What Is Salesforce Koa? A CRM Reasoning Model, Explained for Calling Teams 

Salesforce Koa is first frontier reasoning model built specifically for CRM. Announced at Dreamforce 2026 and developed with NVIDIA, it’s designed to handle multi-step enterprise workflows and agentic tool use inside Agentforce. The three CRM tasks it was built to reason about, lead qualification, service escalation, and resolution, are almost always phone conversations. 

That’s not a stretch. That’s what Salesforce said. 

Most of the conversation around Salesforce Koa right now is about what it is. Fair enough. It’s new, the terminology is unfamiliar, and “reasoning model” isn’t a phrase most Salesforce admins had to care about before last week. But there’s a second question almost nobody is asking: what does Koa need from your org to actually work? A reasoning model is only as good as the data it reasons over. And one of the largest categories of customer interaction that goes unlogged in most Salesforce orgs is calls. 

That’s where this gets interesting. 

What Is Salesforce Koa?

Most AI models predict. They look at a prompt and figure out the most likely next word, sentence, or output. That works fine for summaries and email drafts. It breaks down when a decision has six steps and each one depends on what came before. 

A reasoning model works differently. It thinks through a problem in sequence, checks its own logic mid-process, and handles tasks where the right answer isn’t on the surface. 

Koa is that model. Built on NVIDIA’s Nemotron 3 Super architecture and post-trained entirely on synthetic data, it draws on what Salesforce described as “almost three decades’ worth of Salesforce business context.” Critically, that post-training used synthetic enterprise scenarios, not your customer records, not your call logs, not any real CRM data belonging to a Salesforce customer. 

Post-training, if you haven’t come across this term: it’s when a capable base model gets taught a specialization. Think of a physician who goes on to specialize in cardiology. The foundational training still applies. But now it gets applied to a specific and complex set of problems. Koa started as a powerful base model, then was trained further to understand CRM-specific workflows from the inside. 

Rohan Kumar, President and Chief Platform and Engineering Officer at Salesforce, introduced Koa at the Dreamforce 2026 keynote. The framing wasn’t “here’s a new AI feature.” It was closer to: here’s the reasoning layer Agentforce was always going to need. 

Why Salesforce Built Its Own Reasoning Model

Marc Benioff made a point at Dreamforce worth sitting with. General-purpose AI models are probabilistic. They “kind of know what’s going on.” They can discuss lead qualification in theory. They aren’t grounded in the actual rules your business runs on. 

That distinction matters when the stakes are a sales call, a service escalation, or a customer about to churn. 

Jayesh Govindarajan, EVP of Salesforce AI and Agentforce, and Silvio Savarese, EVP and Chief Scientist at Salesforce AI Research, described the model’s training differently. They talked about giving Koa knowledge of “what a service escalation looks like from the inside, what qualifying a lead actually involves, what ‘resolved’ means when a customer is on the other end.” 

Read those three examples carefully. Escalation. Lead qualification. Resolution. 

Those aren’t abstract CRM events. They’re what happens on calls. A customer escalates because their issue wasn’t resolved on chat. A lead qualifies because someone asked the right questions on a discovery call. A ticket gets marked resolved because an agent confirmed it with the customer, verbally. The Salesforce reasoning model Govindarajan and Savarese described understands these workflows because those workflows are, at their core, human conversations that get logged in CRM. 

So the bet Salesforce made with Koa is this: a model that understands CRM tasks from the inside will make better multi-step decisions than a general model handed a long prompt and asked to figure it out. That’s what separates a CRM reasoning model from a general-purpose one. 

Koa and NVIDIA Nemotron: What the Partnership Means

Nemotron 3 Super is the base architecture Koa is built on. NVIDIA’s Nemotron series is designed for enterprise AI deployment, which means it’s built with the scale and reliability a company like Salesforce needs when embedding a reasoning model into a platform used by hundreds of thousands of organizations. 

What the partnership signals beyond the architecture is that Salesforce isn’t trying to build a general-purpose AI research lab. They’re building a CRM-native AI stack, and the salesforce nvidia nemotron collaboration is the foundation of the reasoning layer inside it. 

Salesforce hasn’t published details on infrastructure costs, availability timelines beyond their public announcements, or how Koa will be priced. This post will be updated when they do. 

Where Koa Sits in the Salesforce AI Stack

Salesforce announced several interlocking pieces at Dreamforce 2026. Here’s how they connect: 

Agentforce 360 is the platform layer, agents, data, and actions working together to run automated workflows across the enterprise. 

AIforce is the interface layer, how users and teams interact with AI capabilities whether they’re in Lightning, Slack, or another Salesforce surface. 

Claudeforce brings Salesforce data and workflows into Anthropic’s Claude environment, with permissions and business rules intact. 

Koa sits underneath Agentforce 360 as the reasoning model built specifically for it. While other models can plug into Salesforce workflows, Koa is the one trained on CRM context from the ground up. It’s “built for Agentforce,” in Salesforce’s own framing. 

For the fuller picture of everything Salesforce announced, including how these pieces connect for calling teams, our full Dreamforce 2026 recap covers the broader stack and what each announcement actually means in practice 

What Koa Means for Teams That Make Calls in Salesforce

Here’s what most Salesforce Koa coverage will miss. The three use cases Salesforce named when explaining the model, escalation, qualification, resolution, are phone events. Not email threads. Not chat transcripts. Calls. 

So if Salesforce Koa is going to reason about those tasks, it needs call data in Salesforce to reason over. That’s where this becomes practical. 

Routing. A koa reasoning model that understands what qualifying a lead actually involves can decide who gets the call, when, and in what order, based on live CRM data. But it can only do that if routing decisions are informed by real call outcomes. Routing calls on live CRM data becomes genuinely intelligent when there’s a reasoning layer that understands the difference between a warm lead and a curious one. 

Grounding. Salesforce Koa reasons over what’s in your Salesforce org. Calls that aren’t logged, transcribed, and tagged don’t exist to it. A conversation that happened but never made it into a record is a data gap, and data gaps produce bad reasoning. Every call logged and transcribed isn’t just good hygiene. When a model reasons over your CRM data to make decisions, complete call records become the baseline. 

Handoff. When Koa decides a lead should be called now, based on behavior data, qualification signals, and CRM history, something has to make that call. An AI that places the call is how that reasoning converts into an actual customer conversation. And why the call is the bottleneck matters even more when the intelligence layer is a reasoning model, not a rule-based trigger. 

One thing to be direct about: 360 CTI doesn’t integrate with Koa. Koa was announced days before this post was written. Nobody has that integration yet. What’s true right now: Salesforce Koa will reason over the data in your org. The quality of your call data decides the quality of its reasoning. That’s not positioning. That’s how it works. 

The Data You’re Generating Now Is the Context Koa Will Reason Over Later 

Salesforce Koa is genuinely new territory. Salesforce has never shipped a purpose-built CRM reasoning model before, and the details are still emerging. What they’ve told us is that the model was built to understand escalation, qualification, and resolution from the inside. Those are phone tasks. And a model that reasons about CRM workflows can only reason about the data that’s actually in CRM. 

That’s the practical takeaway for any team running calls in Salesforce today. The call records you’re building now, transcribed, dispositioned, tied to the right records, are the context a reasoning model will work with when it’s deployed. Gaps in that data aren’t just reporting problems. They’re gaps in what the model can see.