Add preferred source
Salesforce unveiled Koa, its first in-house CRM reasoning model, at the Dreamforce conference on Tuesday. Built in partnership with Nvidia on the open-weight Nemotron 3 Super, Koa was post-trained by Salesforce using synthetic data modeled on nearly three decades of CRM experience, with no actual customer data involved. The model matches or beats leading models on Salesforce’s CRM benchmark while producing three times fewer errors, and is designed for token-efficient inference to reduce enterprise AI costs. Salesforce retains full control of model weights and runs inference within its own trust boundary. Customer pilots are underway with companies including Formula 1 and UChicago Medicine, with general availability expected in Winter 2026. The launch positions Salesforce to handle complex reasoning tasks internally rather than relying on frontier model providers, while the company continues partnerships with Anthropic and others. Agentforce ARR has surpassed $1.5 billion, up over 240% year over year.
Key Elements
Salesforce(NYSE:CRM) unveiled its first in-house reasoning model on Tuesday, a direct challenge to frontier AI labs that have long supplied the cognitive horsepower behind the company’s enterprise automation platform. Dubbed Koa, the model was developed in partnership with Nvidia(NASDAQ:NVDA) and is purpose-built for the mundane but mission-critical work of updating sales records, routing support tickets, and scheduling follow-ups.
The announcement, made at Salesforce’s annual Dreamforce conference, marks a strategic pivot for the customer relationship management giant. Rather than continuing to route complex reasoning tasks to third-party models from Anthropic or OpenAI, Salesforce can now handle those workloads internally using a model it controls end-to-end. The company retains ownership of the model weights and runs post-training and inference within its own trust boundary, a setup designed to appeal to enterprises wary of sending sensitive data to outside providers.
Koa is built on Nvidia’s open-weight Nemotron 3 Super, which Salesforce post-trained using a proprietary synthetic dataset derived from nearly three decades of CRM experience. The training scenarios span more than 14 industries and cover processes, workflows, and operational policies. Crucially, no actual customer data was used, eliminating concerns about proprietary information leaking through model outputs.
On Salesforce’s internal CRM benchmark, which evaluates real-world tasks rather than general-purpose AI capabilities, Koa matches or outperforms leading models while producing three times fewer errors. The company says the model is already in use internally, including within a Slack agent that helps employees find information and complete routine tasks. Customer pilots are underway with organizations including 1-800Accountant, Baxter Credit Union, Formula 1, UChicago Medicine, and Xero, with general availability expected in Winter 2026 for U.S. regions.
“We’ve built many small task-specific language models, which are part of Agentforce’s portfolio,” Jayesh Govindarajan, EVP of Salesforce AI, told TechCrunch. “But reasoning has always been something that we’ve relied on the frontier model providers for. Until now.”
Govindarajan explained that the absence of a suitable base model had previously prevented Salesforce from training its own enterprise-grade system. “Until Nemotron came along, there was no sovereign American pre-trained model that was available, one, and two, that was state of the art, and, three, that had clear data provenance,” he said, adding that popular Chinese open-weight models lacked the transparency Salesforce required.
The token-efficiency angle is central to the value proposition. Kari Ann Briski, Nvidia’s VP of Generative AI Software for Enterprise, described Nemotron’s architecture as uniquely suited to the economics of AI agents, which often require multiple model calls to complete a single workflow. “It’s kind of the trifecta of things that you need to have: sovereign AI, time to first token, efficient reasoning, for the tokenomics of it all,” she said.
For enterprise customers, that translates to lower AI spending per task compared with sending prompts to general-purpose frontier models. The synthetic training data also means Koa was built from simulated scenarios — including irate customers calling service centers and sales professionals closing deals — rather than real client interactions.
Salesforce is not abandoning its relationships with external model providers. The company recently announced ClaudeForce, a partnership with Anthropic that lets businesses use Claude as their AI interface while keeping enterprise data within Salesforce’s infrastructure. Koa will function as an additional option within Agentforce, Salesforce’s platform for building autonomous agents, and can be automatically routed through an AI gateway depending on the task at hand.
Siemens Deployment Shows Enterprise Scale
Salesforce also announced an expanded partnership with Siemens at Dreamforce, demonstrating how Agentforce is being deployed across complex industrial workflows. Siemens is integrating Agentforce with its Teamcenter Service Lifecycle Management platform to bring engineering data into sales and service operations. The integration helps sales teams identify technically valid upgrades and enables service technicians to locate the correct spare parts for specific equipment.
Siemens is also using Agentforce to handle inbound lead qualification for its 18,000 sellers. The company receives more than 2,500 unqualified leads per month, and Salesforce said Siemens now engages 100% of those leads across 132 countries.
The Koa launch arrives as Agentforce continues to scale rapidly. Salesforce reported that Agentforce annual recurring revenue exceeded $1.5 billion in its fiscal second quarter of 2027, up more than 240% year over year. Combined ARR for Agentforce and Data 360 reached nearly $3.9 billion, up more than 210%.
Market Context
Salesforce shares were down 0.17% on Tuesday morning, while Nvidia shares rose more than 1%. The stock remains in a longer-term uptrend, trading 8.4% above its 20-day simple moving average of $237.11 and 27.6% above its 200-day SMA of $201.38. A golden cross formed in September, with the 50-day SMA crossing above the 200-day SMA, though near-term momentum indicators have cooled.
Key technical levels include resistance at $268.50, near the upper end of the recent range and just below the 52-week high of $269.11, and support at $252.00. The next major catalyst is the company’s earnings report expected on December 2, 2026, with analysts projecting EPS of $3.24 and revenue of $11.45 billion.
Wall Street maintains a Buy consensus on the stock with an average price target of $266.36. Recent analyst actions include Needham maintaining a Buy rating with a $400 target, Cantor Fitzgerald raising its target to $300 with an Overweight rating, and BTIG maintaining Buy at $300.
For Nvidia, the Koa partnership represents another validation of its open-weight model strategy. By providing a sovereign American base model with clear data provenance, Nvidia is positioning itself as the infrastructure layer for enterprises that want AI capabilities without ceding control to closed-model providers.
The competitive implications extend beyond Salesforce. As enterprises increasingly demand models tuned for specific business functions, with transparent training data and predictable token costs, the frontier labs’ one-size-fits-all approach may face growing pressure. Koa is an early but prominent example of how the enterprise AI market is diverging from the general-purpose race.
Once added, BigGo Finance appears first in Google Search Top Stories, so you get the broadest, most up-to-the-minute, and most comprehensive global <a href="https://bitcomme.com/financial-services-support-hub-celebrates-grand-opening/” title=”Financial Services Support Hub celebrates grand opening”>financial news first.
