What Is Salesforce Koa & How it Affects Your Business
Salesforce now has its own AI brain for CRM work. At Dreamforce on 15 September 2026, Salesforce and NVIDIA announced Koa, Salesforce’s first CRM reasoning model, built to power Agentforce agents.2
For two years, most AI agents in Salesforce have run on general-purpose models that know a little about everything. Koa is a different bet: a model trained specifically to understand deals, service cases and the workflows that connect them. In this post, I’ll break down what Koa is, how it was built, what the numbers actually say, and what it means for your org.
What is Salesforce Koa?
Koa is a reasoning model built for one job: helping Agentforce agents complete real CRM work. It is designed to think through complex, multi-step workflows and pick the right tools to get each step done.4
Think of it this way. A general AI model can write a poem, explain physics and summarise an email. Koa is less of a generalist and more of a specialist. It is meant to understand the structure of a deal, the lifecycle of a service case, and how workflows differ from one industry to the next.
Where Koa fits in an Agentforce agent
How Koa was built
Salesforce did not build Koa from scratch. It started from NVIDIA’s open-weight Nemotron 3 Super
model and then post-trained it with its own data and methods.5 Reports put the base at the 120-billion-parameter version of Nemotron 3 Super.9
The interesting part is the training data:
- 100% synthetic. The training set was made up entirely of artificial scenarios covering the reasoning, tool use and decisions agents need for work like generating leads, qualifying opportunities and resolving service cases.1
- Modelled on 27 years of CRM know-how. Salesforce based those scenarios on nearly three decades of enterprise CRM deployments.3
- Broad industry coverage. The scenarios span 14+ industries, and each one includes a persona, a sequence of actions and the tool calls needed to finish the task.6
- Two training stages. Salesforce used supervised fine-tuning and reinforcement learning, running on NVIDIA’s NeMo RL, NeMo Gym and NeMo AutoModel tools.12
Why build a CRM-specific model?
Most CRM requests are not one-step questions. “Update this opportunity, route the case to the right team and schedule a follow-up” is three actions, in order, each needing the right tool. That is where general models tend to slip.
Before Koa, one general model did the heavy multi-step reasoning in Agentforce, even though Salesforce had already moved smaller jobs like intent detection and toxicity screening to smaller models.7 Koa extends that idea to reasoning itself.
There is also a cost angle. Inside Agentforce, an AI gateway can send routine, high-volume CRM actions to the cheaper, CRM-tuned Koa, while tasks needing deeper judgement still go to frontier models.8
So Koa does not replace models like Claude or GPT. It sits alongside them.
How well does Koa perform?
Salesforce’s headline claim is that Koa makes three times fewer errors on CRM actions while matching or beating leading models. That comes from Salesforce’s own CRM benchmark, a set of real-world tasks like updating an opportunity, routing a case or scheduling a follow-up.4
The technical report paints a more balanced picture. Koa improves on its open-weight base, with the biggest gains in multi-turn tool use. It beats a strong proprietary baseline but still sits below the strongest frontier models.4 On the Tau2Bench customer service benchmark, Koa scored 69.41: ahead of GPT-4.1, but 14.58 points behind GPT-5.5.9
Trust and data privacy
For many companies, this is Koa’s strongest selling point.
- No customer data in training. Because the training data was entirely synthetic, no real customer data was used to teach the model.5
- Salesforce owns the weights. Salesforce controls Koa’s model weights and runs both post-training and inference inside its own trust boundary.3 Customer data does not have to leave Salesforce’s infrastructure when Koa processes it.
- Extra guardrails. A dedicated serving harness adds trust and safety controls around the model, and customer data and reasoning traces are not used to train it.7
How to use Koa in Agentforce
Koa shows up as a managed LLM in your generative AI models catalogue.1 Admins get two levels of control:7
The second option is the practical one. You can run Koa on a single high-volume agent, such as case routing, and compare its error rate against your current model before rolling it out more widely.
Availability and early adopters
Right now, Koa is open only to select pilot customers in Agentforce. General availability is expected in winter 2026, in U.S. regions.2 There is no word yet on other regions, including India.
Salesforce already uses Koa internally to power its Employee Agent. Named pilot customers include:6
Pricing has not been announced, which suggests Koa will likely be billed through existing Agentforce pricing.10
What Koa means for your business
Koa signals a shift in enterprise AI. The first wave was about using the smartest general model available. The next wave may be about owning the best workflow knowledge. Salesforce did not need to build a frontier model; it added decades of CRM context on top of NVIDIA’s foundation.
Salesforce now offers both a first-party model (Koa) and a deeply integrated third-party one (Claude, through Claudeforce). That makes model choice an ongoing governance decision rather than a one-time purchase.11
Questions to Ask Your Salesforce Account Team
Q1Will Koa-powered actions cost the same as frontier-model actions?
Q2: Can we see per-model error rates on our own workflows, not just internal benchmark?
Q3: When will Koa reach regions outside the US?
Q4: How does a Salesforce-owned model affect our flexibility to switch platforms later?
That last point matters. A Salesforce-controlled model adds value to staying on the platform, but it also raises the cost of leaving it.11
Final thoughts
Koa is one of the most credible attempts yet to prove that a smaller, specialised model can beat a big general one at a narrow job. It has real training infrastructure behind it, a strong data-privacy story, named pilot customers and a clear launch window.
It is not a replacement for frontier models, and its benchmark claims still need outside validation.
But if your business runs on Agentforce, Koa is worth watching closely and testing on one high-volume workflow as soon as it becomes available to you.
FAQs
What is Salesforce Koa?
Salesforce’s first CRM reasoning model, built with NVIDIA on Nemotron 3 Super and designed to power Agentforce agents.
Was my company’s data used to train Koa?
No. Koa was trained entirely on synthetic scenarios, with no customer data.
Does Koa replace Claude or GPT in Agentforce?
No. Koa handles routine, high-volume CRM reasoning, while more complex tasks can still go to frontier models.
When can I use Koa?
It is in pilot now, with general availability expected in winter 2026 in U.S. regions.
How much does Koa cost?
Salesforce has not announced separate pricing yet.
For more insights, updates, and expert tips, follow us on LinkedIn.
SOURCES
