Salesforce is making a deeper push into enterprise artificial intelligence with Koa, a reasoning model designed specifically for the sales, marketing and customer-service tasks that power its Agentforce platform.

The model, announced around Salesforce's Dreamforce 2026 conference, represents a significant change in the company's AI strategy.

Rather than depending entirely on general-purpose models from companies such as OpenAI and Anthropic for complex reasoning, Salesforce is developing a model tailored to the specific work its enterprise customers want AI agents to perform.

According to Salesforce, Koa was built on Nvidia's open-weight Nemotron model and post-trained with Nvidia to handle enterprise tasks such as sales interactions, customer support and multi-step workflows.

The move reflects a broader shift taking place in enterprise AI.

Businesses are increasingly asking not simply whether an AI model is the smartest available, but whether it is secure enough, affordable enough, efficient enough and specialised enough to run business-critical work.

Enterprise AI has different priorities from consumer AI

The AI race has largely been defined by increasingly powerful general-purpose models.

Frontier AI companies have competed to produce systems that can reason, code, analyse documents, solve difficult problems and operate across a huge range of tasks.

Enterprise customers, however, often have a different problem.

A company may not need an AI model capable of solving an advanced mathematics problem if its primary requirement is to handle customer complaints, qualify a sales lead, update records or schedule an appointment.

It needs a model that can perform those tasks reliably while respecting the company's data policies.

That distinction sits at the centre of Salesforce's Koa strategy.

Salesforce says the model is designed around the specific workloads its Agentforce customers need rather than attempting to become another universal model competing across every benchmark.

The company's wider Dreamforce strategy is similarly focused on connecting AI reasoning with enterprise data, business rules, security and controlled execution. Salesforce describes this architecture as an Enterprise AI Harness, designed to allow agents and models to work with governed enterprise information while operating within business policies.

Koa is built on Nvidia's open-weight Nemotron

One of the important pieces of the strategy is the model Salesforce chose as its starting point.

Koa is based on Nvidia's Nemotron, an open-weight model that Salesforce and Nvidia then post-trained for enterprise applications.

That gives Salesforce a middle ground between two extremes.

On one side are closed frontier models controlled by external AI companies.

On the other are completely independent models that a company would have to develop and train largely from scratch.

Using an open-weight foundation gives Salesforce greater control over how the model is adapted and deployed while avoiding the enormous cost and technical challenge of creating a foundation model from zero.

For Salesforce, that control is particularly important because the model will ultimately sit inside a platform handling highly sensitive business information.

Salesforce says Koa was not trained on customer data

One of the most important claims surrounding Koa is that Salesforce did not use actual customer data to train the model.

Instead, Salesforce and Nvidia created synthetic data designed to replicate enterprise situations.

The training environment reportedly simulated scenarios involving customer-service representatives dealing with difficult customers as well as sales professionals attempting to close deals.

That approach is significant because enterprise AI adoption has repeatedly run into the same concern:

What happens to confidential business information when companies send it to an AI model?

Businesses may have customer records, financial information, internal communications, contracts and proprietary processes flowing through their systems.

A model specifically trained on customer information can create additional concerns around data separation, retention and leakage.

Salesforce's approach is therefore to teach Koa how enterprise work behaves without feeding it the actual private information belonging to the companies using Agentforce.

The customer's data remains within Salesforce's controlled environment and can be used to ground the agent's responses and actions without becoming part of the model's general training corpus.

The model is also about reducing AI costs

Another reason Salesforce is developing Koa is economics.

Reasoning models can consume significant numbers of tokens as they work through complex tasks.

For companies operating AI agents at scale, those costs can become substantial.

A customer-service agent handling thousands or millions of interactions does not necessarily need the most expensive general-purpose reasoning model for every request.

If a specialised model can achieve the same result using fewer tokens, the economics can become much more attractive.

Salesforce therefore positions Koa as a model that can deliver the reasoning required for particular enterprise workloads while using fewer tokens than sending every complex task to a large frontier model.

That fits into Salesforce's broader effort to make AI agents economically viable at enterprise scale.

The company's Dreamforce materials emphasise moving AI from experimentation toward scaled business use, while its Enterprise AI Harness is designed to allow companies to use different models and capabilities within a governed architecture.

Salesforce isn't abandoning OpenAI or Anthropic

Koa does not mean Salesforce is closing the door on external AI models.

In fact, the opposite appears to be happening.

Salesforce has been building a broader ecosystem in which different models can be used for different tasks.

Its AI gateway can determine which model should handle a particular request.

A simple task could be handled by a smaller specialised model.

A more complex task could be routed to a more capable frontier model.

And a company could potentially use different models while keeping the underlying business data and workflows inside Salesforce.

