Salesforce is no longer content to be a customer of the frontier labs. At its Dreamforce 2026 conference in San Francisco this week, the company unveiled Koa, a reasoning model built specifically for CRM tasks that it says keeps enterprise data inside its own walls rather than routing it through a third party.
Koa was not trained from scratch. Salesforce post-trained Nvidia’s open Nemotron 3 Super model on synthetic data drawn from roughly thirty years of accumulated business knowledge, then tuned it toward the kind of multi-step work agents need to finish on their own. Rohan Kumar, the company’s chief platform and engineering officer, described the goal during the Tuesday keynote as supporting “long-running agents that execute multiple tasks and complete complex outcomes.”
Salesforce reported that Koa “already matches or exceeds leading model performance on CRM actions with 3x fewer errors” on the company’s own benchmark, a suite its AI Research group built and has circulated across the industry for the past two years. That is a vendor claiming victory on a test it wrote, so the number deserves the same skepticism any self-graded benchmark earns until an outside evaluator runs it.
The pattern behind the number is more interesting than the number itself. Domain-specific models routinely undercut general-purpose ones on cost and hallucination rate simply because they are trained on a narrower slice of the world. Salesforce is betting that narrowness, not raw scale, is what wins inside a CRM workflow.
The launch arrived alongside three other products. AIforce lets Salesforce customers query, assign tasks, and build dashboards through natural language instead of navigating the standard interface. Claudeforce puts Anthropic’s Claude at the front of the platform itself. Thirty seven sales skills ship with it on day one, spanning account management, prospect research, pipeline hygiene, and deal review. Headless 360 opens the platform’s data to customers through APIs, MCP connections, and plugins, so they can work from tools of their choosing without touching salesforce.com directly.
CEO Marc Benioff framed the broader ambition during the keynote: “The SaaS-pocalypse was not about the end of software, but it may be about the end of software that makes humans do all the work.” Read plainly, that is Salesforce arguing its own product category will survive by automating the labor sitting on top of it, not the software itself.
Jason Hiner, writing in The Deep View, places Koa inside a wider move by companies such as Crowdstrike and Thomson Reuters to build their own models rather than lease intelligence from OpenAI, Anthropic, or Google. His read is that intelligence is becoming the asset worth protecting most inside a company, and handing that layer to an outside vendor risks leaking the proprietary data that makes a business defensible, especially to a vendor also serving direct competitors. He also notes why Salesforce stopped short of building a frontier lab: post-training an open model like Nemotron is far cheaper than training one from zero, and gets most of the benefit.
The open question Hiner raises is whether Salesforce eventually turns this into a product, letting other enterprises post-train their own private models on its infrastructure. If it does, the CRM company becomes a picks-and-shovels seller for the exact trend it just demonstrated. For any company weighing whether to lease a frontier model or fine-tune an open one for a narrow internal task, Koa is now the reference case to benchmark against before signing a multi-year AI contract.
Based on reporting and analysis by Jason Hiner for The Deep View, published September 16, 2026.
