Today, Agentforce agents use general-purpose AI models to decide what to do. These models are smart, but they aren’t trained specifically for CRM work like routing a case or qualifying a lead.
At Dreamforce 2026, Salesforce announced Salesforce Koa, its first AI reasoning model built specifically for CRM tasks. Here’s what Koa is, how it was trained, what the test results really show, and what it means for your org.
What Is Salesforce Koa?
Koa is a reasoning model, the part of an AI agent that figures out what steps to take and which tools to use.
Key facts:
- Built on NVIDIA Nemotron. Salesforce didn’t build Koa from scratch. It took NVIDIA’s open Nemotron 3 Super model and trained it further.
- Made for Agentforce. It’s designed to power Agentforce agents.
- Hosted by Salesforce. Koa runs on Salesforce’s own infrastructure.
- No customer data in training. Salesforce says Koa was trained on simulated scenarios, and Salesforce’s research paper adds public data.
Users won’t “open Koa.” It works behind the scenes, and admins choose it like any other AI model.
Why Salesforce Built It
Salesforce compares it to training a new employee. A smart new hire still needs to learn your company’s rules. A general AI model works out every task from scratch and may not do it the same way twice.
CRM work is often repetitive and rule-based. Salesforce believes a model trained on that exact work will be more consistent.
How Koa Was Trained
- Practice in simulations. Salesforce built fake business scenarios across 14+ industries, with simulated customers, some friendly and some frustrated.
- Learning by doing. The model tried tasks again and again, and a scoring system rewarded it for solving the customer’s problem correctly. This method is called reinforcement learning.
- Learning when to stop. Salesforce trained Koa for situations where it doesn’t have the right tool. Instead of guessing, it says what it can’t do or hands the task to a person.
That last point matters. An AI agent that says “I can’t do that” is much safer than one that pretends it did.
What the Test Results Show
Salesforce’s Claim
Salesforce says Koa makes three times fewer errors on CRM tasks in its own benchmark. It doesn’t say which models it compared against.
What Salesforce’s Research Paper Shows
Salesforce’s own research paper gives a more mixed picture:
Test (Salesforce internal) | Koa | GPT-4.1 | GPT-5.5 |
Multi-step CRM tasks completed | 87% | 82% | 91% |
Single-step CRM score | 0.87 | 0.81 | 0.90 |
In plain terms:
- Koa beats GPT-4.1 and the model it was built from.
- Koa is close to, but slightly behind, the strongest models like GPT-5.5.
- On one real-world test set, Koa made up the fewest tool calls of any model tested.
The “three times fewer errors” figure doesn’t appear anywhere in the paper. And all these results come from Salesforce testing its own model, with no independent tests yet.
So Koa looks promising, but it doesn’t beat everything. Test it on your own agents before switching.
FAQs
Was Koa trained on customer data? No. Salesforce says it used simulated and public data only.
Is Koa better than GPT or Claude? It depends on the task. In Salesforce’s paper, it beats GPT-4.1 but trails GPT-5.5 slightly.
Does Koa replace other AI models? No. It’s an extra option. Other model providers remain available.
The Takeaway
Koa is Salesforce’s first AI model trained specifically for CRM work. Its training method is smart, and it’s close to the best models on CRM tasks. But the results come only from Salesforce, and it’s still in pilot.
Next step: Build a small test set from one of your agents now, so you’re ready to compare Koa when it becomes available.
Read our Dreamforce 2026 recap for more announcements.
