Salesforce Koa Explained for Business Leaders
Salesforce Koa is the company's first reasoning model built specifically for CRM work, announced with NVIDIA on 15 September 2026. It is designed to help Agentforce agents carry out multi-step tasks such as updating an opportunity, routing a case or scheduling a follow-up. Koa is still in pilot, but it raises a question every Salesforce customer using AI will face: which model should handle which task, and at what cost? This article explains what Koa is, how it was built, who can use it, and what to do about it now.
What Is Salesforce Koa
A reasoning model is an AI model trained to work through problems step by step and decide which tools to use, rather than only generate text. Koa is a reasoning model tuned for the actions agents take inside Salesforce.
These are the facts Salesforce has published in its Koa announcement:
- Koa was built by post-training NVIDIA Nemotron 3 Super, an existing open model.
- Its training data is synthetic, modelled on nearly three decades of CRM deployments across more than 14 industries. No customer data was used.
- Salesforce controls the model weights and runs training and inference inside its own trust boundary.
- On Salesforce's own CRM benchmark, Koa matches or exceeds leading models on CRM actions with three times fewer errors.
Why Koa Matters for AI Costs and Accuracy
Not every AI task needs the largest model available. Updating a field, summarising a case or drafting a routine follow-up are narrow, repetitive jobs. Larger general-purpose models can do them, but you pay for capability you do not use.
Our analysis is that Koa is part of a wider move towards model routing: sending each task to the model that gives the right accuracy at the lowest cost. Salesforce's Enterprise AI Harness lists "Trusted Models", which routes work based on accuracy, performance, cost and business requirements, as one of its six capabilities. A CRM-specific model fits naturally into that approach.
Salesforce has not yet published Koa pricing, and its announcement does not say how customers will choose between Koa and other models in Agentforce. Any cost saving is therefore a reasonable expectation, not a confirmed figure.
How Koa Was Built and Tested
Salesforce trained Koa using supervised fine-tuning and reinforcement learning on synthetic scenarios. These reflect the reasoning, tool use and decision-making that Agentforce agents perform in real CRM workflows. The benchmark it reports is Salesforce's own suite of CRM tasks.
That matters when you read the performance claim. A vendor benchmark shows how a model performs on the tasks the vendor chose. It is a useful signal, but you should still test Koa on your own processes and data before relying on it.
Who Can Use Koa Today
Koa is already used inside Salesforce, including by an employee agent in Slack. It is now in customer pilots with 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero.
Salesforce expects general availability in winter 2026 in US regions. If your Salesforce org is hosted in the UAE, India, Europe or elsewhere for data residency reasons, confirm with Salesforce when Koa will be available in your region before planning around it.
What Businesses Should Consider
- Region and residency. A model that runs only in US regions may not fit your compliance position.
- Cost visibility. You cannot measure savings without knowing what AI currently costs you per agent and per task.
- Accuracy on your data. Plan a side-by-side test on real workflows rather than relying on the published benchmark alone.
- Change control. Switching the model behind an agent is a change like any other. It needs testing and sign-off.
Questions to Ask Your Salesforce Account Team
- When will Koa be available in the region where our org is hosted?
- How will Koa be priced compared with the models our agents use today?
- Will Agentforce route tasks to Koa automatically, or will we choose per agent?
- Which of our current agent actions would benefit most from Koa?
- How can we run our own accuracy test before switching?
Common Mistakes When Planning Around New AI Models
- Waiting for the perfect model. Agents built well today can switch models later. Data and process work does not go out of date.
- Treating benchmark results as guarantees. Your data and processes are different from any benchmark.
- Not tracking usage. Without a baseline, no one can tell whether a cheaper model saved money.
How Alu Cloud Consulting Can Help
We help Salesforce customers set up the measurement and governance that make model decisions straightforward: usage tracking per agent, test scenarios for key workflows, and a clear process for changing models. Our Salesforce ROI calculator is a useful place to start building the business case, and our Dreamforce 2026 summary shows where Koa fits among this year's launches.
Want to know whether Koa could reduce your Agentforce costs? Speak to our team.
Ready to talk about your Salesforce project?
Every engagement starts with a free discovery call. No pressure, just an honest conversation about where you are and what you are trying to build.
Frequently Asked Questions
What is Salesforce Koa?
Koa is Salesforce's first CRM reasoning model, announced with NVIDIA on 15 September 2026. It was built by post-training NVIDIA Nemotron 3 Super on synthetic CRM data and is designed to help Agentforce agents complete multi-step tasks such as updating opportunities, routing cases and scheduling follow-ups.
Was Koa trained on customer data?
No. Salesforce says no customer data was used to train Koa. Its training data was built from synthetic scenarios modelled on nearly three decades of CRM deployments across more than 14 industries. Salesforce also controls the model weights and runs training and inference inside its own trust boundary.
Is Koa available in the UAE or India?
Not yet. Koa is in customer pilots and Salesforce expects general availability in winter 2026 in US regions. If your Salesforce org is hosted in the UAE, India or another region for data residency reasons, ask your account team for regional timing before planning around Koa.
How much does Koa cost?
Salesforce has not published Koa pricing. It has also not said how customers will choose between Koa and other models inside Agentforce. Any cost saving from using a smaller, CRM-specific model is a reasonable expectation rather than a confirmed figure, so start tracking your current AI usage cost per agent now.
How accurate is Koa?
Salesforce reports that Koa matches or exceeds leading models on CRM actions with three times fewer errors, measured on its own CRM benchmark. Vendor benchmarks are a useful signal, but they reflect the vendor's chosen tasks. Test Koa on your own workflows and data before relying on it in production.