpose frontier models are built to perform well across many domains, but enterprise buyers often end up paying for broad capability they do not fully use.

Salesforce and Nvidia are making the case that a narrower, business-trained reasoning model can deliver more value in the contexts that matter most. If a support agent, a sales assistant or a workflow coordinator can do its job with fewer tokens and stronger privacy guarantees, the trade-off may be worth it.

In that sense, Koa reflects a larger rethinking in AI procurement: companies are moving from “what is the smartest model?” to “what is the right model for this task, this data policy and this budget?”

Comparison at a glance

What this means for Salesforce customers

For customers already using Salesforce’s AI tools, Koa could make it easier to keep more of their automated work inside a controlled environment. That is especially relevant for companies in regulated industries or those handling sensitive customer information, where the path from experimentation to production can be slowed by legal and security reviews.

The model may also improve the economics of scaling AI assistants across large support and sales teams. If the same output can be achieved using fewer tokens, businesses may be able to deploy more agents without seeing costs rise as sharply as they would with a larger frontier model.

Still, Koa’s real-world value will depend on performance in production, not just on the promise of efficiency. Enterprise customers will want proof that it can handle messy conversations, edge cases and long workflows without losing accuracy or creating new operational problems.

What comes next

Koa’s debut suggests that the next phase of enterprise AI may be less about singular breakthroughs and more about specialization. The most valuable systems may be the ones that combine strong base models, synthetic or controlled training data, strict governance and platform-level routing that can choose among several models on the fly.

Salesforce is trying to position itself at the center of that architecture. With Koa, it is no longer only the company that helps businesses connect CRM records to third-party AI. It is now also a company that can offer its own reasoning layer for enterprise work.

That could put pressure on rivals across the stack, from model labs to enterprise software vendors. And it may force a more important question than which AI model is biggest: which one is actually built for the business problem at hand?