For most of Agentforce’s history, one general model did the reasoning for almost every task: Intent detection, prompt-injection screening, toxicity, evaluation, reranking, and so on. Over the past eighteen months we changed that, building a series of smaller models, each built for a specific job. But we left all the core multi-step reasoning with the general model.
Now that changes. Meet Koa, our first AI reasoning model optimized for enterprise work, running inside the Salesforce trust boundary. We co-engineered Koa with NVIDIA, post-training their NVIDIA Nemotron 3 Super model for the most common enterprise tasks. Customers can now use Koa to run their AI agents‘ processes.
A reasoning model for enterprise work
Reasoning is how an agent thinks through complex problems, evaluates its own logic, determines what tools to call, and verifies its path before providing an answer or taking an action. A frontier model does this well, but it reasons about a refund policy, a physics problem, and a travel plan with the same general intelligence, working each one through from first principles. That range is the point of a general model, and it is also its limitation. A frontier model reasons about your business from the outside, applying broad intelligence to work it has never actually done.
Consider how a person becomes good at their job. A new employee gets onboarded to learn policies and procedures, credit thresholds and compliance lines, and the specific ways a company qualifies a lead or clears a refund. Then they do the work enough times to know it cold, and they apply what they learned. That is what expertise is: a body of rules and hard-won practice, learned on the job and applied with consistency.
But a general model always starts from first principles. Every time it gets a task, it works it out from scratch, even if it’s done it a thousand times. And because it reasons probabilistically, it may not do it the same way every time. That’s great for tasks that require creativity. But governed, repeatable work favors a model that has learned the exact work to be done, and always leverages that expertise.
That is the model we set out to build.
How we built it
Training a reasoning model from scratch takes enormous data and time. Fortunately, we didn’t have to do that. NVIDIA had already built Nemotron 3 Super, a 120-billion-parameter open model that gave us two things closed frontier models never do. The first is direct access to the weights, so we could post-train it and adapt it to the exact work we wanted done. The second is provenance: NVIDIA published the datasets Nemotron was trained on, so we could see what went into the foundation instead of taking a vendor’s word for it. That gives our customers a trusted backbone — an open, US-trained model we control and can inspect, not a black box.
There are two ways to teach a model a job. The first is to have it imitate transcripts of work accomplished successfully — supervised fine-tuning, the industry’s default. That is like onboarding by handbook: read and memorize how something gets done. But this approach yielded limited gains. That’s because enterprise agents run multi-turn, tool-using workflows — too complicated to learn just by reading about them.
So instead, we made it do the work. Our teams built simulated enterprise workflows across 14+ industries, including manufacturing, financial services, healthcare, and travel. We generated synthetic customers with moods and personas, some cooperative and some frustrated. Tools responded to the agent’s calls, and a judge scored whether the customer’s problem actually got resolved. If it didn’t, the model tried again, running each scenario thousands of times, attempting to improve its score. This reinforcement learning approach improved the capability that matters most: multi-turn tool use, the ability to carry a task across many steps, tools, and turns of conversation.
To accomplish this, we drew on 27 years of experience running the same workflows we wanted the model to learn. We know 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. That knowledge is what a general model has never had, and it is exactly what we could give Koa.
One scenario deserves special mention, because it is the one enterprises worry about most. In our simulations, we deliberately put the model in situations where the right tool wasn’t available: the customer asked for something the agent cannot currently do. In those situations, a general model tends to fudge its response: call a similar-sounding tool that changes data it shouldn’t, or worse, confirm an action that never happened. We trained Koa to respond like a good employee should: stop, tell the customer plainly what it can’t do yet, ask for what it needs, and pick the task back up the moment, or hand it off to a human teammate when the situation calls for judgment a model shouldn’t make alone.. Knowing when not to act is enterprise expertise, too.
Take something an agent does every day: qualifying a new lead. The work is not open-ended. Your company has a definition of a qualified lead, a sequence for checking it against the account’s history and firmographics, a threshold for routing to sales, and a rule for what happens when the lead falls short. A general model reasons its way through those rules and definitions from the outside, and can reach a slightly different conclusion each time. A model trained on the work applies the same definition, to the hundredth lead as to the first. That is the difference between a model that is smart about qualification in general and one that qualifies leads the way your business does.
Koa doesn’t work alone. It builds on a portfolio of domain-specialized models, developed by Salesforce Engineering and AI Research, including HyperClassifier, TextEval, and most recently Moirai. By orchestrating these purpose-built models alongside frontier LLMs, we route targeted jobs—like intent classification, toxicity screening, and search reranking—to the exact model built for them, while Koa handles multi-step enterprise reasoning.
What we’re seeing so far
Koa is already running at Salesforce inside our employee workflows. We’re now moving into customer pilots with a small group of customers who are using it to run agents in service, sales, and commerce. Early benchmarking on CRM specific tasks such as updating an opportunity, routing a case, or scheduling a follow-up are encouraging, outperforming the strongest general-purpose models.
The part that stays yours
Another benefit to Koa: it runs entirely inside the Salesforce trust boundary, on infrastructure we operate. Customer data never leaves the customer’s control. None of the simulated work used to teach Koa uses real customer data: not the service desks, not the synthetic customers, not the scored resolutions. We didn’t need it. Twenty-seven years of running these processes is what told us how to build the simulation in the first place.
Protecting your data was core to how Koa was built. It’s also critical to how it runs: your data and your traces never train the model. The record of how your agents reason through your business, the accumulated pattern of how your company actually works, that’s the edge no competitor can copy. When you reason on a rented frontier model, that trace becomes fuel to improve the model. When you reason on Koa, your data and your traces stay yours.
Learn more
The right intelligence for the work in front of you, and the control to decide it. It’s already the choice our pilot customers are making across regulated industries — financial services, healthcare, travel, accounting — running agents on a model that’s been optimized for enterprise work. Learn more about Koa here.