That approach turns Salesforce from simply being an AI-model developer into something potentially more important: an AI traffic controller for enterprise software.

The company can determine which model is appropriate for a task rather than forcing customers to choose one model for everything.

ClaudeForce shows the other side of the strategy

Salesforce's partnership with Anthropic provides a useful example.

The company recently announced ClaudeForce, allowing businesses to use Anthropic's Claude within Salesforce workflows while keeping enterprise data inside Salesforce's infrastructure.

That means Salesforce is simultaneously doing two things.

It is building its own specialised reasoning capability with Koa while integrating external frontier models such as Claude.

The strategy is not necessarily to replace every external AI model.

It is to make Salesforce the layer through which businesses access whichever model is most appropriate for a particular job.

That could become increasingly important as companies stop asking, “Which AI model should we use?” and start asking, “Which model should handle this particular task?”

This is where the AI gateway becomes important

An AI gateway may sound like a technical infrastructure detail, but it could become one of the most strategically important pieces of Salesforce's AI platform.

Imagine a customer-service operation receiving 100,000 requests.

The majority might be straightforward.

A smaller percentage could require deeper reasoning.

A few might involve highly sensitive information or require strict compliance controls.

Rather than sending all 100,000 requests to the same expensive model, the gateway can determine which model or workflow is appropriate for each task.

That creates an AI architecture that resembles cloud computing more than the current chatbot model.

Companies do not need every workload to run on the most powerful machine.

They need the right amount of computing power for each job.

Salesforce is attempting to apply the same principle to AI reasoning.

Nvidia benefits from the model shift too

The partnership is also significant for Nvidia.

The chipmaker is increasingly positioning itself not only as the supplier of GPUs used to train AI models, but as a provider of the software and model ecosystem around enterprise AI.

Nemotron gives companies an open-weight foundation they can adapt to specific requirements.

For enterprises, that can provide more control over model behaviour, deployment and data governance.

For Nvidia, it creates another route for its technology to become embedded deeper into corporate AI infrastructure.

The result is an ecosystem in which Nvidia provides the underlying computing and model foundation, while companies such as Salesforce specialise that technology around specific industries and workflows.

The real battle may be over AI infrastructure

The Koa announcement points to an important change in the AI industry.

The first phase of generative AI was largely about building the biggest and most capable models.

The next phase could be about making those models economically useful inside real businesses.

That means solving problems around:

• inference costs;

• data governance;

• security;

• model selection;

• latency;

• reliability;

• compliance;

• integration with existing software; and

• predictable business outcomes.

A company may not care whether its customer-service AI ranks first on a general reasoning benchmark.

It cares whether the agent answers customers correctly, follows company policies, updates the right records and does so without exposing confidential information.

That is a very different definition of AI performance.

Salesforce's bet is on specialised intelligence

Koa therefore represents more than Salesforce creating another language model.

It is a bet that specialisation will matter as much as raw intelligence in enterprise AI.

A smaller model trained specifically around sales and customer-service workflows could potentially outperform a much larger general-purpose model on the tasks Salesforce customers actually need.

If that happens, enterprises may have less reason to send every task to expensive frontier models.

That could change the economics of enterprise AI.

Instead of one giant model doing everything, businesses could operate a portfolio of models, each selected according to cost, complexity, security and performance.

Salesforce's AI gateway and broader Enterprise AI Harness are designed for precisely that kind of environment, where multiple models can operate within a common layer of enterprise data, controls and workflows.

The bigger question

Salesforce is not trying to win the AI race by becoming the next OpenAI or Anthropic.

Its opportunity may be different.

If businesses increasingly adopt AI agents for sales, customer service, marketing and other workflows, Salesforce controls an important part of the environment in which those agents operate.

Koa gives the company another tool inside that environment.

Claude gives it another.

Other models can provide others.

The strategic advantage may therefore come from deciding which model gets used, for which task, with what data and under which security rules.

That could be more valuable to an enterprise customer than having access to one universally powerful AI model.

And it explains why Salesforce's latest AI push is ultimately less about building a chatbot and more about building the infrastructure that determines how businesses use AI.

The enterprise AI race may no longer be about finding one model that can do everything. It may be about building systems that know which model should do what.

Salesforce's most interesting move isn't necessarily Koa itself. It's the architecture around it.

The company is effectively saying enterprises shouldn't have to choose between closed frontier models, open models and specialised models.

They should be able to use all of them — with an AI gateway deciding which model handles each job while Salesforce controls the surrounding data, permissions and workflows.

That could become a powerful business model.

The winners of enterprise AI may not be the companies with the single smartest model. They may be the companies that become the operating layer connecting many models to real business processes.