Salesforce just made a bold bet at Dreamforce 2026: general AI models are not good enough for enterprise work, and the company is building its own to prove it.

On September 15, Salesforce and NVIDIA announced Koa, Salesforce’s first purpose-built CRM reasoning model. Built on NVIDIA’s Nemotron 3 Super architecture and post-trained on synthetic data derived from nearly three decades of CRM deployments across 14 industries, Koa is designed to handle the kind of complex, multi-step reasoning that general language models regularly fumble when applied to real business workflows.

The headline number: in Salesforce’s internal CRM benchmark, Koa delivers three times fewer errors than leading general models on CRM tasks. That is not a marginal improvement. For businesses running AI agents through Agentforce, fewer errors means fewer human interventions, fewer deals falling through the cracks, and fewer costly mistakes in customer-facing processes.

What Koa Actually Does

Koa is not a chatbot add-on. It is the reasoning engine underneath Agentforce agents, enabling them to work through multi-step workflows without getting confused by the real-world messiness of business data.

Think about what a sales agent actually needs to do: it needs to understand a deal’s history, compare it against similar opportunities, recommend the right next step, and then take action through Salesforce’s tools. That is not one question with one answer. That is a chain of reasoning with business context baked in at every step.

General AI models were not trained on this. They were trained on the internet. Koa was trained on patterns from 27 years of actual CRM deployments, which means it has internalized what good sales judgment and service workflows actually look like in practice.

Salesforce also controls the model weights and runs all inference entirely within its own trust boundary. Customer data is not used to train the model. For regulated industries, that distinction matters enormously.

The Missionforce Extension

Alongside Koa, Salesforce announced that Nemotron-based models will be available inside Missionforce, its platform for government and regulated enterprise clients. This extends Koa’s capabilities to air-gapped networks, private clouds, and secure deployment environments where commercial AI services typically cannot reach.

Koa is in pilot with select enterprise customers now and will reach general availability in US regions in winter 2026.

What This Means for Business

The launch of Koa is part of a broader shift happening across enterprise software. The era of plugging a general AI model into your systems and hoping it figures things out is ending. The companies that will win are the ones training domain-specific models on years of proprietary workflow data.

This matters for any business evaluating AI agents right now. A generic AI assistant might handle a simple question well. But when the task involves reasoning across customer history, contract details, compliance rules, and next-best-action logic, domain-specific models consistently outperform general ones. Koa is Salesforce’s answer to that problem.

For businesses running on Salesforce, the practical implication is that Agentforce agents powered by Koa should make fewer errors in the workflows you actually care about: sales cycles, customer service escalations, pipeline forecasting, and quote-to-cash processes.

For businesses not on Salesforce, the lesson is the same one that every vertical AI deployment is teaching right now: the more your AI is trained on the specific context of your industry and your workflows, the better it performs. Generic is not enough.

This is the direction Enterprise DNA has been building toward with Omni by Enterprise DNA. AI agents that understand your business context, your data, and your workflows outperform generic tools every time. Koa is Salesforce’s billion-dollar proof point.

---

If you are exploring how AI agents could be designed around your specific business data and workflows, that is exactly what the Omni Advisory team works through with clients. Book a session to talk through what domain-specific AI implementation could look like for your organisation.